Compare commits

...
258 Commits
Author SHA1 Message Date
GitHub Actions acfe799285 chore: bump Claude Code to 2.1.330 and Agent SDK to 0.3.330 2026-11-20 22:34:58 +00:00
3c38c04401 Use the base branch's version of files that the workflow's inputs name (#2099)
On pull requests from authors without write access, the action already
takes .claude, .mcp.json and similar files from the base branch. Do the
same for files and directories that the settings, plugin_marketplaces and
claude_args inputs name, where the pull request changes them, and keep
the session from editing them.

Co-authored-by: Claude <[email protected]>
2026-11-20 14:16:54 -08:00
GitHub Actions c84e279f85 chore: bump Claude Code to 2.1.329 and Agent SDK to 0.3.329 2026-11-19 22:15:27 +00:00
GitHub Actions f431adaf77 chore: bump Claude Code to 2.1.328 and Agent SDK to 0.3.328 2026-11-19 01:15:13 +00:00
5a85689334 Keep prompt runs to the actor's own comments and the step's output files in place (#2098)
* Don't run a prompt when someone else's comment or review is edited or deleted

docs/security.md says that editing or deleting another person's comment or review
doesn't start a run, even with a prompt. That held with track_progress only. It now
holds for every prompt. New comments and reviews run as before, and label_trigger
still doesn't apply to prompts without track_progress.

Co-authored-by: mhegazy <[email protected]>

* Keep a session from linking the step's output files elsewhere

GITHUB_OUTPUT and GITHUB_STEP_SUMMARY keep what a session writes to them, but a link
put in their place is replaced by the file from before the session, because the action
adds to them afterwards.

Co-authored-by: mhegazy <[email protected]>

---------

Co-authored-by: Claude <[email protected]>
2026-11-18 08:47:34 -08:00
23cd101ac6 Remove the unused prepare entrypoint (#2097)
action.yml has run src/entrypoints/run.ts since #898, and nothing imports prepare.ts.

Co-authored-by: Claude <[email protected]>
2026-11-18 08:47:21 -08:00
6c3921da9b Send Bun commands in claude.yml to the safety review (#2096)
The workflow no longer lists Bun commands as allowed tools, so the review decides on
them. The system prompt says what to do when one is denied.

Co-authored-by: Claude <[email protected]>
2026-11-18 08:47:11 -08:00
e5812cd756 Add an allow_fork_prs input (#2094)
Set to "false", the action skips pull requests from forks, and workflow_run
events for a fork's run, before it checks anything out. The default keeps
today's behavior.

Co-authored-by: Claude <[email protected]>
2026-11-17 16:09:22 -08:00
47bf91ccca Skip fork pull requests in claude.yml (#2092)
In this repository, a mention on a pull request from a fork no longer starts
Claude. A first job looks up where the pull request's branch is, because the
event for a comment doesn't say.

Co-authored-by: Claude <[email protected]>
2026-11-17 16:09:18 -08:00
87ff2368a1 Apply the session's git settings to the action's own commands (#2093)
The action's own git commands now use safe.bareRepository=explicit, as the
session does. On pull requests that restrict_untrusted_prs applies to, the
checkout leaves Git LFS files that aren't downloaded yet as pointer files.

Co-authored-by: Claude <[email protected]>
2026-11-17 16:08:55 -08:00
GitHub Actions 469588f42e chore: bump Claude Code to 2.1.327 and Agent SDK to 0.3.327 2026-11-17 23:18:14 +00:00
6070f08d2b Skip images when a description or comment has been edited (#2091)
Images are downloaded for a description, comment or review only if its text
on GitHub is still the text the run uses. Otherwise the action logs a warning
and leaves the image links as they are.

Co-authored-by: Claude <[email protected]>
2026-11-17 20:46:41 +00:00
8c07613f66 Use the default branch's config on pull requests from authors without write access (#2088)
* Use the default branch's config on pull requests from authors without write access

When a pull request's author does not have write access and it targets a branch
other than the default one, the config that the action restores now comes from
the base_branch input or the default branch instead of the pull request's base.
The base is still fetched so that the two can be compared.

Fetch and check out the config branch by its full ref name, so that a tag of the
same name or a narrow refspec can't leave the session without config.

Document that pr-N holds a fork's commits and should stay out of CI that holds
secrets.

Co-authored-by: mhegazy <[email protected]>

* Keep the other branch's config out of commits and the base comparable

The commit wrapper leaves out files that the action took from a branch other than
the pull request's base. HEAD's history is fetched again if fetching the config
branch cut it, and a base branch that can't be fetched fails the run as before.
A bot's pull request into a pr-N branch no longer takes its config from it.

Co-authored-by: mhegazy <[email protected]>

---------

Co-authored-by: Claude <[email protected]>
2026-11-17 12:45:54 -08:00
GitHub Actions cc98b3ae24 chore: bump Claude Code to 2.1.326 and Agent SDK to 0.3.326 2026-11-16 22:11:51 +00:00
0b40c94f8c Read the run step's script whole and turn off Bun's transpiler cache (#2060)
Bash now reads the rest of the run step's script before it starts Bun, and exits
right after. The action's own scripts run without Bun's transpiler cache.

Co-authored-by: Claude <[email protected]>
2026-11-16 13:37:51 -08:00
d1f58ad5e6 Use the base branch's AGENTS.md on pull requests (#2059)
Claude Code can load AGENTS.md as project instructions, as it does CLAUDE.md.
On pull requests the root AGENTS.md now comes from the base branch too. The
pull request's version is kept under .claude-pr/.

Co-authored-by: Claude <[email protected]>
2026-11-16 11:02:42 -08:00
fe99289f39 Keep deny rules that hold commas or parentheses (#2033)
* Keep deny rules that hold commas or parentheses

A narrower deny rule for a git command, such as git add on a path with
parentheses, wasn't carried over whole to the wrapper script that runs it. One
with a comma could lose its end when the same text also appeared in the rule
added for the wrapper.

Co-authored-by: mhegazy <[email protected]>

* Find deny rules next to line breaks and across values

A rule after a comma and a line break is read whole too, and a rule written as
two quoted values is carried over again.

Co-authored-by: mhegazy <[email protected]>

---------

Co-authored-by: Claude <[email protected]>
2026-11-16 11:02:36 -08:00
514ab7e6d3 Close gaps in using the base branch's npm and Bun config (#2026)
On pull requests from authors without write access, check package.json as it is on disk where .gitattributes converts it, take .gitattributes from the index, keep Bash in the repository root, and match restored directories exactly. Other directories fields than bin no longer count as commands.

Co-authored-by: Claude <[email protected]>
2026-11-16 11:01:31 -08:00
bcbcedd16c Start Bun with the action's own tsconfig again (#2058)
The run step, the post steps and the MCP servers pass --tsconfig-override
next to --no-env-file and --config. The Bun message that led to removing it
in #1315 is gone from 1.4.3, and the action pins 1.4.4.

Co-authored-by: Claude <[email protected]>
2026-11-16 19:01:23 +00:00
9fa68854da Use the base branch's npm and Bun config on outside pull requests (#2020)
* Use the base branch's npm and Bun config on outside pull requests

Config from the base branch, such as a hook or an MCP server, can start npm or Bun
in the checkout. On pull requests from authors without write access, what those tools
read from the repository root now comes from the base branch wherever it differs.

Co-authored-by: mhegazy <[email protected]>

* Leave already restored directories alone when replacing tool config

The replacement could rewrite files that were restored from the base branch just
before. It also now keeps a new package's name, and handles a directory in a config
file's place and a file that is re-encoded on checkout.

Co-authored-by: mhegazy <[email protected]>

---------

Co-authored-by: Claude <[email protected]>
2026-11-13 22:23:52 +00:00
GitHub Actions f3b070c881 chore: bump Claude Code to 2.1.325 and Agent SDK to 0.3.325 2026-11-13 20:41:04 +00:00
4d75e5f390 Make the git wrappers independent of the session's environment (#2018)
* Make the git wrappers independent of the session's environment

Tag mode's wrappers for git add, rm, commit and push now take their settings from
files the action writes, and always work on the repository the action checked out.
Git settings passed through the environment no longer reach them, so a workflow that
relies on one, such as GIT_SSH_COMMAND, needs its own allow rule.

Co-authored-by: mhegazy <[email protected]>

* Keep narrow deny rules and other folders covered with the git wrappers

A narrower deny rule such as git add -A now carries over to the wrapper. Global git
config refuses a repository found in a plain folder for the session, since the
settings no longer sit in the environment. The add wrapper also takes --chmod, and
the wrappers use the home directory the action recorded.

Co-authored-by: mhegazy <[email protected]>

---------

Co-authored-by: Claude <[email protected]>
2026-11-12 15:13:42 -08:00
GitHub Actions 7ca5eaf924 chore: bump Claude Code to 2.1.323 and Agent SDK to 0.3.323 2026-11-12 22:13:11 +00:00
274a39507b Let Claude read CI results on outside pull requests (#2009)
* Add restrict_untrusted_prs input and enable it in claude.yml

A command like `bun test` runs whatever the checkout contains, in a job that
holds the workflow's tokens. On a pull request from an outside contributor,
that code is theirs. The new input denies common code-running commands when
the pull request's author does not have write access, even if claude_args
allows them. It is off by default, and this repository's claude.yml opts in.

Co-authored-by: mhegazy <[email protected]>

* Deny code-running commands on untrusted PRs by default

Turn restrict_untrusted_prs on by default, so every workflow gets it without
opting in. Only an explicit false turns it off, so a mistyped value keeps the
restriction.

Stop logging the content of claude_args or raw lookup errors: warnings now
report a rule count and the HTTP status. Also retry rate-limited 403s, skip the
restriction when tag mode branches from the base on a closed or merged PR, warn
when permission checks are turned off, and parse unquoted Bash rules the way
the SDK option parser does.

Co-authored-by: mhegazy <[email protected]>

* Add an upgrade note for the restrict_untrusted_prs default

Workflows on a floating tag or branch pick up the new default without changing
anything, and the only signal is a log warning, so call that out in the docs.

Co-authored-by: mhegazy <[email protected]>

* Let @claude read CI results on outside pull requests

Test commands are denied there, so maintainers approve the CI run and Claude
reports its result instead. Document that path, and give ci.yml a read-only
token on every run.

Co-authored-by: qing <[email protected]>

---------

Co-authored-by: Claude <[email protected]>
Co-authored-by: mhegazy <[email protected]>
2026-11-12 21:08:36 +00:00
8a29bc6f0a Call out breaking changes in release notes and titles (#2013)
* Call out breaking changes in release notes and titles

Workflows on the floating v1 tag get each release straight away, and notes
listed only pull request titles. Pull requests labeled breaking-change now get
their own section, and the release title says so.

Co-authored-by: mhegazy <[email protected]>

* Start release notes from the last published release

A failed run can leave a tag with no release. Starting from the highest tag
would drop that run's pull requests from the next notes, and could lose the
breaking changes flag.

Co-authored-by: mhegazy <[email protected]>

---------

Co-authored-by: Claude <[email protected]>
2026-11-12 09:52:32 -08:00
e52db45098 Only trust text and runs from the person who passed the write check (#2002)
* Only trust text and runs from the person who passed the write check

A maintainer's triage action could turn someone else's text into the trusted request, or start a run they did not ask for.

- Issue bodies and comments become the request only when their author is the actor.
- Editing or deleting someone else's comment or review no longer starts a run.
- With a prompt and track_progress, label_trigger and assignee_trigger still apply.
- workflow_run also checks the owner of the fork the run came from.
- A bot's push to a pull request also checks the pull request's author.
- The comment cutoff for a label or assignment uses the actor's own event.

Co-authored-by: mhegazy <[email protected]>

* Keep bot activity quiet and let workflows accept any label

Review found that the new checks failed jobs with nothing to do and left some paths open.

- The check on a pull request's author now runs only when a run would start, so bot activity with no trigger still skips quietly.
- An empty label_trigger accepts any label when a prompt is set, instead of crashing.
- Editing or deleting someone else's comment or review no longer starts a run with a prompt either.
- The comment cutoff for an assignment follows the person who made it, even if the event's actor is the assignee.

Co-authored-by: mhegazy <[email protected]>

---------

Co-authored-by: Claude <[email protected]>
2026-11-12 09:50:10 -08:00
75b2776d96 Stop forced pushes and file reads through the git allow rules (#2006)
* Stop forced pushes and file reads through the git allow rules

Tag mode lets Claude run its push wrapper and git commit without asking, and both accepted more than a normal commit and push.

- The push wrapper rejects a ref starting with +, which git treats as a forced update.
- A new commit wrapper replaces the git commit allow rule. It accepts only a literal message and basic flags, so no flag can read a file into a commit.

Co-authored-by: mhegazy <[email protected]>

* Keep deny rules working with the git wrappers

Deny rules match command text, so denying git commit or git push no longer blocked the wrappers that run them. The commit wrapper also stops offering to rewrite a commit, because a pushed one could not be published again.

Co-authored-by: mhegazy <[email protected]>

* Read deny rules the same way the CLI receives them

Unquoted rules, the equals form and mid-word wildcards all reached the CLI but were missed when carrying deny rules over to the git wrappers, so those workflows could still commit and push.

Co-authored-by: mhegazy <[email protected]>

* Only let the push wrapper update the session's own branch (#2008)

* Only let the push wrapper update the session's own branch

The wrapper accepted any valid branch name, so a prompt-injected session could update main or another person's branch once it moved that branch locally. It now pushes only to the branch the run was set up on.

Co-authored-by: mhegazy <[email protected]>

* Repair the tag mode tests after merging

The merge left the test file's two new test groups interleaved, so it did not parse.

Co-authored-by: mhegazy <[email protected]>

---------

Co-authored-by: Claude <[email protected]>

* Note the pr-N branch for fork pull requests in the push docs

On a pull request from a fork, the session's branch is pr-N in the base
repository, not the pull request's own branch. A pull request opened from the
default branch can also update that branch, so branch protection still matters.

---------

Co-authored-by: Claude <[email protected]>
2026-11-12 17:49:56 +00:00
650432d396 Keep fork branch names away from this repository's branches (#2007)
* Keep fork branch names away from this repository's branches

With commit signing on, agent mode wrote commits through the API to the branch named by a fork PR. A fork branch named release targeted this repository's release, and main#pwn targeted main, because the # cut the URL short.

- Fork PRs in agent mode now use pr-N as tag mode does.
- Branch names are encoded in the Git references API URLs.

Co-authored-by: mhegazy <[email protected]>

* Use pr-N for comments and reviews on pull requests

A comment on a pull request fell through to the default branch, so signed commits from that run landed there. Branch lookups also become exact, because a prefix match made pr-5 fail once pr-50 existed.

Co-authored-by: mhegazy <[email protected]>

* Use issue-N for signed commits on plain issues

Prompt runs on an issue wrote commits through the API to the workflow's
branch, usually the default branch. They now use their own branch, as pull
request runs do. CLAUDE_BRANCH still takes precedence.

Co-authored-by: mhegazy <[email protected]>

---------

Co-authored-by: Claude <[email protected]>
2026-11-12 09:49:27 -08:00
462c312141 Stop a session from leaving programs for later steps to run (#2003)
* Add restrict_untrusted_prs input and enable it in claude.yml

A command like `bun test` runs whatever the checkout contains, in a job that
holds the workflow's tokens. On a pull request from an outside contributor,
that code is theirs. The new input denies common code-running commands when
the pull request's author does not have write access, even if claude_args
allows them. It is off by default, and this repository's claude.yml opts in.

Co-authored-by: mhegazy <[email protected]>

* Deny code-running commands on untrusted PRs by default

Turn restrict_untrusted_prs on by default, so every workflow gets it without
opting in. Only an explicit false turns it off, so a mistyped value keeps the
restriction.

Stop logging the content of claude_args or raw lookup errors: warnings now
report a rule count and the HTTP status. Also retry rate-limited 403s, skip the
restriction when tag mode branches from the base on a closed or merged PR, warn
when permission checks are turned off, and parse unquoted Bash rules the way
the SDK option parser does.

Co-authored-by: mhegazy <[email protected]>

* Add an upgrade note for the restrict_untrusted_prs default

Workflows on a floating tag or branch pick up the new default without changing
anything, and the only signal is a log warning, so call that out in the docs.

Co-authored-by: mhegazy <[email protected]>

* Stop a session from leaving programs for later steps to run

Steps after the session, such as the checkout action's cleanup, hold the job's credentials and trust files the session can change.

- GITHUB_ENV, GITHUB_PATH and git's config files are restored when the session ends.
- Git ignores core.fsmonitor while the session runs and for the action's own commands afterwards.
- Git hooks are off for the action's own commits, and during sessions on restricted pull requests.
- The token revocation step reads the token from the environment, since a session can write step outputs and an expression in a script could run its commands.

Co-authored-by: mhegazy <[email protected]>

* Close more ways a session could affect later steps

Review found gaps in the post-session restore, and one regression for Git LFS.

- Restoring can no longer hang on a FIFO or read a device, and removes anything planted at a tracked path.
- Config files pulled in through include.path are restored too, which covers the config actions/checkout keeps there.
- The action's own commits upload Git LFS objects, because the hook that normally does it is off.
- Hooks go off before an untrusted pull request is checked out, not only when the session starts.
- A session can no longer use a step output or ~/.curlrc to stop or redirect token revocation.

Co-authored-by: mhegazy <[email protected]>

* Move the session guard import away from the permissions import

This keeps the change from conflicting with the permissions check being added to the same import block.

Co-authored-by: mhegazy <[email protected]>

---------

Co-authored-by: Claude <[email protected]>
2026-11-12 17:49:00 +00:00
ead3158f27 Keep the token out of .git/config when bots are allowed (#2005)
* Keep the token out of .git/config when bots are allowed

Allowed bots skip the write check and relay text other people wrote, so a prompt injection in that text could read the token from the origin URL. Workflows with allowed_bots now use the credential helper that allowed_non_write_users already uses.

A later step that pushes with the origin remote now needs its own credentials in those workflows.

Test contexts now get their own copy of the default inputs, because one test file's allowedBots value was leaking into later tests.

Co-authored-by: mhegazy <[email protected]>

* Stop the default test context sharing its inputs

createMockContext and createMockAutomationContext still returned the shared default inputs when a test passed no overrides, so an allowedBots value set by one test file reached the credential tests.

Co-authored-by: mhegazy <[email protected]>

* Limit the credential helper to this server and to runs with a token

The helper answered for every host git contacted, so it could hand the token to
another server. It also answered with an empty password in later steps, which
stopped git from trying their own credentials. Origin now names a username, so
rewrite rules on the server's prefix can't bypass the helper, and the helper
works from a path with a space.

Co-authored-by: mhegazy <[email protected]>

* Say which credentials work for later steps using origin

Origin now names a username, so a credential stored under another username is
not found. The docs and the log line now list what does work.

Co-authored-by: mhegazy <[email protected]>

---------

Co-authored-by: Claude <[email protected]>
2026-11-12 09:24:16 -08:00
8565ca8f18 Run sessions in auto mode by default (#2004)
* Run sessions in auto mode by default

A safety review now decides tool calls that no allow rule covers, instead of those calls being denied outright. Set --permission-mode in claude_args to opt out.

- Runs from mentions used to default to acceptEdits, and runs with a prompt to default.
- Mention runs switch to acceptEdits and warn when auto mode isn't available, such as on an older model, so they keep their file edits.
- Sessions with a custom ANTHROPIC_BASE_URL keep the earlier defaults, because the check depends on what a gateway passes through.

Co-authored-by: mhegazy <[email protected]>

* Update the permission mode comment in claude.yml

The action now defaults to auto mode, so the comment no longer describes acceptEdits being replaced.

Co-authored-by: mhegazy <[email protected]>

* Keep the earlier default when non-write users are allowed

Anyone's text reaches these runs, so a call the earlier defaults deny shouldn't be left to the safety review. Workflows can still opt in with the flag. The FAQ also no longer says Bash is disabled by default.

Co-authored-by: mhegazy <[email protected]>

* Keep a mode or gateway set in the settings input

Workflows that set a permission mode or base URL in the settings input now keep the earlier defaults. Otherwise auto mode would silently replace a chosen mode, or start behind a gateway that may not support it.

Co-authored-by: mhegazy <[email protected]>

* Keep earlier defaults for allowed bots and settings files

Allowed bots skip the write check and relay other people's text, so they get the same default as non-write users. A settings file no longer helps choose the default, because a pull request can change it before the base copy is restored.

Co-authored-by: mhegazy <[email protected]>

* Keep earlier defaults with a custom Claude Code executable

A version that predates auto mode exits on the flag, so runs that supply their
own executable failed at startup. They now need to opt in. The edit fallback
also applies when auto mode is set in the equals form.

Co-authored-by: mhegazy <[email protected]>

---------

Co-authored-by: Claude <[email protected]>
2026-11-12 09:24:04 -08:00
3fd089526c Deny code-running commands on PRs from authors without write access (#2001)
* Add restrict_untrusted_prs input and enable it in claude.yml

A command like `bun test` runs whatever the checkout contains, in a job that
holds the workflow's tokens. On a pull request from an outside contributor,
that code is theirs. The new input denies common code-running commands when
the pull request's author does not have write access, even if claude_args
allows them. It is off by default, and this repository's claude.yml opts in.

Co-authored-by: mhegazy <[email protected]>

* Deny code-running commands on untrusted PRs by default

Turn restrict_untrusted_prs on by default, so every workflow gets it without
opting in. Only an explicit false turns it off, so a mistyped value keeps the
restriction.

Stop logging the content of claude_args or raw lookup errors: warnings now
report a rule count and the HTTP status. Also retry rate-limited 403s, skip the
restriction when tag mode branches from the base on a closed or merged PR, warn
when permission checks are turned off, and parse unquoted Bash rules the way
the SDK option parser does.

Co-authored-by: mhegazy <[email protected]>

* Add an upgrade note for the restrict_untrusted_prs default

Workflows on a floating tag or branch pick up the new default without changing
anything, and the only signal is a log warning, so call that out in the docs.

Co-authored-by: mhegazy <[email protected]>

---------

Co-authored-by: Claude <[email protected]>
2026-11-12 09:14:31 -08:00
d5f7970ae2 Pass structured output to shell steps through env (#1963)
* Pass structured output to shell steps through env

Claude's answer can contain quotes, which break a script when GitHub writes the value into it before the shell reads it. The example, the docs and the test workflow now pass it through env, and a test checks that they keep doing so.

Co-authored-by: inigo <[email protected]>

* Keep fromJSON out of env in the structured output docs

Co-authored-by: inigo <[email protected]>

---------

Co-authored-by: Claude <[email protected]>
2026-11-10 13:07:47 -08:00
GitHub Actions 4fa93a88de chore: bump Claude Code to 2.1.322 and Agent SDK to 0.3.322 2026-11-10 02:30:41 +00:00
GitHub Actions 9eb012b6d0 chore: bump Claude Code to 2.1.321 and Agent SDK to 0.3.321 2026-11-09 19:22:42 +00:00
Chris NormanandGitHub 115e0b2c49 release: push tags with the release deploy key (enables tag-ruleset enforcement) (#1650)
* release: push tags with the release deploy key

So the tag ruleset (terraform-config#81310) can be enforced: GITHUB_TOKEN can never bypass a ruleset, a deploy key can. No change to how releases are cut.

* release: SHA-pin actions/checkout in the deploy-key jobs

The checkout steps now hold the ruleset-bypass deploy key, so resolve actions/checkout by full SHA (v6.0.2, same pin as sync-base-action.yml) rather than the mutable v6 tag.
2026-11-06 14:21:18 -05:00
GitHub Actions 61a9c274fa chore: bump Claude Code to 2.1.319 and Agent SDK to 0.3.319 2026-11-06 17:05:12 +00:00
GitHub Actions c0949c50a7 chore: bump Claude Code to 2.1.318 and Agent SDK to 0.3.318 2026-11-04 00:03:23 +00:00
GitHub Actions 4845bacbb3 chore: bump Claude Code to 2.1.317 and Agent SDK to 0.3.317 2026-11-03 01:27:13 +00:00
GitHub Actions d535f5172d chore: bump Claude Code to 2.1.316 and Agent SDK to 0.3.316 2026-10-31 01:03:26 +00:00
GitHub Actions 6d5ae526d2 chore: bump Claude Code to 2.1.314 and Agent SDK to 0.3.314 2026-10-30 01:07:44 +00:00
GitHub Actions d70510b174 chore: bump Claude Code to 2.1.312 and Agent SDK to 0.3.312 2026-10-27 20:25:35 +00:00
98a9e39fe1 docs: use claude-plugins-official in plugin examples (#1932)
The plugin examples point at the older claude-code-plugins marketplace. New workflows should install from claude-plugins-official, which is where these plugins are developed.

Co-authored-by: Claude <[email protected]>
2026-10-27 12:51:22 -07:00
GitHub Actions 8041f1d17e chore: bump Claude Code to 2.1.311 and Agent SDK to 0.3.311 2026-10-26 21:53:19 +00:00
925afb5377 chore: bump Bun to 1.4.4 (#1916)
Pin oven-sh/setup-bun to Bun 1.4.4 in action.yml, base-action/action.yml
and the CI workflow (the prettier job previously floated on latest), and
update @types/bun to 1.4.3, the newest published types, in both packages.

Co-authored-by: GitHub Actions <[email protected]>
2026-10-24 00:45:41 +00:00
GitHub Actions 16f9f71ab9 chore: bump Claude Code to 2.1.310 and Agent SDK to 0.3.310 2026-10-23 23:40:57 +00:00
Octavian GuzuandGitHub 8d70a0dd26 Pin the PR version a run checks out (#1915)
Squash merge of 2 commits
2026-10-23 09:26:23 -07:00
GitHub Actions ae68417a74 chore: bump Claude Code to 2.1.309 and Agent SDK to 0.3.309 2026-10-22 22:20:35 +00:00
GitHub Actions 6b6870e649 chore: bump Claude Code to 2.1.308 and Agent SDK to 0.3.308 2026-10-22 04:36:28 +00:00
GitHub Actions aa8a7738e4 chore: bump Claude Code to 2.1.307 and Agent SDK to 0.3.307 2026-10-21 22:19:04 +00:00
GitHub Actions a5a423f75c chore: bump Claude Code to 2.1.306 and Agent SDK to 0.3.306 2026-10-21 03:18:30 +00:00
GitHub Actions e4a6faad49 chore: bump Claude Code to 2.1.305 and Agent SDK to 0.3.305 2026-10-20 20:48:51 +00:00
GitHub Actions db77d4d3c9 chore: bump Claude Code to 2.1.304 and Agent SDK to 0.3.304 2026-10-19 21:37:01 +00:00
GitHub Actions 5eb090f3c3 chore: bump Claude Code to 2.1.303 and Agent SDK to 0.3.303 2026-10-18 22:29:02 +00:00
GitHub Actions f27be6b65a chore: bump Claude Code to 2.1.302 and Agent SDK to 0.3.302 2026-10-16 20:54:52 +00:00
GitHub Actions 374f254f5e chore: bump Claude Code to 2.1.301 and Agent SDK to 0.3.301 2026-10-15 19:24:46 +00:00
GitHub Actions a26b07d8c7 chore: bump Claude Code to 2.1.300 and Agent SDK to 0.3.300 2026-10-14 18:27:27 +00:00
GitHub Actions 34968b5981 chore: bump Claude Code to 2.1.299 and Agent SDK to 0.3.299 2026-10-14 02:51:42 +00:00
GitHub Actions 1d6de8cb0c chore: bump Claude Code to 2.1.296 and Agent SDK to 0.3.296 2026-10-09 19:32:29 +00:00
GitHub Actions 2dca132ff0 chore: bump Claude Code to 2.1.295 and Agent SDK to 0.3.295 2026-10-08 19:51:21 +00:00
GitHub Actions 38c80c1a32 chore: bump Claude Code to 2.1.294 and Agent SDK to 0.3.294 2026-10-08 17:21:17 +00:00
GitHub Actions 6fed3ca145 chore: bump Claude Code to 2.1.293 and Agent SDK to 0.3.293 2026-10-07 18:13:38 +00:00
GitHub Actions 58985842b8 chore: bump Claude Code to 2.1.292 and Agent SDK to 0.3.292 2026-10-06 19:02:41 +00:00
GitHub Actions 86d88e619d chore: bump Claude Code to 2.1.291 and Agent SDK to 0.3.291 2026-10-06 03:57:41 +00:00
GitHub Actions 6c690188da chore: bump Claude Code to 2.1.290 and Agent SDK to 0.3.290 2026-10-05 23:35:40 +00:00
GitHub Actions cab360f656 chore: bump Claude Code to 2.1.289 and Agent SDK to 0.3.289 2026-10-03 23:10:04 +00:00
GitHub Actions ed670b4cf9 chore: bump Claude Code to 2.1.288 and Agent SDK to 0.3.288 2026-10-02 20:23:19 +00:00
GitHub Actions 97c5347339 chore: bump Claude Code to 2.1.287 and Agent SDK to 0.3.287 2026-10-01 18:01:56 +00:00
GitHub Actions 12dd8d74c7 chore: bump Claude Code to 2.1.286 and Agent SDK to 0.3.286 2026-09-30 19:10:48 +00:00
a8cb0db4d8 ci: security hardening for GitHub Actions workflows that call Claude (#1867)
* Run workflows that call Claude on the egress-firewall runner in auto permission mode

Move every workflow job that signs in to the Claude API from
ubuntu-latest to GitHub's egress-firewall runner (ubuntu-24.04-firewall),
and pass --permission-mode auto to the Claude Code action in each of
them except claude.yml.

In auto permission mode Claude Code reviews each tool call that needs
permission and that the allowed tools do not cover, and runs it only if
Claude Code's safety review passes it. Until now these runs, which have nobody to
ask, refused every such call. The issue triage step, which anyone can
start by opening an issue, also passes a --disallowedTools list for
tools it never needs.

claude.yml answers @claude mentions, and for those the action sets
--permission-mode acceptEdits itself. Its permission mode is unchanged.

Allowed tools, settings, models, triggers and permissions are unchanged.
The network allow list for the firewall follows in the next change.

* Add the network allow list, a security check and CLAUDE.md guidance

- .github/egress-firewall.yaml: the hosts that jobs on the
  egress-firewall runner may reach, in enforce mode, each with what
  uses it.
- .github/workflows/workflow-hardening.yml and
  .github/scripts/check_workflow_hardening.py: a check that fails when a
  job that calls Claude is not on the egress-firewall runner, does not
  pass --permission-mode auto, or when the allow list is missing, empty, not in
  enforce mode or names a host with '*'.
  claude.yml is listed as exempt from the permission mode rule, with
  the reason.
- CLAUDE.md: a "Security hardening for GitHub Actions" section so that
  new and edited workflows keep these protections.

* Pin the model for issue triage

The step named no model, so it ran on Claude Code's default model. That
default changed with Claude Code 2.1.280 to a model that is not available
to this CI account. Pass --model claude-opus-5, the model this repository's test-*.yml
workflows (ANTHROPIC_MODEL) and the Python SDK's example runs are
pinned to for the same reason.

This is independent of the security hardening in the two changes before
it.

* Check settings files and --settings for a permission mode

The security check looked for defaultMode only in an inline 'settings'
input. Also read the file a 'settings' input or a --settings flag in
claude_args names, when it is inside the repository, and fail if it sets
a permission mode.

* Flag defaultMode anywhere in a job that calls Claude

A settings file that an earlier step writes at run time can't be read by
the step-level check, so also fail when the job's definition mentions
defaultMode at all.

* Fail the check when settings name a file outside the repository

The check can't read such a file, so it can't tell whether it sets a
permission mode.

* Run @claude mentions in auto permission mode

The action sets --permission-mode acceptEdits for @claude mentions and
appends the workflow's claude_args after it, so the --permission-mode auto
in claude.yml wins. Drop claude.yml from EXEMPT_FROM_AUTO_MODE.

The check now also fails when a step in auto mode names a model older than
claude-opus-4-6: Claude Code does not run auto mode on those and falls back
to its default permission mode.

---------

Co-authored-by: Claude <[email protected]>
2026-09-30 06:33:49 +00:00
GitHub Actions fd1c128679 chore: bump Claude Code to 2.1.285 and Agent SDK to 0.3.285 2026-09-29 19:29:07 +00:00
GitHub Actions 8ce9314fa9 chore: bump Claude Code to 2.1.284 and Agent SDK to 0.3.284 2026-09-28 18:04:13 +00:00
GitHub Actions 756cc22e19 chore: bump Claude Code to 2.1.283 and Agent SDK to 0.3.283 2026-09-25 21:50:53 +00:00
GitHub Actions 9171db3e57 chore: bump Claude Code to 2.1.282 and Agent SDK to 0.3.282 2026-09-24 20:26:46 +00:00
GitHub Actions 8cf3482550 chore: bump Claude Code to 2.1.281 and Agent SDK to 0.3.281 2026-09-23 19:39:47 +00:00
46a42b4323 ci: pin the integration tests to claude-opus-5 (#1853)
The test workflows name no model, so they run on the Claude Code CLI's
default model. Since the bump to Claude Code 2.1.280 that default is Opus 5.5,
which the CI account's API access refuses, so the tests fail and the release
job never runs. Setting ANTHROPIC_MODEL at the workflow level keeps the tests
independent of the CLI's default model, the same fix the Agent SDK for Python
used.

Co-authored-by: Claude Code <[email protected]>
2026-09-23 09:39:39 -07:00
GitHub Actions 9ca9355b36 chore: bump Claude Code to 2.1.280 and Agent SDK to 0.3.280 2026-09-22 16:39:34 +00:00
GitHub Actions cfc3eb22bf chore: bump Claude Code to 2.1.278 and Agent SDK to 0.3.278 2026-09-19 03:11:10 +00:00
GitHub Actions 4036a180cf chore: bump Claude Code to 2.1.277 and Agent SDK to 0.3.277 2026-09-18 18:07:24 +00:00
GitHub Actions a4f54ef2c5 chore: bump Claude Code to 2.1.276 and Agent SDK to 0.3.276 2026-09-18 02:13:13 +00:00
GitHub Actions 2261fcfc88 chore: bump Claude Code to 2.1.275 and Agent SDK to 0.3.275 2026-09-17 22:33:23 +00:00
GitHub Actions 3b8197d3d4 chore: bump Claude Code to 2.1.274 and Agent SDK to 0.3.274 2026-09-17 00:12:40 +00:00
GitHub Actions 7b0b255830 chore: bump Claude Code to 2.1.273 and Agent SDK to 0.3.273 2026-09-15 20:23:44 +00:00
GitHub Actions bf38e86e58 chore: bump Claude Code to 2.1.272 and Agent SDK to 0.3.272 2026-09-15 00:43:27 +00:00
GitHub Actions 51db78a4b8 chore: bump Claude Code to 2.1.271 and Agent SDK to 0.3.271 2026-09-14 22:13:27 +00:00
GitHub Actions 9cdae7f0d9 chore: bump Claude Code to 2.1.270 and Agent SDK to 0.3.270 2026-09-12 19:46:49 +00:00
GitHub Actions 56cf60fde4 chore: bump Claude Code to 2.1.269 and Agent SDK to 0.3.269 2026-09-11 19:18:51 +00:00
GitHub Actions 0a8d3c9443 chore: bump Claude Code to 2.1.268 and Agent SDK to 0.3.268 2026-09-10 20:39:16 +00:00
GitHub Actions 19dda84776 chore: bump Claude Code to 2.1.267 and Agent SDK to 0.3.267 2026-09-09 20:11:01 +00:00
GitHub Actions 5ccc3a35a6 chore: bump Claude Code to 2.1.266 and Agent SDK to 0.3.266 2026-09-08 23:56:12 +00:00
GitHub Actions 0d0e0876d3 chore: bump Claude Code to 2.1.265 and Agent SDK to 0.3.265 2026-09-08 20:38:21 +00:00
GitHub Actions 9c5ddab2e6 chore: bump Claude Code to 2.1.263 and Agent SDK to 0.3.263 2026-09-06 02:55:18 +00:00
GitHub Actions d75b94d5ad chore: bump Claude Code to 2.1.261 and Agent SDK to 0.3.261 2026-09-04 19:58:53 +00:00
GitHub Actions ef8bb1e43b chore: bump Claude Code to 2.1.260 and Agent SDK to 0.3.260 2026-09-03 23:48:49 +00:00
GitHub Actions fa2b2666b7 chore: bump Claude Code to 2.1.259 and Agent SDK to 0.3.259 2026-09-02 22:34:41 +00:00
GitHub Actions 8251c103ac chore: bump Claude Code to 2.1.258 and Agent SDK to 0.3.258 2026-09-01 22:33:40 +00:00
GitHub Actions 781d62e9d5 chore: bump Claude Code to 2.1.257 and Agent SDK to 0.3.257 2026-09-01 17:54:28 +00:00
GitHub Actions 833fb0f8c9 chore: bump Claude Code to 2.1.252 and Agent SDK to 0.3.252 2026-08-31 19:47:59 +00:00
GitHub Actions a874e9ecd7 chore: bump Claude Code to 2.1.251 and Agent SDK to 0.3.251 2026-08-28 18:20:42 +00:00
GitHub Actions a60f3e1db3 chore: bump Claude Code to 2.1.250 and Agent SDK to 0.3.250 2026-08-28 00:50:11 +00:00
GitHub Actions e8c2d7c16c chore: bump Claude Code to 2.1.248 and Agent SDK to 0.3.248 2026-08-27 22:12:44 +00:00
GitHub Actions 70fec18385 chore: bump Claude Code to 2.1.247 and Agent SDK to 0.3.247 2026-08-27 00:09:24 +00:00
GitHub Actions 1f291e1cfe chore: bump Claude Code to 2.1.246 and Agent SDK to 0.3.246 2026-08-25 22:32:51 +00:00
Muhammad Abdullah khanandGitHub 76ac41a83e fix: encode branch names in GitHub links (#1713)
* fix: encode branch names in GitHub links

* style: format branch URL helper
2026-08-25 11:02:49 -07:00
Jeremy SchoemakerandGitHub 8ef9699156 fix: bound download_job_log against a stalled log fetch (#1719)
The download_job_log MCP tool called
client.actions.downloadJobLogsForWorkflowRun() with no timeout and no
AbortController. @octokit/rest@21 runs on Node's native fetch, which has
no default timeout, and Octokit only cancels a request when the caller
passes request.signal. If the log blob fetch stalls, that await never
resolves and never rejects.

This tool is always enabled in tag mode (src/modes/tag/index.ts), so a
"fix the failing CI" run that calls get_ci_status -> get_workflow_run_details
-> download_job_log can hang on this one await with nothing to recover it.
It's headless, so the run only ends when the Actions job-level
timeout-minutes kills it, burning the whole job budget with the tracking
comment stuck at "Claude Code is working...".

Sibling fetch in src/github/utils/image-downloader.ts (fetchImage) already
got this treatment in #1625 via a timeout-driven AbortController. Same
shape of call: fetch a GitHub-hosted resource by ID from untrusted PR/CI
content. This mirrors that fix for github-actions-server.ts.

Extracted the download+write logic into an exported downloadJobLog()
function (with an injectable timeoutMs) so the timeout path is directly
testable, and guarded the module's entrypoint side effects with
import.meta.main, matching the pattern already used by the other
entrypoints in src/entrypoints/.
2026-08-25 08:46:02 -07:00
Yauheni PapovichandGitHub 791545dab1 fix: allow parentheses in valid branch names (#1710)
Accept parentheses while preserving existing branch-name security checks.

Add regression coverage for scoped branch names.

Refs anthropics/claude-code-action#1709
2026-08-25 08:45:50 -07:00
2d7a787fbd fix: use paths in delete_files prompt example (#1702)
* fix: use paths param in delete_files prompt example

The tag-mode prompt told the model to call delete_files with
"files", but the MCP tool schema and handler expect "paths".

Co-authored-by: Cursor <[email protected]>

* test: prove delete_files prompt against the live MCP schema

The old {files} payload is rejected by the same Zod shape the
tool registers; the generated prompt example now parses cleanly.

Co-authored-by: Cursor <[email protected]>

* test: cover delete_files schema edges and the sibling commit_files tool

Confirm the old files-only payload still fails, types and required
fields are enforced, and commit_files was not inverted by the fix.

Co-authored-by: Cursor <[email protected]>

---------

Co-authored-by: RESILIENCE Agentic Solutions <[email protected]>
Co-authored-by: Cursor <[email protected]>
2026-08-25 08:45:33 -07:00
Mohammed AlkindiandGitHub b58c16b325 chore: add .gitattributes to normalize line endings (#1708)
On Windows, git's default core.autocrlf=true checks out CRLF, and the contributor commands in CONTRIBUTING.md then fail: bun run format:check reports 191 files because Prettier defaults to endOfLine "lf", and three tests that match LF-terminated content fail.

All tracked blobs are already LF, so this changes checkout behavior only and produces no renormalization diff. Every CI job runs on ubuntu-latest, where it is a no-op.
2026-08-25 08:45:13 -07:00
GitHub Actions 16b3b310c3 chore: bump Claude Code to 2.1.245 and Agent SDK to 0.3.245 2026-08-25 05:14:05 +00:00
GitHub Actions 6bcfb8263a chore: bump Claude Code to 2.1.241 and Agent SDK to 0.3.241 2026-08-25 03:43:35 +00:00
GitHub Actions b62c7454dc chore: bump Claude Code to 2.1.241 and Agent SDK to 0.3.243 2026-08-25 03:38:24 +00:00
GitHub Actions e5ad3c7725 chore: bump Claude Code to 2.1.243 and Agent SDK to 0.3.243 2026-08-24 23:41:35 +00:00
GitHub Actions c81e3bc69d chore: bump Claude Code to 2.1.241 and Agent SDK to 0.3.241 2026-08-23 00:53:06 +00:00
GitHub Actions 24dcd50c05 chore: bump Claude Code to 2.1.240 and Agent SDK to 0.3.240 2026-08-22 14:45:56 +00:00
GitHub Actions dcb57747bf chore: bump Claude Code to 2.1.239 and Agent SDK to 0.3.239 2026-08-21 19:55:45 +00:00
492d2d78ee fix: teach claude_args --allowedTools in the signed prompt (#1704)
allowed_tools was removed in v1.0. The tag-mode prompt still named it
as the way to enable Bash under commit signing.

Co-authored-by: RESILIENCE Agentic Solutions <[email protected]>
Co-authored-by: Cursor <[email protected]>
2026-08-20 16:17:53 -07:00
2ca5fb4027 fix: surface resolved model limits (#1608)
Signed-off-by: ulofiai <[email protected]>
Co-authored-by: ulofiai <[email protected]>
2026-08-20 16:17:36 -07:00
f3f2789f0a fix(mcp): recognize mcp__github aggregate selector for GitHub MCP server initialization (#1657)
* fix(mcp): accept shorthand selectors for GitHub MCP server initialization

## Problem

Signed-off-by: anish <[email protected]>

* address review feedback: fix prettier formatting

Signed-off-by: anish <[email protected]>

---------

Signed-off-by: anish <[email protected]>
Co-authored-by: anish <[email protected]>
2026-08-20 16:17:25 -07:00
Gautam SharmaandGitHub 6a5f1d8e0a fix(cleanup): keep the base-branch config revert out of the auto-commit (#1677)
restoreConfigFromBase replaces .claude/, CLAUDE.md and the other sensitive
paths with the PR base branch's versions, then deliberately unstages them so
the revert does not reach a commit. checkAndCommitOrDeleteBranch then ran a
bare `git add -A`, which staged them again and pushed a silent revert of the
PR author's own config onto their branch, under a commit message that says
only "Auto-commit: Save uncommitted changes from Claude".

restoreConfigFromBase now returns the paths it restored. run.ts threads them
through updateCommentLink into checkAndCommitOrDeleteBranch, which excludes
them via pathspec from both the staging and the git status check.

The exclusion is driven by what was actually restored rather than applied
unconditionally. This path also runs for issues, where no restore happens and
Claude may legitimately have been asked to edit CLAUDE.md or
.claude/settings.json; excluding those there would silently drop the work —
trading one silent-data-loss bug for another. Reverting the fix, dropping the
status scoping, and switching to an unconditional exclusion each fail the new
tests.

The status check is scoped the same way as the staging: when the reverted
config is the only dirty entry there is no real work, so the branch is now
correctly treated as empty and deleted instead of receiving a pure revert.

Reachable on a closed or merged PR where Claude left uncommitted changes with
use_commit_signing false — the only combination where a restore has run and
claudeBranch is set.

Fixes #1669
2026-08-20 16:17:03 -07:00
HyunSooandGitHub 39ad3c8977 fix(github): honor GITHUB_GRAPHQL_URL for the GraphQL client (#1575)
The REST client honors GITHUB_API_URL, but the GraphQL client derived its
base URL from GITHUB_API_URL as well and ignored the standard
GITHUB_GRAPHQL_URL variable that GitHub Actions provides. On standard GitHub
Enterprise Server this still worked because @octokit/graphql rewrites a
".../api/v3" REST base to ".../api/graphql", but any deployment whose GraphQL
endpoint is not derivable from the REST base (custom proxy, separate host)
sent GraphQL requests to the wrong URL.

Honor GITHUB_GRAPHQL_URL independently and fall back to GITHUB_API_URL when it
is unset, so behavior is unchanged for github.com and standard GHES. A single
trailing "/graphql" is stripped because @octokit/graphql appends its own.

Add wire-level regression tests that run the real client factory in a fresh
process and assert the final request URLs and Authorization headers;
constructor-option assertions are insufficient because @octokit/graphql
rewrites the path after the client is constructed.
2026-08-20 15:45:21 -07:00
GitHub Actions 3f854a8fb5 chore: bump Claude Code to 2.1.238 and Agent SDK to 0.3.238 2026-08-20 20:33:55 +00:00
GitHub Actions 5ee796a55f chore: bump Claude Code to 2.1.237 and Agent SDK to 0.3.237 2026-08-20 00:54:23 +00:00
Tem RevilandGitHub cff8d3c8f0 fix(git-config): neutralize checkout credential in include-based config (#1526)
configureGitAuth() removed the actions/checkout auth header with
`git config --unset-all http.<server>/.extraheader`, which only edits the
repo-local config. Since actions/checkout v6.0.0 (backported to v5.0.1 and
v4.3.1) the header is written to a separate file under RUNNER_TEMP and
pulled in via include.path, so --unset-all on the local config is a no-op:
the code logged "No existing authentication headers to remove" while the
checkout credential (usually the workflow GITHUB_TOKEN) stayed usable by
git for the rest of the job.

Also clear the header from every included file, so it is neutralized under
both the pre-v6 (local) and v6+ (include) layouts. Includes that do not
define the header are left untouched.

Fixes #1510
2026-08-19 17:23:23 -07:00
GitHub Actions e2a4b761cd chore: bump Claude Code to 2.1.236 and Agent SDK to 0.3.236 2026-08-19 20:05:23 +00:00
65b50df083 fix(github): match bot actors in comment filters using GraphQL __typename (#1616)
`exclude_comments_by_actor` and `include_comments_by_actor` never matched
any bot. Both the documented `*[bot]` wildcard and exact entries such as
`dependabot[bot]` silently did nothing.

GitHub's GraphQL API returns the bare login for App actors ("dependabot"),
while REST and the GitHub UI append a suffix ("dependabot[bot]"). Filter
patterns are written in the suffixed form, so matching a GraphQL login
against them could never succeed and `actor.endsWith("[bot]")` was dead
code.

Request `__typename` on the Actor-typed author selections and normalize
App actors to their suffixed name via `resolveActorName()` before matching.
Normalizing at the filter boundary fixes the wildcard and exact-match cases
together, and leaves the author names shown in the prompt unchanged.

The existing test mocked `login: "scanner[bot]"`, a payload GraphQL never
produces, which is why the gap was invisible. It now mocks the real shape
(`__typename: "Bot", login: "scanner"`) and fails without this fix.

The commit author selection is left alone: it is a GitCommit, not an Actor.

Fixes #1514

Co-authored-by: Claude Opus 5 <[email protected]>
2026-08-18 17:24:26 -07:00
JuwanandGitHub 0a80d21df7 fix: strip unused ALL_INPUTS environment variable from Claude subprocess env (#1692) 2026-08-18 17:23:38 -07:00
JuwanandGitHub 54eadc2f72 fix(security): unify secret redaction in public comment outputs (#1693)
Ensure all public issue, PR, and inline comments apply redactSecrets() in addition to sanitizeContent() before submitting payloads to the GitHub API. This aligns public comment output with error log and step-summary redaction policies, preventing potential leakage of Anthropic API keys, AWS credentials, Slack tokens, JWTs, and GitHub tokens.
2026-08-18 17:23:11 -07:00
GitHub Actions d40ddef4c0 chore: bump Claude Code to 2.1.235 and Agent SDK to 0.3.235 2026-08-18 20:39:22 +00:00
GitHub Actions 459ad358ae chore: bump Claude Code to 2.1.234 and Agent SDK to 0.3.234 2026-08-17 20:23:54 +00:00
Abhinav Kumar SinghandGitHub d721746d68 fix: bound image attachment downloads (#1625) 2026-08-14 16:47:46 -07:00
5da4c76dde fix: bump shell-quote to 1.8.4 to remediate CVE-2026-9277 (#1557)
shell-quote 1.8.3 (root and base-action dependency) is affected by
CVE-2026-9277, a CRITICAL severity vulnerability. 1.8.4 contains the fix.

Co-authored-by: Ashwin Bhat <[email protected]>
2026-08-14 16:46:34 -07:00
Ashwin BhatandGitHub a2cac87e27 ci: skip Claude-backed test jobs on fork PRs (#1655)
* ci: skip Claude-backed test jobs on fork PRs

Jobs that run the action against the Claude API authenticate via workload
identity federation, which fork PRs cannot mint an OIDC token for, so they
always failed on external contributions. Gate each such job on the PR head
repo matching the base repo; push and workflow_dispatch runs are unaffected.

No-Verification-Needed: CI workflow config only, exercised by Actions on the PR

* test: pin the bare remote's initial branch in fetch-depth test

The shallow-clone case created its bare remote with a plain git init, so
HEAD pointed at whatever init.defaultBranch resolves to (master on CI)
while the test only pushed main. git clone --depth=1 implies
--single-branch, and with a dangling remote HEAD it produces an empty,
non-shallow clone, so the is-shallow assertion failed on runners whose
default branch is not main.

No-Verification-Needed: test-only change
2026-08-14 16:40:34 -07:00
Rishav NaskarandGitHub b49813d0e7 feat(context): include diffHunk in PR review comment context (#1584)
* feat(context): include diffHunk in PR review comment context

Review comments arrived with only path and line, so the code they were
written against was missing from the prompt. Fetch diffHunk in the PR
GraphQL query and render it under the comment as a diff block.

The hunk is PR-authored content, so it goes through sanitizeContent like
the comment body. Comments without a hunk are unchanged.

Fixes #855

* test(formatter): cover outdated review comments with an empty diff hunk

GitHub returns diffHunk: "" (not null) for comments whose line no longer
exists in the diff, so the render guard has to reject empty strings too.
Found running the real query against anthropics/claude-code-action#1025.
2026-08-14 16:32:44 -07:00
9678fce999 fix(base-action): add ~/.local/bin to $GITHUB_PATH after auto-install (#1643)
## Problem

Signed-off-by: anish <[email protected]>
Co-authored-by: anish <[email protected]>
2026-08-14 16:31:36 -07:00
Madan kumarandGitHub ed186becce fix: only limit fetch depth when the checkout is already shallow (#1647)
restoreConfigFromBase and setupBranch pass --depth to every git fetch. On a
checkout made with fetch-depth: 0 that does not just cap the download: it
truncates the history already present and marks the repository shallow, which
drops the merge base with the base branch. `git log origin/<base>..HEAD` then
silently includes commits that are already merged, and
`git diff origin/<base>...HEAD` fails with "no merge base" — the two commands
the prompt tells Claude to run to scope its work to the PR.

Gate the flag on `git rev-parse --is-shallow-repository`, so a checkout that is
already shallow (the fetch-depth: 1 default) keeps the same depth behaviour and
the fetch savings it was added for, while a full checkout stays full.

Fixes #1642
2026-08-14 16:31:26 -07:00
05ee4b30d7 Harden delete_files MCP tool: validate paths within repo root (#1636)
Mirror the path validation already performed by the commit_files tool.
delete_files previously only normalized absolute paths against CWD and
passed relative paths through unchecked; it now runs each path through
validatePathWithinRepo, rejecting "../" traversal and symlinked escapes
for consistency and defense-in-depth.

Co-authored-by: Claude Opus 4.8 <[email protected]>
2026-08-14 16:31:15 -07:00
GitHub Actions 9d7150bc8a chore: bump Claude Code to 2.1.233 and Agent SDK to 0.3.233 2026-08-14 22:21:44 +00:00
GitHub Actions e63208cb98 chore: bump Claude Code to 2.1.232 and Agent SDK to 0.3.232 2026-08-13 23:30:40 +00:00
GitHub Actions dc33e8a15b chore: bump Claude Code to 2.1.231 and Agent SDK to 0.3.231 2026-08-13 08:39:40 +00:00
GitHub Actions c58ad32088 chore: bump Claude Code to 2.1.229 and Agent SDK to 0.3.229 2026-08-12 20:57:42 +00:00
Henrique PiresandGitHub dfb8fc798e fix(mcp): detect binary files by content instead of extension allowlist (#1633) 2026-08-11 16:35:46 -07:00
a2489efcb9 fix(summary): keep every text block in structured tool results (#1619)
formatResultContent recognized structured tool output shaped like
`[{ type: "text", text: "..." }]` but read only `parsedContent[0].text`.
When a tool result split its output across several text blocks, the step
summary showed the first and silently dropped the rest, so extra findings,
file paths and follow-up instructions vanished from the rendered
Claude Code Report while remaining in the execution transcript.

Collect the text from every block instead of just the first. Blocks of other
types, such as images, are skipped rather than stringified into the summary.

Fixes #1572

Co-authored-by: Claude Opus 5 <[email protected]>
2026-08-11 16:35:38 -07:00
Abhinav Kumar SinghandGitHub 8b8745859f fix: paginate GitHub Actions MCP responses (#1629) 2026-08-11 16:35:34 -07:00
GitHub Actions 239e3a7308 chore: bump Claude Code to 2.1.228 and Agent SDK to 0.3.228 2026-08-11 20:03:37 +00:00
GitHub Actions 5ef2e550a4 chore: bump Claude Code to 2.1.227 and Agent SDK to 0.3.227 2026-08-10 23:16:28 +00:00
GitHub Actions 6b082c4193 chore: bump Claude Code to 2.1.226 and Agent SDK to 0.3.226 2026-08-08 02:48:25 +00:00
GitHub Actions 7ff6806c8e chore: bump Claude Code to 2.1.225 and Agent SDK to 0.3.225 2026-08-08 01:14:59 +00:00
NickNojiriandGitHub 751e003832 fix(branch): collapse empty path segments in branch_name_template (#1539)
A branch_name_template that places {{description}} (or another variable)
next to a slash crashes the run when the variable resolves to an empty
string. An issue/PR title with no ASCII-alphanumeric content — emoji-only,
CJK-only, or punctuation-only — makes extractDescription() return "", so a
template like "{{prefix}}{{description}}/{{entityNumber}}" produces
"claude//123". validateBranchName rejects consecutive (and leading/trailing)
slashes, and the thrown error propagates uncaught out of setupBranch,
aborting the entire run.

Normalize the templated result before the empty-result check: collapse runs
of slashes and drop any leading/trailing slash. Single-slash and dash
separators are untouched, so existing template behavior is unchanged; a
template that collapses to empty still falls back to the default format.

This is distinct from the {{label}} sanitization tracked in #1491 (and its
open PRs), which deliberately leave {{description}} alone — so this path
remained broken. Fixes the whole empty-segment class regardless of variable.

Adds regression tests for emoji-only and CJK-only titles, a trailing empty
segment, and a direct validateBranchName assertion proving the run no longer
aborts.
2026-08-07 07:59:52 -07:00
Rishav NaskarandGitHub b704dd3960 fix(branch): validate generated branch name under commit signing (#1582)
The non-signing path validated newBranch before checkout, but the
use_commit_signing path passed it straight to the file ops server, so an
invalid branch_name_template surfaced only as a 422 "Reference name is
not valid" on the first commit.

Validate once after the name is resolved so both paths fail early with
the same message.

Fixes #1573
2026-08-07 07:59:43 -07:00
Minh VuandGitHub ecf573bd65 fix: expose conclusion output (#1549) 2026-08-07 07:58:52 -07:00
Minh VuandGitHub 7764306e92 fix: stop retrying deterministic ref updates (#1551) 2026-08-07 07:58:19 -07:00
Minh VuandGitHub 5dd098c551 ci: pass allowed tools through claude args (#1552) 2026-08-07 07:57:29 -07:00
4c4309a064 fix(cache): disable setup-bun cache to avoid 5-retry HTML-error burn (#1580)
The upstream oven-sh/setup-bun action saves with a deterministic key
(Bun version) that isn't ref-aware. On every second-and-subsequent run
against the same PR ref, the GitHub cache API rejects the duplicate
key+ref with a 409 (HTML body), and @actions/cache treats the unparsable
response as transient and burns ~20-30s on 5 retries before warning.

The 35 MB Bun binary downloads in 2-3s, so disabling the cache is a net
wallclock win and removes the noisy warning that fires on every PR push
after the first.

Closes #1252

Co-authored-by: Mukunda Rao Katta <[email protected]>
2026-08-07 07:57:18 -07:00
Minh VuandGitHub c4190dbd78 docs: update base action inputs (#1550) 2026-08-07 07:56:51 -07:00
Jeremy SchoemakerandGitHub 9a2db97708 docs: fix broken Bedrock anchor in cloud-providers.md (#1579)
The link pointed at `#for-aws-bedrock:`, which does not exist on the
github-actions docs page, so it landed readers at the top of a long page
instead of the Bedrock section.

The current heading id is `#using-with-amazon-bedrock-and-google-cloud`.
2026-08-07 07:56:43 -07:00
Sahil GuptaandGitHub 2df67d2c33 fix: match label_trigger case-insensitively (#1576)
label_trigger used a case-sensitive exact comparison, so a workflow
configured with label_trigger: "claude-task" did not fire when an issue
received a label named "Claude-Task" (the same label name with different
casing).

GitHub label names are unique without regard to case, so comparing without
case is unambiguous. It also matches the trigger_phrase check in the same
function, which is already case-insensitive.

Compare labelName and labelTrigger with toLowerCase(), and add a test
covering a mixed-case label.

Fixes #1571
2026-08-07 07:55:32 -07:00
Takaki SatoandGitHub d573b167d3 fix: support labeled action for pull_request events in track_progress (#1586)
Adds "labeled" to the valid pull_request actions for track_progress,
mirroring the existing support for issue events. Previously, adding
a label to a PR (e.g. to trigger a label-driven Claude review) would
fail validation even though the same pattern works for issues.

Fixes #1585
2026-08-07 07:55:20 -07:00
leepokaiandGitHub 0a5f191964 fix: handle null files field from GraphQL on very large PRs (#1593)
GitHub's GraphQL API returns files: null (with no errors entry, and
changedFiles misreported as 0) when a PR's diff is too large to compute.
The unguarded pullRequest.files.nodes dereference in the fetcher crashed
the action with 'TypeError: null is not an object', and the formatter had
the same latent crash on prData.files.nodes.length.

Widen the GitHubPullRequest type to files | null so the compiler enforces
guards, degrade gracefully in the fetcher with a warning, and render the
file count as unavailable (not '0 files') in the formatter.

Fixes #1587
2026-08-07 07:55:10 -07:00
ulofiaiandGitHub 6ef6450f51 fix: enforce max turns from claude args (#1607) 2026-08-07 07:55:06 -07:00
GitHub Actions 1623c36729 chore: bump Claude Code to 2.1.224 and Agent SDK to 0.3.224 2026-08-07 04:02:00 +00:00
Ashwin BhatandGitHub 96e281f4d9 Run checkout auth cleanup when API commit signing is enabled (#1597)
* Run checkout auth cleanup when API commit signing is enabled

* Derive git-config test expectations from GITHUB_SERVER_URL

No-Verification-Needed: test-only change
2026-08-06 10:18:34 -07:00
Ashwin BhatandGitHub e1fc925862 Scope the config snapshot to files inside the working tree (#1596)
* Scope config snapshot to files inside the working tree

* Record excluded snapshot entries as placeholders instead of links

* Limit linked snapshot content to unmodified tracked files and tracked directories

File targets reached through a link are included only when their content is
unchanged from HEAD, and directory targets only when they contain tracked
files; anything else is recorded as a single placeholder. Adds tests for a
sensitive path that links to a tracked directory, links to untracked
directories, and links to tracked files modified after checkout.
2026-08-06 10:17:40 -07:00
Ashwin BhatandGitHub 0aee57ab82 Redact common credential patterns from published run output (#1595)
* Redact common credential patterns from published run output

* Handle color codes and escape sequences ahead of redacted values

Vendor-prefixed formats no longer require a leading word boundary, so a
value that follows an ANSI SGR terminator or a serialized JSON escape is
still matched. AWS key ids keep a boundary but also accept those cases.
sanitizeContent goes back to GitHub-only redaction for inbound content,
and the failure annotation is redacted like the tracking comment.

* Coerce non-string text content before redacting tool results

No-Verification-Needed: one-line coercion in a formatting helper plus regression test
2026-08-06 10:17:25 -07:00
GitHub Actions c038e4dcde chore: bump Claude Code to 2.1.223 and Agent SDK to 0.3.223 2026-08-06 00:53:11 +00:00
Ashwin BhatandGitHub 4c04887769 Invoke the formatter directly from the format hook (#1594)
* Invoke the formatter directly from the format hook

The PostToolUse format hook now runs prettier directly with a pinned
version and --no-config instead of going through the package.json
"format" script, so the hook resolves the same way regardless of the
scripts and formatter config in the checked-out tree. Output matches
the previous "bun run format" (both .prettierrc files are empty).

Also documents which paths the action restores from the PR base branch
and recommends keeping base-branch hooks self-contained.

No-Verification-Needed: config, comment, and doc-only change

* Qualify the self-contained hook guidance for Bun-only runners

Note in docs/security.md and the restore-config JSDoc that bunx runs the
tool under node when node is on PATH, but on a Bun-only runner Bun runs
the script itself and reads bunfig.toml (preload etc.) from the
checkout, so that file and .npmrc are runtime config from the PR head.

No-Verification-Needed: comment- and doc-only change

* Exclude .claude-pr from prettier

No-Verification-Needed: prettierignore-only change
2026-08-05 15:38:47 -07:00
GitHub Actions 9db594c7a0 chore: bump Claude Code to 2.1.222 and Agent SDK to 0.3.222 2026-08-04 22:40:25 +00:00
Ashwin BhatandGitHub acb0385805 Check collaborator permissions for workflow_run events (#1590)
The write-permission gate previously only ran for issue/PR entity
events. Apply it to workflow_run events as well, checking both the
workflow actor and the actor recorded on the upstream run when they
differ. allowed_non_write_users and the github_token override behave
the same as for entity events. Document the behavior for workflow_run
pipelines.
2026-08-04 10:05:34 -07:00
Ashwin BhatandGitHub b80a0f042f Match downloaded images to their source URLs by asset identifier (#1588)
* Match downloaded images to source URLs by asset identifier

* Derive the signed URL asset identifier from the parsed path
2026-08-04 10:04:47 -07:00
Ashwin BhatandGitHub 6fb6bb6858 Pin bun config for MCP server processes (#1589) 2026-08-04 10:03:43 -07:00
Ashwin BhatandGitHub b2963b9127 Derive trigger timestamps for issues and pull_request events (#1592)
* Derive trigger timestamps for issues and pull_request events

For issues labeled/assigned triggers, look up the matching event in the
issue's event history to get the exact time of the label/assignment,
falling back to the payload's updated_at/created_at when the lookup
fails. issues opened uses issue.created_at; pull_request opened uses
pull_request.created_at and other pull_request actions use updated_at.

* Ignore issue label/assign events older than the payload snapshot

A matching labeled/assigned event that predates the webhook payload's
issue.updated_at cannot be the event that fired the webhook, so fall
back to the payload timestamps instead of adopting it as the boundary.
2026-08-03 19:15:50 -07:00
GitHub Actions 86180fa9e4 chore: bump Claude Code to 2.1.221 and Agent SDK to 0.3.221 2026-08-04 00:15:10 +00:00
GitHub Actions be7b93b190 chore: bump Claude Code to 2.1.220 and Agent SDK to 0.3.220 2026-07-25 01:36:28 +00:00
GitHub Actions e0cf66d1d2 chore: bump Claude Code to 2.1.219 and Agent SDK to 0.3.219 2026-07-24 17:14:37 +00:00
GitHub Actions 44423bdec7 chore: bump Claude Code to 2.1.218 and Agent SDK to 0.3.218 2026-07-22 21:27:29 +00:00
KeisukeYamashitaandGitHub b00a3414fd fix: share one exchanged WIF credential across spawned Claude processes (#1407)
* fix: share one exchanged WIF credential across spawned Claude processes

GitHub OIDC tokens are single-use at the Anthropic token-exchange
endpoint (the same jti cannot be exchanged twice). With plugins
configured, the action spawns several short-lived claude processes
(plugin marketplace add, one plugin install per plugin, then the main
query). Each resolved federation from bare env vars and exchanged the
same identity-token file independently: the first exchange succeeded
and every later process got 401 (jti_reused), which the main query
retried for ~3 minutes before failing the job.

The SDK only enables its on-disk credentials cache when federation is
loaded from a profile config file, not from bare env vars. Write a
profile pointing at the identity-token file and select it via
ANTHROPIC_CONFIG_DIR / ANTHROPIC_PROFILE so the first process exchanges
once and the rest reuse the cached access token. The env vars are kept
as a fallback for CLIs that predate profile support.

* fix: scope the WIF credential cache per federation config

Address review feedback on the shared-credentials-cache fix:

- Embed a fingerprint of the federation inputs (rule, org, service
  account, workspace, base URL, scope) in the config dir name. The SDK
  cache reuses a token on expires_at alone and RUNNER_TEMP is per-job,
  so a later step with different federation inputs would silently reuse
  the first step's token. service_account_id and scope are included
  beyond the reviewed list because both are sent in the exchange
  request body and change which credential is minted.
- Skip the action-managed profile with a warning when the operator has
  already set ANTHROPIC_CONFIG_DIR or ANTHROPIC_PROFILE.
- Shrink the profile to the minimal file-backed form; the CLI's bundled
  SDK gap-fills the federation fields from the env vars the action
  already exports (verified against the pinned 2.1.173 binary).
- Remove the token dir in stop() so the identity token and the cached
  exchanged credential don't outlive the step.
- Document that cache sharing relies on the plugin subprocesses
  spawning sequentially.
2026-07-22 07:02:41 -07:00
GitHub Actions fa7e2f0a29 chore: bump Claude Code to 2.1.217 and Agent SDK to 0.3.217 2026-07-21 21:35:48 +00:00
GitHub Actions b76a0776ae chore: bump Claude Code to 2.1.216 and Agent SDK to 0.3.216 2026-07-20 22:14:27 +00:00
GitHub Actions af0559ee4f chore: bump Claude Code to 2.1.215 and Agent SDK to 0.3.215 2026-07-19 02:56:28 +00:00
GitHub Actions 3553f84341 chore: bump Claude Code to 2.1.214 and Agent SDK to 0.3.214 2026-07-18 01:20:51 +00:00
GitHub Actions 700e7f8316 chore: bump Claude Code to 2.1.212 and Agent SDK to 0.3.212 2026-07-17 00:27:04 +00:00
PaarthandGitHub 3e807ec379 fix: handle null comment/review author from deleted accounts (#1490)
GitHub's GraphQL author field is null when the account behind a
comment, review, PR, or issue has been deleted (the ghost user). The
action typed author as non-null and read author.login directly, so a
single comment from a deleted account threw and was swallowed into a
generic 'Failed to fetch PR/issue data', failing the entire run.

Make author nullable on the four affected types and fall back to
'ghost' at each login read. With the type nullable, tsc flags every
dereference, so all sites are covered.
2026-07-15 21:00:22 -07:00
2988cbe14a docs: map custom_instructions to --append-system-prompt (#1480) (#1484)
* docs: map custom_instructions to --append-system-prompt (#1480)

The v1 migration guide mapped the v0 `custom_instructions` input to
`claude_args: --system-prompt`, but these have different semantics:
`custom_instructions` *appended* to Claude Code's default system prompt,
while `--system-prompt` *replaces* it entirely. Users who followed the
guide silently lost the whole built-in system prompt (tool-usage guidance,
sub-agent conventions, etc.), keeping only their few custom lines.

Fixes #1480:
- Map `custom_instructions` -> `--append-system-prompt` (matches v0 append
  semantics) in the deprecated-inputs table, the migration example, and the
  checklist.
- Correct the claude_args options table: `--system-prompt` replaces the
  entire prompt; add an `--append-system-prompt` row for append behavior.

Docs-only; no code changes.

Co-Authored-By: Claude Opus 4.8 <[email protected]>

* docs: fix remaining custom_instructions migration references (#1480)

Update usage.md, faq.md, and configuration.md to map custom_instructions
to --append-system-prompt, matching the migration-guide fix. The
override_prompt row is left unchanged since replacement semantics may be
intended there.

Co-Authored-By: Claude Fable 5 <[email protected]>

---------

Co-authored-by: Claude Opus 4.8 <[email protected]>
2026-07-15 20:35:32 -07:00
a1c0599a9c fix(format): filter out thinking_tokens system messages from step summary (#1479)
## Summary

Signed-off-by: anish <[email protected]>
Co-authored-by: anish <[email protected]>
2026-07-15 20:27:23 -07:00
5bfa96a5b0 fix: allow leading underscore in branch names (valid per git-check-ref-format) (#1486)
Branch names starting with an underscore (e.g. _release/v1.2.3) are valid
per git check-ref-format but were rejected by validateBranchName's
first-character whitelist. Since setupBranch validates a PR's baseRefName
after checkout, the action failed on every open PR targeting such a
branch. A leading underscore carries no option-injection risk (only a
leading dash does, which is still rejected separately).

Co-authored-by: Claude Fable 5 <[email protected]>
2026-07-15 20:27:12 -07:00
Jianke LINandGitHub 214a70611b fix: map claude_args model to SDK options (#1474) 2026-07-15 20:22:15 -07:00
5f509a1c1f fix(sanitizer): strip alt text from reference-style markdown images (#1488)
stripMarkdownImageAltText removed alt text from inline images
(![alt](url)) but not reference-style images (![alt][ref]), because the
regex requires the "](" of the inline form. Alt text is a
hidden-instruction channel that reaches the prompt via sanitizeContent,
so the reference-style form let it survive.

Add a matching replace for the reference-style form (![alt][ref] ->
![][ref]), preserving the [ref] label so the image definition still
resolves. Adds regression tests.

Co-authored-by: Contributor <[email protected]>
Co-authored-by: Claude Opus 4.8 (1M context) <[email protected]>
2026-07-15 20:21:12 -07:00
PaarthandGitHub e64308ff97 fix: sanitize {{label}} in branch name templates (#1492)
A scoped label like area:permissions was substituted into the branch
name verbatim, producing a ":" that validateBranchName rejects. Because
the branch setup block catches that error and calls process.exit(1), the
whole run died. {{description}} was already sanitized via extractDescription;
{{label}} was the only free-text variable that skipped it.

Add a sanitizeLabel helper (replaces invalid-char runs with a hyphen so
scoped labels stay readable) and apply it before substitution, falling back
to entityType when a label sanitizes to empty. Adds regression tests that
also assert the result passes validateBranchName.
2026-07-15 20:21:02 -07:00
58dc33d9ad test: cover prepareContext validation error branches (#1460)
* test: cover prepareContext validation error branches

create-prompt.test.ts exercised only happy paths; the ~20 validation
guards in prepareContext (missing PR_NUMBER, unsupported event type,
unsupported issue action, missing claude branch, etc.) had no coverage.

Adds a "prepareContext validation errors" block asserting the thrown
messages for the reachable guards, using the existing createMockContext
helper.

Co-Authored-By: Claude Opus 4.8 <[email protected]>

* test: cover comments/common link and body builders

`src/github/operations/comments/common.ts` had no direct test coverage,
though its exports are live code used by create-initial.ts and
update-with-branch.ts. This adds unit tests for all four exports:
SPINNER_HTML, createJobRunLink, createBranchLink, and createCommentBody.

Assertions are built from the imported GITHUB_SERVER_URL so they hold on
GHES as well as github.com. Pure test additions — no production changes.

Co-Authored-By: Claude Opus 4.8 <[email protected]>

---------

Co-authored-by: Claude Opus 4.8 <[email protected]>
2026-07-15 20:20:53 -07:00
evol1228andGitHub ae650f8355 docs: fix dead example links in custom-automations.md (#1513) 2026-07-15 20:18:08 -07:00
GitHub Actions 1298632ce7 chore: bump Claude Code to 2.1.211 and Agent SDK to 0.3.211 2026-07-15 23:10:52 +00:00
GitHub Actions 1253134445 chore: bump Claude Code to 2.1.210 and Agent SDK to 0.3.210 2026-07-14 23:46:09 +00:00
NickNojiriandGitHub 4f07c81564 fix(sanitizer): redact GitHub user-to-server (ghu_) tokens (#1502)
redactGitHubTokens covers ghp_, gho_, ghs_, ghr_, and github_pat_
tokens but misses ghu_ (GitHub App user-to-server tokens), one of the
documented GitHub token prefixes. A ghu_ token appearing in issue or
PR content passed through sanitization unredacted.

Add the ghu_ pattern, mirroring the existing 40-character token
patterns, with unit tests including the git-credential URL form.
2026-07-14 13:06:18 -07:00
GitHub Actions f1bd27ca5b chore: bump Claude Code to 2.1.209 and Agent SDK to 0.3.209 2026-07-14 06:36:42 +00:00
GitHub Actions a08f8913d5 chore: bump Claude Code to 2.1.208 and Agent SDK to 0.3.208 2026-07-14 01:11:18 +00:00
972a512078 fix(sdk): fail step when result has is_error:true despite success subtype (#1496)
Treat subtype success with is_error:true as a failed run so CI does not
show a misleading green check when the review never actually ran.

Fixes #1495

Co-authored-by: syf2211 <[email protected]>
2026-07-13 09:01:51 -07:00
GitHub Actions e90deca476 chore: bump Claude Code to 2.1.207 and Agent SDK to 0.3.207 2026-07-11 00:52:42 +00:00
GitHub Actions 536f2c32a3 chore: bump Claude Code to 2.1.206 and Agent SDK to 0.3.206 2026-07-09 23:35:14 +00:00
GitHub Actions 37b464ce72 chore: bump Claude Code to 2.1.205 and Agent SDK to 0.3.205 2026-07-08 21:22:46 +00:00
GitHub Actions ba0aafd430 chore: bump Claude Code to 2.1.204 and Agent SDK to 0.3.204 2026-07-08 00:28:40 +00:00
GitHub Actions 0fe28cdb64 chore: bump Claude Code to 2.1.203 and Agent SDK to 0.3.203 2026-07-07 21:07:00 +00:00
GitHub Actions f87768c6d2 chore: bump Claude Code to 2.1.202 and Agent SDK to 0.3.202 2026-07-06 22:52:16 +00:00
Ashwin BhatandGitHub 58a2944bbc chore: fix prettier formatting (#1463) 2026-07-04 09:31:09 -07:00
Akhilesh AroraandGitHub beb753ed72 fix: propagate curl failures in install pipeline (#1241)
installClaudeCode() pipes `curl -fsSL | bash -s --`. Bash exits with
the status of the last command, so when curl fails (429 rate limit,
403, or connection error) `bash -s` still exits 0 on empty stdin and
the action logs "Claude Code installed successfully". The 3-attempt
retry loop never triggers because the first attempt looks successful,
and the run later dies with "Executable not found in $PATH: claude".

Prefix the pipeline with `set -o pipefail;` so curl's non-zero exit
propagates through the pipe and the retry loop can actually kick in.
Extracted into buildInstallCommand() with regression tests covering
both the old buggy shape and the fixed one.
2026-07-03 22:38:53 -07:00
JerryLeeandGitHub 235b39bf21 fix: preserve repeated add-dir flags in claude args (#1256) 2026-07-03 22:38:30 -07:00
石岳峰andGitHub d060ddc963 fix(restore): handle symlinked CLAUDE.md paths during config snapshot (#1441)
When snapshotting PR-authored sensitive paths into .claude-pr/, cpSync with
dereference:true throws ENOENT if a symlink target is missing on the PR head
(e.g. .claude/CLAUDE.md -> ../AGENTS.md). Fall back to copying the symlink
itself so restoreConfigFromBase can continue and restore trusted base versions.

Fixes #1398
2026-07-03 22:38:23 -07:00
tarunag10andGitHub 0f07aee435 Use modern noreply email for co-author trailers (#1369) 2026-07-03 22:38:02 -07:00
a221ad2dd9 Drop buffered inline comment when it is posted live (#1405) (#1412)
When classify_inline_comments is enabled, create_inline_comment buffers calls
without confirmed=true. The model frequently re-issues the call with confirmed=true
after reading the buffered reply, which posts the comment live but leaves the
original buffered entry behind. The post-session replay step then posts it again,
so every inline comment lands twice.

Reconcile the buffer on a live post: after a confirmed comment is created, remove
any buffered entry matching the same path, line, startLine and body so it cannot be
replayed. Extracts the reconciliation into src/mcp/inline-comment-buffer.ts (the MCP
server module starts a server on import) and adds unit tests.

Co-authored-by: archievi <[email protected]>
2026-07-03 22:37:36 -07:00
GitHub Actions 558b1d6cab chore: bump Claude Code to 2.1.201 and Agent SDK to 0.3.201 2026-07-03 23:51:09 +00:00
GitHub Actions 01872ccc02 chore: bump Claude Code to 2.1.200 and Agent SDK to 0.3.200 2026-07-03 16:53:31 +00:00
GitHub Actions 769e3bdff9 chore: bump Claude Code to 2.1.199 and Agent SDK to 0.3.199 2026-07-02 23:36:20 +00:00
GitHub Actions 6c0083bb72 chore: bump Claude Code to 2.1.198 and Agent SDK to 0.3.198 2026-07-01 20:46:29 +00:00
846d5d8993 Add agent-approval-check composite action (#1429)
* Add agent-approval-check composite action

Require N human approvals on PRs that contain agent-authored commits.
Posts an agent-approval-check commit status that repos mark as a
required check on protected branches.

This is a sanitized port of the check Anthropic runs internally on
every agent-authored PR — same detection rules, /approve <sha>
comment flow, sibling-PR-same-SHA guard, and fail-closed semantics,
with the Anthropic-specific path exemptions and kill-switch removed
and config moved to action inputs.

Co-Authored-By: Claude <[email protected]>

* Drop stray internal acronym from comment

* agent-approval-check: require write-access approvers, pin deps, pagination + doc fixes

🏠 Remote-Dev: homespace

* agent-approval-check: prettier

🏠 Remote-Dev: homespace

* agent-approval-check: verify approver write permission via REST; commits(last:100); docstring

🏠 Remote-Dev: homespace

* agent-approval-check: use headRefOid; drop pull_request_review trigger and correct threat-model docs

🏠 Remote-Dev: homespace

* agent-approval-check: stale-notification wording, no-retry-on-4xx, docstring API-call count

🏠 Remote-Dev: homespace

* agent-approval-check: fail-closed sibling guard on commits-ordering edge; drop stale 'reviewed' from README

🏠 Remote-Dev: homespace

* agent-approval-check: drop hardcoded API-call counts from logs; clarify author write-access requirement in README

🏠 Remote-Dev: homespace

* agent-approval-check: count all agent-email commits (close-reopen bypass); validate REQUIRED_APPROVALS>=1; exempt_head_branches warning

🏠 Remote-Dev: homespace

---------

Co-authored-by: Claude <[email protected]>
Co-authored-by: Octavian Guzu <[email protected]>
2026-06-30 14:09:57 -07:00
GitHub Actions fad22eb3fa chore: bump Claude Code to 2.1.197 and Agent SDK to 0.3.197 2026-06-30 17:57:23 +00:00
GitHub Actions 4633baf526 chore: bump Claude Code to 2.1.196 and Agent SDK to 0.3.196 2026-06-29 23:27:52 +00:00
GitHub Actions a92e7c70a4 chore: bump Claude Code to 2.1.195 and Agent SDK to 0.3.195 2026-06-26 21:30:09 +00:00
tarunag10andGitHub f8076dc008 fix: bound app token revocation cleanup (#1437) 2026-06-25 20:57:44 -07:00
GitHub Actions 5211368122 chore: bump Claude Code to 2.1.193 and Agent SDK to 0.3.193 2026-06-25 21:46:38 +00:00
GitHub Actions 428971d2ec chore: bump Claude Code to 2.1.191 and Agent SDK to 0.3.191 2026-06-24 21:59:09 +00:00
GitHub Actions 74eedf1a18 chore: bump Claude Code to 2.1.190 and Agent SDK to 0.3.190 2026-06-24 15:55:36 +00:00
GitHub Actions 80b3182633 chore: bump Claude Code to 2.1.187 and Agent SDK to 0.3.187 2026-06-23 21:05:06 +00:00
360be9c8fc fix: allow @ in branch names (valid per git-check-ref-format) (#1411)
`validateBranchName` rejects branch names containing `@`, even though
`git check-ref-format` permits `@` and GitHub itself accepts such
branches. PRs whose head or base branch contains an `@` fail validation
in-process before any git operation, so the action errors out
immediately.

Branch names with `@` show up in real workflows: ticket conventions
like "TICKET-123@add-feature" (#998), leading-prefix conventions like
"@hotfix/...", and agent tooling that appends "@<sessionid>" (#1305).
There is no workaround other than renaming the branch, which is often
not under the user's control.

Branch names are never passed through a shell (git calls use
execFileSync argv arrays), so `@` carries no injection risk. This is
the same reasoning used to add `#` in #1167, `+` in #1248, and `,` in
#1310. The bare name "@" (HEAD shorthand in git revision syntax) and
the "@{" reflog sequence are still rejected.

- Add `@` to the validateBranchName whitelist regex, including the
  leading position (the leading-character rule blocks option injection
  via `-`, which `@` cannot cause)
- Reject the bare name "@" with a dedicated check
- Update the surrounding comment, JSDoc, and error message to match
- Add test cases for @-containing names and bare "@"

Fixes #998

Co-authored-by: Claude Fable 5 <[email protected]>
2026-06-22 14:42:36 -07:00
e452eb9dce test: cover format-turns content-type fallbacks and system_other handling (#1421)
Adds unit tests for previously-uncovered branches in
src/entrypoints/format-turns.ts:

- detectContentType: malformed-JSON fall-through (objects and arrays)
  and the default python classification for non-python/non-js code
- formatResultContent: non-string inputs (number, plain object)
- groupTurnsNaturally / formatGroupedContent: the system_other path
  for non-init system turns

Tests only; no source changes. format-turns.ts line coverage rises
from ~86% and the file's non-CLI logic is now fully exercised.

Co-authored-by: hk <[email protected]>
Co-authored-by: Claude Opus 4.8 <[email protected]>
2026-06-22 14:41:41 -07:00
Steven Zimmerman, CPAandGitHub 6b8063043e fix: filter PR reviews and inline review comments to trigger time (#1385)
Issue/PR comments (#512) and the issue/PR body (#710) are filtered to the
trigger timestamp so content created or edited after an authorized trigger
cannot be injected into Claude's prompt (TOCTOU protection). Reviews and
inline review comments were not: fetchGitHubData returned reviewData filtered
by actor only, and formatReviewComments renders it into the prompt, so a
review submitted or edited after the trigger reached Claude verbatim.

filterReviewsToTriggerTime already existed (added alongside the comment filter
in #512) but was only wired to the image-download list, never to the returned
reviewData.

Filter reviewData.nodes through filterReviewsToTriggerTime and each review's
inline comments through filterCommentsToTriggerTime, alongside the existing
actor filter, then build the review image-processing lists from those
already-filtered nodes (removing a now-redundant second filter pass).
Strengthen the two integration tests to assert post-trigger and edited-after
reviews/comments are dropped.
2026-06-22 14:41:33 -07:00
GitHub Actions 30544b6743 chore: bump Claude Code to 2.1.186 and Agent SDK to 0.3.186 2026-06-22 20:38:23 +00:00
GitHub Actions 2fee155104 chore: bump Claude Code to 2.1.185 and Agent SDK to 0.3.185 2026-06-20 21:00:13 +00:00
GitHub Actions 51705da45e chore: bump Claude Code to 2.1.183 and Agent SDK to 0.3.183 2026-06-19 01:21:31 +00:00
GitHub Actions 806af32823 chore: bump Claude Code to 2.1.181 and Agent SDK to 0.3.181 2026-06-17 22:08:47 +00:00
Ryan NoonanandGitHub 0a08a86780 fix: skip workflow validation token exchange failures (#1417) 2026-06-17 13:53:35 -07:00
GitHub Actions 9dd8b95a39 chore: bump Claude Code to 2.1.179 and Agent SDK to 0.3.179 2026-06-16 20:23:28 +00:00
GitHub Actions 4d7e1f0cd8 chore: bump Claude Code to 2.1.178 and Agent SDK to 0.3.178 2026-06-15 21:37:45 +00:00
3d9f0dc7dc fix(mcp): align allowed-tools parser with SDK option parser (#1373)
parseAllowedTools (used to decide which GitHub MCP servers to install)
hand-rolled a regex parse of claude_args, while the tools actually
granted to Claude are parsed by parseClaudeArgsToExtraArgs in
base-action/src/parse-sdk-options.ts using shell-quote. The two parsers
diverged on two inputs (#1357):

- Multiple values after a single flag: for
  `--allowedTools "Read" "Grep" "mcp__github__get_commit"` the regex
  captured only "Read", so the github MCP server was not installed even
  though mcp__github__get_commit was granted — tool calls then failed.
- Commented-out lines: the regex counted tools on `#`-prefixed lines
  that the SDK parser strips, installing servers that were never used.

Reimplement parseAllowedTools on the same shell-quote tokenizer and the
same "accumulating flag consumes all consecutive non-flag values"
semantics, stripping comment lines first, so the install decision agrees
with the tools that are actually granted. Unquoted glob patterns (e.g.
`mcp__github__*`), which shell-quote yields as glob objects, are
recovered to their literal text to preserve existing behavior.

Closes #1357

Co-authored-by: bymle <[email protected]>
2026-06-13 22:49:34 -07:00
a5e5d3b82e fix(parse-sdk-options): prevent shell-quote from collapsing unquoted Bash(X:*) rules to bare Bash (#1350)
* fix(parse-sdk-options): prevent shell-quote from collapsing unquoted Bash(X:*) rules to bare Bash

shell-quote's parse() tokenizes unquoted `(`, `)` as control operators
and barewords containing `*` as glob ops, all returned as non-string
objects. parseClaudeArgsToExtraArgs filtered those out, so an unquoted
`--allowedTools View,Bash(gh:*),Bash(cat:*)` collapsed to bare `Bash` —
silently widening scoped permission rules to unrestricted Bash(*).

Escape shell control metachars to Unicode private-use placeholders
before parse() and restore after; extract .pattern from glob ops.
Preserves existing quote/whitespace handling.

Co-Authored-By: Claude Opus 4.7 <[email protected]>

* ci: retrigger

Co-Authored-By: Claude Opus 4.7 <[email protected]>

---------

Co-authored-by: Claude Opus 4.7 <[email protected]>
2026-06-13 22:45:14 -07:00
GitHub Actions d5726de019 chore: bump Claude Code to 2.1.177 and Agent SDK to 0.3.177 2026-06-13 01:25:48 +00:00
GitHub Actions 56fa348258 chore: bump Claude Code to 2.1.176 and Agent SDK to 0.3.176 2026-06-12 21:54:44 +00:00
Ashwin BhatandGitHub 82d95d45af Add pr-stamp-sweep review workflow (#1409)
* Add pr-stamp-sweep review workflow

Adds a reusable workflow script that fans out one reviewer agent per
candidate PR to judge whether it can be approved as-is, then runs an
adversarial verification pass on each approval recommendation before
including it in the final list.

Reviewers read pre-fetched PR dossiers from /tmp/claude/pr-sweep/<n>.md
(metadata, body, comments, full diff) rather than calling gh directly,
and verify diff claims against the local checkout. PR numbers are
passed via args; the script fails fast if none are provided.

* Sharpen security checklist in reviewer prompts

Replace the general security-history note with explicit attack classes
both reviewer stages must check: prompt injection (untrusted content
reaching Claude's context, including via tool output), code execution
(shell commands, eval/spawn, workflow expressions), path traversal,
and credential exposure.
2026-06-11 21:54:48 -07:00
GitHub Actions 0cb4f3e5e7 chore: bump Claude Code to 2.1.175 and Agent SDK to 0.3.175 2026-06-12 04:24:54 +00:00
8551f4b0aa fix(image-downloader): detect image type from magic bytes (#1396)
GitHub serves pasted attachments from /user-attachments/assets/<uuid>
with no file extension, so getImageExtension() silently defaulted to
".png". When the bytes are actually JPEG/GIF/WebP the downloaded file is
mislabeled, and the Read tool then sends a base64 image whose declared
media_type doesn't match its magic bytes — which the Anthropic API
rejects with `400 invalid_request_error` ("image was specified using the
image/png media type, but the image appears to be a image/jpeg image").

Sniff the real format from the buffer's magic bytes after download and
only fall back to the URL-based extension when the signature is
unrecognized. Adds a regression test for a JPEG at an extensionless URL.

Co-authored-by: Claude Opus 4.8 (1M context) <[email protected]>
2026-06-11 21:23:35 -07:00
looooownandGitHub eba921ff6f docs: fix execution file parsing example (#1297) 2026-06-11 21:19:57 -07:00
36617bd48b fix(sanitizer): match attribute quotes by type to avoid mangling content (#1371)
stripHiddenAttributes used the pattern `["'][^"']*["']` for each quoted
attribute, which matches an opening quote of either type and stops at the
first quote of either type. When a value contained the other quote
character — e.g. an apostrophe inside a double-quoted attribute like
`title="We'll do it"` — the match terminated at the apostrophe, so the
wrong span was removed and the surrounding text was corrupted (e.g.
`<Tooltip title="We'll do it" placement="top">` became
`<Tooltipll do it" placement="top">`).

This surfaced via the github_inline_comment MCP tool: suggestion blocks
whose code lines contain quotes were mangled before posting (#1366).

Match each quoted form per quote type (`"[^"]*"` and `'[^']*'`),
mirroring stripMarkdownLinkTitles, so a value may freely contain the
other quote character. The unquoted fallback is unchanged.

Closes #1366

Co-authored-by: bymle <[email protected]>
2026-06-11 21:19:44 -07:00
24b915648e Include labels in formatContext() output for issues and PRs (#1298)
The formatted_context block sent to the agent omitted labels for both
issues and pull requests, even though the GraphQL queries already
fetched them. This caused agents to incorrectly report "no labels"
when labels existed, breaking any workflow that routes on label state
(e.g., drift-fix routing on drift:* labels in a Drift Watcher pattern).

Changes:
- Add 'PR Labels:' line to PR context output
- Add 'Issue Labels:' line to issue context output
- Both emit 'none' when no labels are present (explicit > omitted)
- Bump labels(first: 1) → labels(first: 100) in both queries; the
  previous cap meant only one label would appear even after the
  formatter fix
- Update existing tests + add 'with labels' tests for both PR and
  issue branches

Discovered while building a GitHub Actions workflow that uses this
action for drift-watcher routing in an internal seed framework
(joshpayne-joby/slim-routines#6). The agent self-diagnosed the gap
by inspecting this action's source — a satisfying full-loop.

Co-authored-by: Claude Opus 4.7 <[email protected]>
2026-06-11 21:19:15 -07:00
Kyungil ParkandGitHub 9441a7fe22 fix: clear stale claude-prompts dir before each write (#1288)
Previously, prompt files at `${RUNNER_TEMP}/claude-prompts/` from a prior
invocation could persist on non-ephemeral self-hosted runners (where the
documented `RUNNER_TEMP` cleanup contract is not reliably honored). In
particular, `claude-user-request.txt` is only written by `create-prompt`
when a user request exists; an `agent`-mode invocation does not overwrite
it, so a stale value left by an earlier mention-mode job in another repo
would leak into a later agent-mode job's effective context on the same
runner agent.

Fix: `rm -rf` the directory before `mkdir` in both write sites
(`src/create-prompt/index.ts`, `src/modes/agent/index.ts`). Idempotent,
safe on hosted runners (where the dir is already empty), and self-heals
on self-hosted runners.

Closes #1287
2026-06-11 21:18:57 -07:00
KiwiandGitHub b371255139 pin setup-bun path for post steps (#1365)
Signed-off-by: kiwigitops <[email protected]>
2026-06-11 21:18:24 -07:00
Stephen CobbeandGitHub 84d317e8f9 fix: break SDK iterator after result message to prevent hang (#1339)
In some workflow contexts — reliably reproducible for us on
pull_request-triggered runs of this action — the Claude Agent SDK
query() async iterator does not close after the terminal result
message is emitted. The for-await loop in runClaudeWithSdk blocks
indefinitely after Claude has finished its work, until the workflow's
timeout-minutes cap kills the job.

Symptoms observed in production (4× in our scan-reviewer workflow):
- Claude completes successfully: SDK emits { type: "result",
  subtype: "success", ... } with the cost / turns / duration set.
- The action then sits with zero log output for the rest of
  timeout-minutes (we measured 18-19 min of dead time after result).
- The job is cancelled at timeout. writeExecutionFile is never
  called → no claude-execution-output.json → cost-tracker and other
  post-steps see nothing.
- Run shows as cancelled, even though Claude did its work and any
  verdict it posted via gh tools already landed.

Author-mode (workflow_dispatch) runs from the same codebase
terminate cleanly the same day, so the hang is specific to certain
event triggers.

By SDK contract the result message is terminal — no further messages
follow. Break out of the loop immediately after capturing it,
regardless of whether the upstream iterator ever closes. If the SDK
is later fixed to close cleanly in all contexts, this break becomes
a no-op.
2026-06-11 21:17:47 -07:00
Nikita KirsanovandGitHub cd59d5df0d fix: fall back to inherited env for auth when inputs are empty (#1342)
The "Run Claude Code Action" step maps the auth inputs into env
unconditionally:

    ANTHROPIC_API_KEY: ${{ inputs.anthropic_api_key }}
    CLAUDE_CODE_OAUTH_TOKEN: ${{ inputs.claude_code_oauth_token }}

When the input is empty — whether because the caller supplied auth via
the step `env:` block (as reported in #676) or because the `with:` value
resolves empty in some runner/secret configurations — this assignment
overwrites the inherited env value with an empty string. validate-env
then fails with the misleading "Either ANTHROPIC_API_KEY or
CLAUDE_CODE_OAUTH_TOKEN is required" error even though the caller did
provide a token.

Fall back to the inherited env var when the input is empty, mirroring the
existing `${{ env.X }}` pattern already used a few lines below for
ANTHROPIC_BASE_URL / ANTHROPIC_CUSTOM_HEADERS. The input still takes
precedence; nothing changes for workflows that pass auth via `with:`.

Fixes #676
2026-06-11 21:17:22 -07:00
8046d850b5 docs(faq): correct rebase FAQ to match actual behavior (#1370)
The "Why won't Claude rebase my branch?" FAQ told users they could
enable rebasing by passing `--allowedTools "Bash(git rebase:*)"` via
claude_args. This does not work: the system prompt built in
src/create-prompt/index.ts unconditionally instructs Claude that it
cannot merge, rebase, or perform branch operations beyond creating and
pushing commits, so Claude declines rebase requests regardless of the
allowed tools.

Update the FAQ to describe the actual behavior and point users to the
real workaround (rebase locally or via the Claude Code CLI).

Closes #1286

Co-authored-by: bymle <[email protected]>
2026-06-11 21:16:55 -07:00
ee2b19d882 test: add unit tests for parseGitHubContext and context type guards (#1404)
* test: add unit tests for parseGitHubContext and context type guards

Covers all supported webhook event types (entity and automation),
the pull_request_target normalization, isPR detection for comments
on pull requests, env-derived input defaults and parsing, and the
nine type guard functions. Raises src/github/context.ts line
coverage from 26.5% to 100%.

Co-Authored-By: Claude Fable 5 <[email protected]>

* test: assert all env-derived inputs to close mutation testing gaps

Mutation testing (StrykerJS, 138 mutants) showed 22 surviving mutants
in the env input parsing of parseGitHubContext: fields like
trackProgress, includeFixLinks, allowedBots and the comment actor
filters were never asserted. Asserting every input field in both the
defaults and the explicit-values tests, plus covering the optional
chaining on payload.repository, brings the mutation score for
src/github/context.ts from 84.06% to 100%.

Co-Authored-By: Claude Fable 5 <[email protected]>

---------

Co-authored-by: Claude Fable 5 <[email protected]>
2026-06-11 21:16:45 -07:00
GitHub Actions ebcdfe6dc6 chore: bump Claude Code to 2.1.174 and Agent SDK to 0.3.174 2026-06-12 01:17:25 +00:00
GitHub Actions 0f97b95b65 chore: bump Claude Code to 2.1.173 and Agent SDK to 0.3.173 2026-06-11 05:43:00 +00:00
GitHub Actions eee73e2ae5 chore: bump Claude Code to 2.1.172 and Agent SDK to 0.3.172 2026-06-10 21:01:40 +00:00
232c9a15f4 Drop --tsconfig-override from Bun invocations to avoid runtime crash (#1315)
* Drop --tsconfig-override from Bun invocations to avoid runtime crash

Passing --tsconfig-override to `bun run` triggers a Bun runtime bug
("Internal error: directory mismatch for directory .../tsconfig.json")
that aborts the action with exit code 1 before any work is done.

Bun already auto-discovers the action's own tsconfig.json by walking up
the directory tree from the entry file, so the override is redundant —
the workspace's tsconfig is never an ancestor of the action checkout.
Dropping the flag preserves tsconfig resolution while avoiding the crash.

Refs: oven-sh/bun#25730

https://claude.ai/code/session_01L763e4S7zBnzDqmYYEJS1A

* Fix prettier formatting in create-prompt/index.ts

Removes redundant outer parentheses that were tripping format:check.
Pre-existing on main; unrelated to the action.yml change but needed
to keep CI green on this branch.

https://claude.ai/code/session_01L763e4S7zBnzDqmYYEJS1A

---------

Co-authored-by: Claude <[email protected]>
2026-06-09 14:12:16 -07:00
GitHub Actions 11ba60486e chore: bump Claude Code to 2.1.170 and Agent SDK to 0.3.170 2026-06-09 17:27:26 +00:00
GitHub Actions 593d7a5c4e chore: bump Claude Code to 2.1.169 and Agent SDK to 0.3.169 2026-06-08 21:57:59 +00:00
GitHub Actions fbda2eb1bd chore: bump Claude Code to 2.1.168 and Agent SDK to 0.3.168 2026-06-06 23:42:46 +00:00
GitHub Actions 64de744025 chore: bump Claude Code to 2.1.167 and Agent SDK to 0.3.167 2026-06-06 01:34:14 +00:00
GitHub Actions 410165836e chore: bump Claude Code to 2.1.166 and Agent SDK to 0.3.166 2026-06-06 00:56:12 +00:00
GitHub Actions 41ea7642c1 chore: bump Claude Code to 2.1.165 and Agent SDK to 0.3.165 2026-06-05 05:45:57 +00:00
GitHub Actions 0b1b620029 chore: bump Claude Code to 2.1.163 and Agent SDK to 0.3.163 2026-06-04 21:54:01 +00:00
GitHub Actions 70a6e5256e chore: bump Claude Code to 2.1.162 and Agent SDK to 0.3.162 2026-06-03 21:32:44 +00:00
155 changed files with 19012 additions and 1151 deletions
+1 -1
View File
@@ -5,7 +5,7 @@
"hooks": [
{
"type": "command",
"command": "bun run format"
"command": "bunx [email protected] --no-config --write ."
}
],
"matcher": "Edit|Write|MultiEdit"
+162
View File
@@ -0,0 +1,162 @@
export const meta = {
name: "pr-stamp-sweep",
description:
"Review candidate PRs for stampability, then adversarially verify security of stamp candidates",
whenToUse:
"Sweep candidate PRs for stampability: per-PR review + adversarial security verify. Requires pre-fetched PR dossiers in /tmp/claude/pr-sweep/<n>.md and args {prs: [...]}.",
phases: [
{ title: "Review", detail: "one reviewer agent per PR" },
{
title: "Verify",
detail: "adversarial security skeptic per stamp candidate",
},
],
};
// PRECONDITION: before invoking this workflow, pre-fetch each candidate PR to
// /tmp/claude/pr-sweep/<n>.md, containing the PR's metadata, body, existing
// reviews/comments, and the full diff (e.g. via `gh pr view` + `gh pr diff`).
// Sandboxed agents can't reliably call gh themselves, so they read these
// dossier files instead. Pass the PR numbers as args: {prs: [<PR numbers>]}.
const REVIEW_SCHEMA = {
type: "object",
properties: {
number: { type: "number" },
verdict: { type: "string", enum: ["stamp", "skip", "needs-discussion"] },
category: {
type: "string",
description: "docs | tests | bugfix | nicety | security-fix | other",
},
summary: {
type: "string",
description: "1-2 sentence plain-language summary of what the PR does",
},
reasoning: {
type: "string",
description: "why this verdict — correctness, scope, quality",
},
behaviorChange: {
type: "string",
description: 'what user-visible behavior changes, or "none"',
},
concerns: { type: "array", items: { type: "string" } },
securitySensitive: {
type: "boolean",
description:
"true if it touches auth, sanitization, parsers of untrusted input, actor checks, file restore, or shell construction",
},
duplicateOf: {
type: "string",
description: "PR number(s) this duplicates, or empty string",
},
},
required: [
"number",
"verdict",
"category",
"summary",
"reasoning",
"behaviorChange",
"concerns",
"securitySensitive",
"duplicateOf",
],
};
const VERDICT_SCHEMA = {
type: "object",
properties: {
number: { type: "number" },
safeToStamp: { type: "boolean" },
findings: {
type: "array",
items: { type: "string" },
description:
"concrete security/correctness problems found, empty if clean",
},
confidence: { type: "string", enum: ["high", "medium", "low"] },
},
required: ["number", "safeToStamp", "findings", "confidence"],
};
if (!args || !Array.isArray(args.prs) || args.prs.length === 0)
throw new Error(
"pass {prs: [<PR numbers>]} as args; pre-fetch each PR to /tmp/claude/pr-sweep/<n>.md first",
);
const prs = args.prs;
log(`Reviewing ${prs.length} candidate PRs`);
const results = await pipeline(
prs,
(n) =>
agent(
`You are reviewing open PR #${n} on anthropics/claude-code-action to decide if it is safe for a maintainer to approve ("stamp") with minimal further discussion.
The full PR (metadata, body, existing reviews/comments, and complete diff) is in /tmp/claude/pr-sweep/${n}.md — read it first. The repo is checked out at the current working directory. Read the actual current source files the diff touches to verify the diff applies cleanly conceptually and the claims in the PR body are true. Do NOT modify anything or run git commands that change state.
Context about this repo:
- It's a GitHub Action that runs Claude on issues/PRs. It processes UNTRUSTED content (PR bodies, comments, branch names, file contents from forks). Treat any change touching content sanitization, actor/bot allowlists, config restoration, prompt construction, or shell command construction as high-risk.
- Most candidate PRs are from EXTERNAL contributors. Treat the diff with suspicion: look for subtle malicious changes, weakened validation, injection vectors, overly broad permissions, or changes whose description doesn't match the code.
- Runtime is Bun; strict TypeScript (noUnusedLocals/noUnusedParameters). Tests are unit tests run with bun test.
Stamp criteria (ALL must hold):
1. Small, focused, and the code does exactly what the title/body says.
2. No major behavior change — bug fixes restoring intended behavior, docs fixes, test-only additions, and small niceties qualify. New inputs/features, behavior redesigns, or large refactors do NOT.
3. Correct: you verified the logic against the actual current source, not just the diff. Check edge cases.
4. No security concern. Check explicitly for: prompt injection (untrusted text reaching Claude's prompt without sanitization), code execution (untrusted data reaching shell commands, eval/spawn, or GitHub workflow expressions), path traversal (untrusted input influencing filesystem paths), credential exposure (tokens reaching logs, comments, or attacker-readable output), weakened validation or permission checks, and suspicious hunks unrelated to the stated purpose.
5. Wouldn't break the public API of base-action/ or action.yml output wiring.
If the PR is a docs change, verify the docs claims against the actual code behavior. If test-only, check tests actually pass conceptually (assert the right things, match real implementations) and don't weaken or skip anything.
Verdicts: "stamp" = approve as-is; "needs-discussion" = plausible but has questions/issues worth a comment; "skip" = too big, wrong, redundant, or risky.
If this PR appears to duplicate another open PR (same fix, same files), still judge it on its own merits but note the duplication in duplicateOf.
Return structured output only.`,
{ label: `review:#${n}`, phase: "Review", schema: REVIEW_SCHEMA },
),
(review, n) => {
if (!review) return null;
if (review.verdict !== "stamp") return { review, verify: null };
return agent(
`You are an adversarial security skeptic. Another reviewer recommended APPROVING open PR #${n} on anthropics/claude-code-action. Your job is to REFUTE that recommendation — find any reason it should NOT be stamped.
Their assessment: ${JSON.stringify(review)}
Read the full PR at /tmp/claude/pr-sweep/${n}.md and the touched source files in the current working directory. This repo processes untrusted PR/issue content from forks; anything that lets untrusted content reach Claude's prompt, a shell command, a workflow expression, or a filesystem path unsanitized is a critical vulnerability.
Hunt specifically for:
- Subtle malice or scope creep: hunks that don't match the stated purpose, weakened validation, regex changes that widen acceptance, removed escaping.
- Prompt injection: untrusted data (comment bodies, branch names, file contents, command output, downloaded files) reaching Claude's prompt or context without sanitization, including indirect routes like tool output Claude later reads.
- Code execution: untrusted data reaching shell commands, eval/spawn argv, GitHub workflow \${{ }} expressions, or API call templates; new process spawning; path traversal letting untrusted input write or read outside intended directories.
- Credential exposure: tokens or secrets flowing into logs, posted comments, error messages, env passed to untrusted code, or files Claude can read.
- Logic errors the first reviewer missed: off-by-one, wrong polarity, unhandled edge cases (empty strings, unicode, very long inputs).
- Supply-chain angles: pinned versions that don't match the claimed SHA/tag, new dependencies, fetched URLs.
- For docs PRs: claims that would mislead users into insecure configurations.
- For test-only PRs: tests that codify wrong behavior, or that would mask future regressions.
If the diff pins a version/SHA, verify the claim is plausible from local information; flag if unverifiable. Be strict: if uncertain whether something is a real problem, lean toward reporting it as a finding with your uncertainty noted. Only return safeToStamp=true if you genuinely failed to find any disqualifying issue.
Return structured output only.`,
{ label: `verify:#${n}`, phase: "Verify", schema: VERDICT_SCHEMA },
).then((v) => ({ review, verify: v }));
},
);
const clean = results.filter(Boolean);
const stamped = clean.filter(
(r) => r.review.verdict === "stamp" && r.verify && r.verify.safeToStamp,
);
const demoted = clean.filter(
(r) => r.review.verdict === "stamp" && (!r.verify || !r.verify.safeToStamp),
);
const discuss = clean.filter((r) => r.review.verdict === "needs-discussion");
const skipped = clean.filter((r) => r.review.verdict === "skip");
log(
`stamp: ${stamped.length}, demoted by verifier: ${demoted.length}, needs-discussion: ${discuss.length}, skip: ${skipped.length}`,
);
return { stamped, demoted, discuss, skipped };
+3
View File
@@ -0,0 +1,3 @@
# Keep the working tree LF on every platform: Prettier defaults to
# endOfLine "lf" and several tests match LF-terminated content.
* text=auto eol=lf
+29
View File
@@ -0,0 +1,29 @@
# Hosts that jobs on GitHub's egress-firewall runner (runs-on: ubuntu-24.04-firewall)
# may reach. All other hosts are blocked, apart from any that GitHub's firewall allows
# by default. Add a host only when a workflow step needs it, name it in full (no '*'), and say what uses it.
mode: enforce
allow:
# Claude API: model requests, trading the workflow's GitHub identity token for a
# short-lived API token, and fetching the Claude GitHub App's token in claude.yml and
# claude-review.yml
- api.anthropic.com
# GitHub API: calls made by the Claude Code action, by gh in scripts/gh.sh and
# scripts/edit-issue-labels.sh, and by the /review-pr command
- api.github.com
# Claude Code install script: fetched by the Claude Code action, and by
# test-custom-executables.yml ("Install Claude Code manually")
- claude.ai
# Claude Code binary, downloaded by the install script
- downloads.claude.ai
# Bun and Node.js release downloads (oven-sh/setup-bun and actions/setup-node)
- release-assets.githubusercontent.com
# npm packages (bun install in action.yml, base-action/action.yml and
# test-mcp-servers.yml)
- registry.npmjs.org
# Bun install script (test-custom-executables.yml, "Install Bun manually")
- bun.sh
# apt packages bubblewrap and socat, which the Claude Code action installs when a
# workflow admits users without write access (issue-triage.yml)
- azure.archive.ubuntu.com
- archive.ubuntu.com
- security.ubuntu.com
+11
View File
@@ -0,0 +1,11 @@
# Groups the notes of each release. Merged pull requests with no label go under
# "Other changes". .github/workflows/release.yml reads the first heading to flag
# the release's title, so keep the two in step.
changelog:
categories:
- title: "⚠️ Breaking changes"
labels:
- breaking-change
- title: Other changes
labels:
- "*"
+339
View File
@@ -0,0 +1,339 @@
#!/usr/bin/env python3
"""Fail if a workflow job that calls Claude, or .github/egress-firewall.yaml, breaks a rule in CLAUDE.md,
"Security hardening for GitHub Actions". Run from the repository root. A job calls Claude when it runs the
Claude Code action, or when it or a local action it uses mentions ANTHROPIC_FEDERATION_RULE_ID.
"""
import json
import pathlib
import re
import shlex
import subprocess
import sys
FIREWALL_RUNNER = "ubuntu-24.04-firewall"
WORKFLOW_DIR = pathlib.Path(".github/workflows")
POLICY_PATH = pathlib.Path(".github/egress-firewall.yaml")
SIGN_IN_MARKER = "anthropic_federation_rule_id"
CLAUDE_ACTIONS = ("anthropics/claude-code-action", "anthropics/claude-code-base-action")
HELP = 'See CLAUDE.md, "Security hardening for GitHub Actions".'
# Claude Code runs auto mode only on claude-opus-4-6 and newer models. On an older model it
# falls back to its default permission mode, with no safety review.
AUTO_MODE_MIN_VERSION = (4, 6)
# The version in a model name: claude-opus-4-6, claude-sonnet-4-5-20250929, claude-3-5-sonnet-latest.
MODEL_VERSION = re.compile(
r"claude-(?:(?:opus|sonnet|haiku)-(\d+)(?:-(\d{1,2}))?"
r"|(\d+)(?:-(\d{1,2}))?-(?:opus|sonnet|haiku))(?![\d.])"
)
# Key: "<workflow file name>:<job id>". Value: why that job is exempt from the table's rule.
EXEMPT_FROM_FIREWALL_RUNNER: dict[str, str] = {}
EXEMPT_FROM_AUTO_MODE: dict[str, str] = {}
def stop(message: str):
sys.exit(f"::error::{message}")
def load_yaml(path: pathlib.Path):
"""Parse a YAML file with PyYAML, or with the yq command if PyYAML is absent."""
try:
import yaml
except ImportError:
try:
result = subprocess.run(
["yq", "-o=json", ".", str(path)], check=True, capture_output=True, text=True
)
except FileNotFoundError:
stop(
f"Cannot read {path}: Python has no 'yaml' module and no 'yq' command was found. "
"Add a step that runs 'pip install pyyaml' before this check."
)
except subprocess.CalledProcessError:
stop(f"Cannot read {path}: 'yq' could not parse it. Check that the file is valid YAML.")
return json.loads(result.stdout)
try:
with path.open(encoding="utf-8") as handle:
return yaml.safe_load(handle)
except yaml.YAMLError as error:
stop(f"Cannot read {path}: it is not valid YAML ({error}).")
def contains_marker(node) -> bool:
"""Whether any key or string under node contains SIGN_IN_MARKER, ignoring case."""
if isinstance(node, dict):
return any(contains_marker(k) or contains_marker(v) for k, v in node.items())
if isinstance(node, list):
return any(contains_marker(item) for item in node)
return isinstance(node, str) and SIGN_IN_MARKER in node.lower()
def load_local_action(uses: str):
"""The parsed action file of a local action (uses: ./path), or None."""
if not uses.startswith("./"):
return None
for name in ("action.yml", "action.yaml"):
action_file = pathlib.Path(uses) / name
if action_file.is_file():
action = load_yaml(action_file)
return action if isinstance(action, dict) else {}
return None
def steps_of(job: dict) -> list[dict]:
return [step for step in job.get("steps") or [] if isinstance(step, dict)]
def runs_claude_code_action(step: dict) -> bool:
"""Whether the step runs the Claude Code action: the published action, or a
local action that accepts a claude_args input."""
uses = str(step.get("uses", ""))
if uses.lower().startswith(CLAUDE_ACTIONS):
return True
action = load_local_action(uses)
return action is not None and "claude_args" in (action.get("inputs") or {})
def job_calls_claude(job: dict) -> bool:
if contains_marker(job):
return True
for step in steps_of(job):
if runs_claude_code_action(step):
return True
action = load_local_action(str(step.get("uses", "")))
if action is not None and contains_marker(action):
return True
return False
def permission_mode_problem(step: dict, exempt: bool, inherited_env: dict) -> str | None:
"""The message for a step whose permission mode is wrong, or None if it is right.
A step must set auto mode, on a model that supports it. A step of a job in
EXEMPT_FROM_AUTO_MODE must set no mode at all. inherited_env is the workflow's and the
job's 'env'.
"""
inputs = step.get("with") or {}
lines = str(inputs.get("claude_args", "")).splitlines()
text = " ".join(line for line in lines if not line.strip().startswith("#"))
try:
args = shlex.split(text, comments=True)
except ValueError:
return "'claude_args' has a quote that is never closed. Close it"
modes = []
for index, arg in enumerate(args):
if arg == "--dangerously-skip-permissions":
return (
"remove '--dangerously-skip-permissions' from 'claude_args': "
"it turns off permission checks"
)
if arg == "--permission-mode":
modes.append(args[index + 1] if index + 1 < len(args) else "")
elif arg.startswith("--permission-mode="):
modes.append(arg.split("=", 1)[1])
if exempt and modes:
return (
"remove '--permission-mode' from 'claude_args': this job is listed in "
"EXEMPT_FROM_AUTO_MODE (.github/scripts/check_workflow_hardening.py), and a job listed "
"there must not set a permission mode"
)
if not modes and not exempt:
return "add '--permission-mode auto' to 'claude_args' under the step's 'with:'"
for mode in modes:
if mode == "":
return (
"'claude_args' has '--permission-mode' with nothing after it. "
"Write '--permission-mode auto'"
)
if mode != "auto":
return (
f"'claude_args' has '--permission-mode {mode}'. "
"Change it to '--permission-mode auto'"
)
env = {**inherited_env, **(step.get("env") or {})}
models = [
("the step's 'model'", inputs.get("model", "")),
("ANTHROPIC_MODEL", env.get("ANTHROPIC_MODEL", "")),
]
models += [
(f"'{flag}' in 'claude_args'", value)
for flag in ("--model", "--fallback-model")
for value in flag_values(args, flag)
]
settings = [("the step's 'settings'", inputs.get("settings", ""))]
settings += [
("'--settings' in 'claude_args'", value) for value in flag_values(args, "--settings")
]
for where, value in settings:
text = settings_text(value)
if text is None:
return (
f"{where} names a file outside the repository, which this check cannot read. "
"Use inline settings or a file inside the repository"
)
if "defaultMode" in text:
return f"remove 'defaultMode' from {where}: settings must not set a permission mode"
try:
parsed = json.loads(text) if text else {}
except ValueError:
parsed = {}
if isinstance(parsed, dict) and "model" in parsed:
models.append((f"'model' in {where}", parsed["model"]))
if not exempt:
for where, model in models:
if predates_auto_mode(str(model or "")):
return (
f"{where} is '{model}', which Claude Code does not run in auto mode: it "
"falls back to the default permission mode. Use claude-opus-4-6 or a newer model"
)
return None
def flag_values(args: list[str], flag: str) -> list[str]:
"""The values a flag in claude_args is given, as '--flag value' or '--flag=value'."""
values = [
args[index + 1] for index, arg in enumerate(args) if arg == flag and index + 1 < len(args)
]
return values + [arg.split("=", 1)[1] for arg in args if arg.startswith(f"{flag}=")]
def predates_auto_mode(model: str) -> bool:
"""Whether the model is older than AUTO_MODE_MIN_VERSION. A name with no version, such as
'opus' or 'default', stands for a current model, except 'haiku' (claude-haiku-4-5)."""
if model.strip().lower() == "haiku":
return True
match = MODEL_VERSION.search(model.lower())
if not match:
return False
major, minor = match.group(1, 2) if match.group(1) else match.group(3, 4)
return (int(major), int(minor or 0)) < AUTO_MODE_MIN_VERSION
def settings_text(value) -> str | None:
"""The settings JSON a 'settings' value stands for: the value itself, or the contents of
the file it names when it is a path inside the repository. None when it names a path
outside the repository."""
text = str(value or "").strip()
if not text or text.startswith("{"):
return text
path = pathlib.Path(text)
root = pathlib.Path.cwd().resolve()
try:
resolved = path.resolve()
resolved.relative_to(root)
except (OSError, ValueError):
return None
if resolved.is_file():
return resolved.read_text(encoding="utf-8", errors="replace")
return text
def check_job(file_name: str, job_id: str, job: dict, workflow_env: dict) -> list[str]:
key = f"{file_name}:{job_id}"
where = f".github/workflows/{file_name}: job '{job_id}'"
errors = []
runs_on = job.get("runs-on")
if isinstance(runs_on, list) and len(runs_on) == 1:
runs_on = runs_on[0]
if key in EXEMPT_FROM_FIREWALL_RUNNER:
print(
f"The egress-firewall runner is not required for job '{job_id}' in {file_name}. "
f"Reason: {EXEMPT_FROM_FIREWALL_RUNNER[key]}."
)
elif runs_on != FIREWALL_RUNNER:
if "runs-on" not in job:
has = "no 'runs-on'"
elif isinstance(job["runs-on"], str):
has = f"'runs-on: {job['runs-on']}'"
else:
has = "a 'runs-on' list or group"
errors.append(
f"{where} calls Claude, so it must have 'runs-on: {FIREWALL_RUNNER}'. "
f"It has {has}. {HELP}"
)
exempt = key in EXEMPT_FROM_AUTO_MODE
if exempt:
print(
f"Auto permission mode is not required for job '{job_id}' in {file_name}. "
f"Reason: {EXEMPT_FROM_AUTO_MODE[key]}."
)
if "defaultMode" in json.dumps(job):
# Catches a settings file that an earlier step of the job writes, which the
# step-level check cannot read.
errors.append(
f"{where} mentions 'defaultMode': settings must not set a permission mode. {HELP}"
)
for index, step in enumerate(steps_of(job), start=1):
if not runs_claude_code_action(step):
continue
inherited_env = {**workflow_env, **(job.get("env") or {})}
problem = permission_mode_problem(step, exempt, inherited_env)
if problem:
step_label = f"step '{step['name']}'" if "name" in step else f"step {index}"
errors.append(f"{where}, {step_label}: {problem}. {HELP}")
return errors
def check_policy() -> list[str]:
if not POLICY_PATH.is_file():
return [
f"{POLICY_PATH} is missing. Jobs on the egress-firewall runner need it "
f"to limit outbound network access. {HELP}"
]
policy = load_yaml(POLICY_PATH)
if not isinstance(policy, dict):
return [
f"{POLICY_PATH} is empty or is not a set of 'name: value' lines. It needs 'mode: enforce' "
f"and an 'allow:' list of hosts. {HELP}"
]
errors = []
if "mode" not in policy:
errors.append(f"{POLICY_PATH}: 'mode' is missing. Add 'mode: enforce'. {HELP}")
elif policy["mode"] != "enforce":
errors.append(
f"{POLICY_PATH}: 'mode' is '{policy['mode']}'. It must be 'enforce'. {HELP}"
)
allow = policy.get("allow")
if not isinstance(allow, list) or not allow:
errors.append(
f"{POLICY_PATH}: the 'allow' list is missing or empty. List under 'allow:' "
f"each host the jobs need. {HELP}"
)
else:
for host in allow:
if "*" in str(host):
errors.append(
f"{POLICY_PATH}: the 'allow' entry '{host}' contains '*'. "
f"Name each host in full. {HELP}"
)
return errors
def main() -> int:
if not WORKFLOW_DIR.is_dir():
stop(f"{WORKFLOW_DIR} not found. Run this check from the repository root.")
errors = []
checked = 0
for path in sorted([*WORKFLOW_DIR.glob("*.yml"), *WORKFLOW_DIR.glob("*.yaml")]):
workflow = load_yaml(path)
jobs = workflow.get("jobs") if isinstance(workflow, dict) else None
for job_id, job in (jobs or {}).items():
if not isinstance(job, dict) or not job_calls_claude(job):
continue
checked += 1
errors.extend(check_job(path.name, job_id, job, workflow.get("env") or {}))
if checked:
errors.extend(check_policy())
for error in errors:
print(f"::error::{error}")
if errors:
return 1
if checked == 0:
print("OK: no workflow job calls Claude, so there was nothing to check.")
else:
print(f"OK: checked {checked} job(s) that call Claude and found no problems.")
return 0
if __name__ == "__main__":
sys.exit(main())
+7 -4
View File
@@ -4,6 +4,9 @@ on:
pull_request:
workflow_call:
permissions:
contents: read
jobs:
test:
runs-on: ubuntu-latest
@@ -12,7 +15,7 @@ jobs:
- uses: oven-sh/setup-bun@v2
with:
bun-version: 1.2.12
bun-version: 1.4.4
- name: Install dependencies
run: bun install
@@ -25,9 +28,9 @@ jobs:
steps:
- uses: actions/checkout@v6
- uses: oven-sh/setup-bun@v1
- uses: oven-sh/setup-bun@v2
with:
bun-version: latest
bun-version: 1.4.4
- name: Install dependencies
run: bun install
@@ -42,7 +45,7 @@ jobs:
- uses: oven-sh/setup-bun@v2
with:
bun-version: 1.2.12
bun-version: 1.4.4
- name: Install dependencies
run: bun install
+9 -1
View File
@@ -6,7 +6,11 @@ on:
jobs:
review:
runs-on: ubuntu-latest
# Skip on fork PRs since they can't mint the OIDC token used for Claude API auth
if: github.event.pull_request.head.repo.full_name == github.repository
# This job calls Claude, so it runs on GitHub's egress-firewall runner, which
# filters the job's outbound network traffic (allow list: .github/egress-firewall.yaml).
runs-on: ubuntu-24.04-firewall
permissions:
contents: read
pull-requests: write
@@ -28,6 +32,10 @@ jobs:
anthropic_service_account_id: ${{ vars.ANTHROPIC_SERVICE_ACCOUNT_ID }}
prompt: "/review-pr REPO: ${{ github.repository }} PR_NUMBER: ${{ github.event.pull_request.number }}"
# --permission-mode auto: a tool call that needs permission and is outside --allowedTools
# below and the allowed-tools list in .claude/commands/review-pr.md runs only if Claude
# Code's safety review passes it.
claude_args: |
--permission-mode auto
--allowedTools "mcp__github_inline_comment__create_inline_comment"
--model "claude-opus-4-7"
+48 -2
View File
@@ -11,18 +11,55 @@ on:
types: [submitted]
jobs:
claude:
# In this repository Claude doesn't run on fork PRs, whatever triggers it. The event for a comment
# on a PR's conversation doesn't say where the PR's branch is, so this job asks GitHub. If it can't
# tell, Claude doesn't run.
fork-check:
if: |
(github.event_name == 'issue_comment' && contains(github.event.comment.body, '@claude')) ||
(github.event_name == 'pull_request_review_comment' && contains(github.event.comment.body, '@claude')) ||
(github.event_name == 'pull_request_review' && contains(github.event.review.body, '@claude')) ||
(github.event_name == 'issues' && (contains(github.event.issue.body, '@claude') || contains(github.event.issue.title, '@claude')))
runs-on: ubuntu-latest
permissions:
pull-requests: read
outputs:
run-claude: ${{ steps.check.outputs.run-claude }}
steps:
- name: Check that a pull request is not from a fork
id: check
shell: bash
env:
GH_TOKEN: ${{ github.token }}
# Empty on an issue
PR_NUMBER: ${{ github.event.pull_request.number || (github.event.issue.pull_request && github.event.issue.number) }}
run: |
same_repo=true
if [ -n "$PR_NUMBER" ]; then
for attempt in 1 2 3; do
same_repo=$(gh api "repos/$GITHUB_REPOSITORY/pulls/$PR_NUMBER" --jq '.head.repo.id == .base.repo.id') && break
[ "$attempt" -lt 3 ] || exit 1
sleep 5
done
fi
if [ "$same_repo" = true ]; then
echo "run-claude=true" >> "$GITHUB_OUTPUT"
else
echo "::notice::Claude does not run on pull requests from forks in this repository."
fi
claude:
needs: fork-check
if: needs.fork-check.outputs.run-claude == 'true'
# This job calls Claude, so it runs on GitHub's egress-firewall runner, which
# filters the job's outbound network traffic (allow list: .github/egress-firewall.yaml).
runs-on: ubuntu-24.04-firewall
permissions:
contents: write
pull-requests: write
issues: write
id-token: write
actions: read # Lets Claude read CI results on pull requests
steps:
- name: Checkout repository
uses: actions/checkout@v6
@@ -39,6 +76,15 @@ jobs:
anthropic_federation_rule_id: ${{ vars.ANTHROPIC_FEDERATION_RULE_ID }}
anthropic_organization_id: ${{ vars.ANTHROPIC_ORGANIZATION_ID }}
anthropic_service_account_id: ${{ vars.ANTHROPIC_SERVICE_ACCOUNT_ID }}
# --permission-mode auto: a tool call that needs permission and that the allowed tools
# do not cover runs only if Claude Code's safety review passes it. This is also the
# action's default, and it is set here so that the workflow hardening check can see it.
# No --allowedTools, because a rule there skips the review: bun test runs whatever the workspace holds.
# --disallowedTools: of the CI tools that actions: read enables, Claude needs only the run results. A pull
# request controls the text of logs and of job and step names.
# --append-system-prompt: how to use those results where commands are denied.
claude_args: |
--allowedTools "Bash(bun install),Bash(bun test:*),Bash(bun run format),Bash(bun typecheck)"
--disallowedTools "mcp__github_ci__download_job_log,mcp__github_ci__get_workflow_run_details"
--append-system-prompt "A safety review decides whether a command such as bun test may run, and the action denies such commands on pull requests whose author is not confirmed to have write access. If one is denied, do not try the others or ask for changes to allowedTools. On an issue, say what you could not run. On a pull request, call mcp__github_ci__get_ci_status and report the newest run named CI, which is ci.yml. Only a success conclusion means it passed. action_required means a maintainer has not approved it yet, and a run without a conclusion has not finished. Say that the result is for the latest commit of the pull request and not for edits you made, and that the pull request can change its own workflows, so a maintainer should check the run."
--model "claude-opus-4-7"
--permission-mode auto
+7 -1
View File
@@ -6,7 +6,9 @@ on:
jobs:
triage-issue:
runs-on: ubuntu-latest
# This job calls Claude, so it runs on GitHub's egress-firewall runner, which
# filters the job's outbound network traffic (allow list: .github/egress-firewall.yaml).
runs-on: ubuntu-24.04-firewall
timeout-minutes: 10
permissions:
contents: read
@@ -33,5 +35,9 @@ jobs:
anthropic_federation_rule_id: ${{ vars.ANTHROPIC_FEDERATION_RULE_ID }}
anthropic_organization_id: ${{ vars.ANTHROPIC_ORGANIZATION_ID }}
anthropic_service_account_id: ${{ vars.ANTHROPIC_SERVICE_ACCOUNT_ID }}
# --permission-mode auto: a tool call that needs permission and is outside the allowed-tools list in
# .claude/commands/label-issue.md runs only if Claude Code's safety review passes it. Tools in
# --disallowedTools never run. --model is set because Claude Code's default model changes between releases.
claude_args: '--model claude-opus-5 --permission-mode auto --disallowedTools "WebFetch,WebSearch,Write,Edit,MultiEdit,NotebookEdit"'
allowed_non_write_users: "*" # Required for issue triage workflow, if users without repo write access create issues
github_token: ${{ secrets.GITHUB_TOKEN }}
+37 -6
View File
@@ -32,9 +32,16 @@ jobs:
next_version: ${{ steps.next_version.outputs.next_version }}
steps:
- name: Checkout code
uses: actions/checkout@v6
# SHA-pinned (same pin as sync-base-action.yml): this step is handed
# the ruleset-bypass deploy key, so it must not resolve a mutable tag.
uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
with:
fetch-depth: 0
# Tag pushes below authenticate as the release deploy key (a
# bypass actor on the "Protect release tags" ruleset) instead of
# GITHUB_TOKEN, which can never bypass a ruleset. The key lives
# only in the main-restricted `production` environment.
ssh-key: ${{ secrets.RELEASE_TAG_DEPLOY_KEY }}
- name: Get latest tag
id: get_latest_tag
@@ -87,12 +94,29 @@ jobs:
if: ${{ !inputs.dry_run }}
env:
GH_TOKEN: ${{ github.token }}
NEXT_VERSION: ${{ steps.next_version.outputs.next_version }}
run: |
next_version="${{ steps.next_version.outputs.next_version }}"
# Start from the last published release, not the highest tag: a failed run
# can leave a tag with no release, and its pull requests would be dropped.
previous_release=$(gh api "repos/$GITHUB_REPOSITORY/releases?per_page=10" \
--jq '[.[] | select(.draft | not) | .tag_name | select(test("^v[0-9]+\\.[0-9]+\\.[0-9]+$"))][0]')
gh release create "$next_version" \
--title "$next_version" \
--generate-notes \
notes_file="$RUNNER_TEMP/release-notes.md"
gh api "repos/$GITHUB_REPOSITORY/releases/generate-notes" \
-f tag_name="$NEXT_VERSION" \
-f previous_tag_name="$previous_release" \
--jq .body > "$notes_file"
# The heading comes from .github/release.yml. Matching it, not the words
# alone, keeps a pull request's title from flagging the release.
title="$NEXT_VERSION"
if grep -qxF "### ⚠️ Breaking changes" "$notes_file"; then
title="$NEXT_VERSION (breaking changes)"
fi
gh release create "$NEXT_VERSION" \
--title "$title" \
--notes-file "$notes_file" \
--latest=false # keep v1 as latest
update-major-tag:
@@ -105,9 +129,16 @@ jobs:
contents: write
steps:
- name: Checkout code
uses: actions/checkout@v6
# SHA-pinned (same pin as sync-base-action.yml): this step is handed
# the ruleset-bypass deploy key, so it must not resolve a mutable tag.
uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
with:
fetch-depth: 0
# Tag pushes below authenticate as the release deploy key (a
# bypass actor on the "Protect release tags" ruleset) instead of
# GITHUB_TOKEN, which can never bypass a ruleset. The key lives
# only in the main-restricted `production` environment.
ssh-key: ${{ secrets.RELEASE_TAG_DEPLOY_KEY }}
- name: Update major version tag
run: |
+13 -4
View File
@@ -17,9 +17,16 @@ permissions:
contents: read
id-token: write
# Pin the model so these tests don't depend on the CLI's default model,
# which can change with each Claude Code release.
env:
ANTHROPIC_MODEL: claude-opus-5
jobs:
test-inline-prompt:
runs-on: ubuntu-latest
# Skip on fork PRs since they can't mint the OIDC token used for Claude API auth
if: github.event_name != 'pull_request' || github.event.pull_request.head.repo.full_name == github.repository
runs-on: ubuntu-24.04-firewall
steps:
- uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
@@ -31,7 +38,7 @@ jobs:
anthropic_federation_rule_id: ${{ vars.ANTHROPIC_FEDERATION_RULE_ID }}
anthropic_organization_id: ${{ vars.ANTHROPIC_ORGANIZATION_ID }}
anthropic_service_account_id: ${{ vars.ANTHROPIC_SERVICE_ACCOUNT_ID }}
allowed_tools: "LS,Read"
claude_args: '--allowedTools "LS,Read" --permission-mode auto'
- name: Verify inline prompt output
run: |
@@ -70,7 +77,9 @@ jobs:
fi
test-prompt-file:
runs-on: ubuntu-latest
# Skip on fork PRs since they can't mint the OIDC token used for Claude API auth
if: github.event_name != 'pull_request' || github.event.pull_request.head.repo.full_name == github.repository
runs-on: ubuntu-24.04-firewall
steps:
- uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
@@ -90,7 +99,7 @@ jobs:
anthropic_federation_rule_id: ${{ vars.ANTHROPIC_FEDERATION_RULE_ID }}
anthropic_organization_id: ${{ vars.ANTHROPIC_ORGANIZATION_ID }}
anthropic_service_account_id: ${{ vars.ANTHROPIC_SERVICE_ACCOUNT_ID }}
allowed_tools: "LS,Read"
claude_args: '--allowedTools "LS,Read" --permission-mode auto'
- name: Verify prompt file output
run: |
@@ -12,9 +12,16 @@ permissions:
contents: read
id-token: write
# Pin the model so these tests don't depend on the CLI's default model,
# which can change with each Claude Code release.
env:
ANTHROPIC_MODEL: claude-opus-5
jobs:
test-custom-executables:
runs-on: ubuntu-latest
# Skip on fork PRs since they can't mint the OIDC token used for Claude API auth
if: github.event_name != 'pull_request' || github.event.pull_request.head.repo.full_name == github.repository
runs-on: ubuntu-24.04-firewall
steps:
- uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
@@ -59,7 +66,7 @@ jobs:
anthropic_service_account_id: ${{ vars.ANTHROPIC_SERVICE_ACCOUNT_ID }}
path_to_claude_code_executable: /home/runner/.local/bin/claude
path_to_bun_executable: /home/runner/.bun/bin/bun
allowed_tools: "LS,Read"
claude_args: '--allowedTools "LS,Read" --permission-mode auto'
- name: Verify custom executables worked
run: |
+13 -3
View File
@@ -12,9 +12,16 @@ permissions:
contents: read
id-token: write
# Pin the model so these tests don't depend on the CLI's default model,
# which can change with each Claude Code release.
env:
ANTHROPIC_MODEL: claude-opus-5
jobs:
test-mcp-integration:
runs-on: ubuntu-latest
# Skip on fork PRs since they can't mint the OIDC token used for Claude API auth
if: github.event_name != 'pull_request' || github.event.pull_request.head.repo.full_name == github.repository
runs-on: ubuntu-24.04-firewall
steps:
- name: Checkout repository
uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
@@ -36,7 +43,7 @@ jobs:
anthropic_federation_rule_id: ${{ vars.ANTHROPIC_FEDERATION_RULE_ID }}
anthropic_organization_id: ${{ vars.ANTHROPIC_ORGANIZATION_ID }}
anthropic_service_account_id: ${{ vars.ANTHROPIC_SERVICE_ACCOUNT_ID }}
claude_args: --allowedTools mcp__test-server__test_tool
claude_args: --allowedTools mcp__test-server__test_tool --permission-mode auto
env:
# Change to test directory so it finds .mcp.json
CLAUDE_WORKING_DIR: ${{ github.workspace }}/base-action/test/mcp-test
@@ -94,7 +101,9 @@ jobs:
echo "✓ All MCP server checks passed!"
test-mcp-config-flag:
runs-on: ubuntu-latest
# Skip on fork PRs since they can't mint the OIDC token used for Claude API auth
if: github.event_name != 'pull_request' || github.event.pull_request.head.repo.full_name == github.repository
runs-on: ubuntu-24.04-firewall
steps:
- name: Checkout repository
uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
@@ -129,6 +138,7 @@ jobs:
anthropic_organization_id: ${{ vars.ANTHROPIC_ORGANIZATION_ID }}
anthropic_service_account_id: ${{ vars.ANTHROPIC_SERVICE_ACCOUNT_ID }}
claude_args: |
--permission-mode auto
--allowedTools mcp__test-server__test_tool
--mcp-config '{"mcpServers":{"test-server":{"type":"stdio","command":"bun","args":["simple-mcp-server.ts"],"env":{}}}}'
env:
+21 -4
View File
@@ -12,9 +12,16 @@ permissions:
contents: read
id-token: write
# Pin the model so these tests don't depend on the CLI's default model,
# which can change with each Claude Code release.
env:
ANTHROPIC_MODEL: claude-opus-5
jobs:
test-settings-inline-allow:
runs-on: ubuntu-latest
# Skip on fork PRs since they can't mint the OIDC token used for Claude API auth
if: github.event_name != 'pull_request' || github.event.pull_request.head.repo.full_name == github.repository
runs-on: ubuntu-24.04-firewall
steps:
- uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
@@ -27,6 +34,7 @@ jobs:
anthropic_federation_rule_id: ${{ vars.ANTHROPIC_FEDERATION_RULE_ID }}
anthropic_organization_id: ${{ vars.ANTHROPIC_ORGANIZATION_ID }}
anthropic_service_account_id: ${{ vars.ANTHROPIC_SERVICE_ACCOUNT_ID }}
claude_args: "--permission-mode auto"
settings: |
{
"permissions": {
@@ -65,7 +73,9 @@ jobs:
fi
test-settings-inline-deny:
runs-on: ubuntu-latest
# Skip on fork PRs since they can't mint the OIDC token used for Claude API auth
if: github.event_name != 'pull_request' || github.event.pull_request.head.repo.full_name == github.repository
runs-on: ubuntu-24.04-firewall
steps:
- uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
@@ -78,6 +88,7 @@ jobs:
anthropic_federation_rule_id: ${{ vars.ANTHROPIC_FEDERATION_RULE_ID }}
anthropic_organization_id: ${{ vars.ANTHROPIC_ORGANIZATION_ID }}
anthropic_service_account_id: ${{ vars.ANTHROPIC_SERVICE_ACCOUNT_ID }}
claude_args: "--permission-mode auto"
settings: |
{
"permissions": {
@@ -99,7 +110,9 @@ jobs:
fi
test-settings-file-allow:
runs-on: ubuntu-latest
# Skip on fork PRs since they can't mint the OIDC token used for Claude API auth
if: github.event_name != 'pull_request' || github.event.pull_request.head.repo.full_name == github.repository
runs-on: ubuntu-24.04-firewall
steps:
- uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
@@ -122,6 +135,7 @@ jobs:
anthropic_federation_rule_id: ${{ vars.ANTHROPIC_FEDERATION_RULE_ID }}
anthropic_organization_id: ${{ vars.ANTHROPIC_ORGANIZATION_ID }}
anthropic_service_account_id: ${{ vars.ANTHROPIC_SERVICE_ACCOUNT_ID }}
claude_args: "--permission-mode auto"
settings: "test-settings.json"
- name: Verify echo worked
@@ -155,7 +169,9 @@ jobs:
fi
test-settings-file-deny:
runs-on: ubuntu-latest
# Skip on fork PRs since they can't mint the OIDC token used for Claude API auth
if: github.event_name != 'pull_request' || github.event.pull_request.head.repo.full_name == github.repository
runs-on: ubuntu-24.04-firewall
steps:
- uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
@@ -178,6 +194,7 @@ jobs:
anthropic_federation_rule_id: ${{ vars.ANTHROPIC_FEDERATION_RULE_ID }}
anthropic_organization_id: ${{ vars.ANTHROPIC_ORGANIZATION_ID }}
anthropic_service_account_id: ${{ vars.ANTHROPIC_SERVICE_ACCOUNT_ID }}
claude_args: "--permission-mode auto"
settings: "test-settings.json"
- name: Verify echo was denied
+34 -18
View File
@@ -12,10 +12,17 @@ permissions:
contents: read
id-token: write
# Pin the model so these tests don't depend on the CLI's default model,
# which can change with each Claude Code release.
env:
ANTHROPIC_MODEL: claude-opus-5
jobs:
test-basic-types:
name: Test Basic Type Conversions
runs-on: ubuntu-latest
# Skip on fork PRs since they can't mint the OIDC token used for Claude API auth
if: github.event_name != 'pull_request' || github.event.pull_request.head.repo.full_name == github.repository
runs-on: ubuntu-24.04-firewall
steps:
- name: Checkout
uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
@@ -36,14 +43,14 @@ jobs:
anthropic_organization_id: ${{ vars.ANTHROPIC_ORGANIZATION_ID }}
anthropic_service_account_id: ${{ vars.ANTHROPIC_SERVICE_ACCOUNT_ID }}
claude_args: |
--permission-mode auto
--allowedTools Bash
--json-schema '{"type":"object","properties":{"text_field":{"type":"string"},"number_field":{"type":"number"},"boolean_true":{"type":"boolean"},"boolean_false":{"type":"boolean"}},"required":["text_field","number_field","boolean_true","boolean_false"]}'
- name: Verify outputs
env:
OUTPUT: ${{ steps.test.outputs.structured_output }}
run: |
# Parse the structured_output JSON
OUTPUT='${{ steps.test.outputs.structured_output }}'
# Test string pass-through
TEXT_FIELD=$(echo "$OUTPUT" | jq -r '.text_field')
if [ "$TEXT_FIELD" != "hello" ]; then
@@ -76,7 +83,9 @@ jobs:
test-complex-types:
name: Test Arrays and Objects
runs-on: ubuntu-latest
# Skip on fork PRs since they can't mint the OIDC token used for Claude API auth
if: github.event_name != 'pull_request' || github.event.pull_request.head.repo.full_name == github.repository
runs-on: ubuntu-24.04-firewall
steps:
- name: Checkout
uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
@@ -96,14 +105,14 @@ jobs:
anthropic_organization_id: ${{ vars.ANTHROPIC_ORGANIZATION_ID }}
anthropic_service_account_id: ${{ vars.ANTHROPIC_SERVICE_ACCOUNT_ID }}
claude_args: |
--permission-mode auto
--allowedTools Bash
--json-schema '{"type":"object","properties":{"items":{"type":"array","items":{"type":"string"}},"config":{"type":"object"},"empty_array":{"type":"array"}},"required":["items","config","empty_array"]}'
- name: Verify JSON stringification
env:
OUTPUT: ${{ steps.test.outputs.structured_output }}
run: |
# Parse the structured_output JSON
OUTPUT='${{ steps.test.outputs.structured_output }}'
# Arrays should be JSON stringified
if ! echo "$OUTPUT" | jq -e '.items | length == 3' > /dev/null; then
echo "❌ Array not properly formatted"
@@ -129,7 +138,9 @@ jobs:
test-edge-cases:
name: Test Edge Cases
runs-on: ubuntu-latest
# Skip on fork PRs since they can't mint the OIDC token used for Claude API auth
if: github.event_name != 'pull_request' || github.event.pull_request.head.repo.full_name == github.repository
runs-on: ubuntu-24.04-firewall
steps:
- name: Checkout
uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
@@ -150,14 +161,14 @@ jobs:
anthropic_organization_id: ${{ vars.ANTHROPIC_ORGANIZATION_ID }}
anthropic_service_account_id: ${{ vars.ANTHROPIC_SERVICE_ACCOUNT_ID }}
claude_args: |
--permission-mode auto
--allowedTools Bash
--json-schema '{"type":"object","properties":{"zero":{"type":"number"},"empty_string":{"type":"string"},"negative":{"type":"number"},"decimal":{"type":"number"}},"required":["zero","empty_string","negative","decimal"]}'
- name: Verify edge cases
env:
OUTPUT: ${{ steps.test.outputs.structured_output }}
run: |
# Parse the structured_output JSON
OUTPUT='${{ steps.test.outputs.structured_output }}'
# Zero should be "0", not empty or falsy
ZERO=$(echo "$OUTPUT" | jq -r '.zero')
if [ "$ZERO" != "0" ]; then
@@ -190,7 +201,9 @@ jobs:
test-name-sanitization:
name: Test Output Name Sanitization
runs-on: ubuntu-latest
# Skip on fork PRs since they can't mint the OIDC token used for Claude API auth
if: github.event_name != 'pull_request' || github.event.pull_request.head.repo.full_name == github.repository
runs-on: ubuntu-24.04-firewall
steps:
- name: Checkout
uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
@@ -206,14 +219,14 @@ jobs:
anthropic_organization_id: ${{ vars.ANTHROPIC_ORGANIZATION_ID }}
anthropic_service_account_id: ${{ vars.ANTHROPIC_SERVICE_ACCOUNT_ID }}
claude_args: |
--permission-mode auto
--allowedTools Bash
--json-schema '{"type":"object","properties":{"test-result":{"type":"string"},"item_count":{"type":"number"}},"required":["test-result","item_count"]}'
- name: Verify sanitized names work
env:
OUTPUT: ${{ steps.test.outputs.structured_output }}
run: |
# Parse the structured_output JSON
OUTPUT='${{ steps.test.outputs.structured_output }}'
# Hyphens should be preserved in the JSON
TEST_RESULT=$(echo "$OUTPUT" | jq -r '.["test-result"]')
if [ "$TEST_RESULT" != "passed" ]; then
@@ -232,7 +245,9 @@ jobs:
test-execution-file-structure:
name: Test Execution File Format
runs-on: ubuntu-latest
# Skip on fork PRs since they can't mint the OIDC token used for Claude API auth
if: github.event_name != 'pull_request' || github.event.pull_request.head.repo.full_name == github.repository
runs-on: ubuntu-24.04-firewall
steps:
- name: Checkout
uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
@@ -246,6 +261,7 @@ jobs:
anthropic_organization_id: ${{ vars.ANTHROPIC_ORGANIZATION_ID }}
anthropic_service_account_id: ${{ vars.ANTHROPIC_SERVICE_ACCOUNT_ID }}
claude_args: |
--permission-mode auto
--allowedTools Bash
--json-schema '{"type":"object","properties":{"done":{"type":"boolean"}},"required":["done"]}'
@@ -284,7 +300,7 @@ jobs:
- test-edge-cases
- test-name-sanitization
- test-execution-file-structure
if: always()
if: ${{ always() && (github.event_name != 'pull_request' || github.event.pull_request.head.repo.full_name == github.repository) }}
steps:
- name: Generate Summary
run: |
+23
View File
@@ -0,0 +1,23 @@
# Fails when a workflow job that calls Claude lacks its security settings.
# The rules are in CLAUDE.md, "Security hardening for GitHub Actions".
name: Security check for workflows that call Claude
on:
pull_request:
push:
branches: [main]
permissions:
contents: read
jobs:
workflow-hardening:
name: Check security settings
runs-on: ubuntu-latest
timeout-minutes: 5
steps:
- name: Checkout repository
uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
- name: Check security settings of workflows that call Claude
run: python3 .github/scripts/check_workflow_hardening.py
+4
View File
@@ -1,4 +1,8 @@
.DS_Store
node_modules
scripts/.push-branch
scripts/.no-hooks
scripts/.restored-paths
scripts/.git-repo
**/.claude/settings.local.json
+3 -1
View File
@@ -1,2 +1,4 @@
# Test fixtures should not be formatted to preserve exact output matching
test/fixtures/
test/fixtures/
# Snapshot of PR-authored config kept for review; do not reformat
.claude-pr/
+14
View File
@@ -42,3 +42,17 @@ Single entrypoint: `src/entrypoints/run.ts` orchestrates everything — prepare
- `moduleResolution: "bundler"` — imports don't need `.js` extensions.
- GitHub API calls should use retry logic (`src/utils/retry.ts`).
- MCP servers are auto-installed at runtime to `~/.claude/mcp/github-{type}-server/`.
## Security hardening for GitHub Actions
Workflow jobs in this repository that call Claude run with three protections. Keep them when you add or edit a workflow.
1. **Egress-firewall runner.** The job has `runs-on: ubuntu-24.04-firewall`, a GitHub-hosted runner that filters the job's outbound network traffic. Do not move a job that calls Claude to another runner.
2. **Network allow list.** `.github/egress-firewall.yaml` lists the hosts those jobs may reach, besides any that GitHub's firewall allows by default. Keep `mode: enforce`, which is what makes the firewall block the rest. Follow that file's header when you add a host.
3. **Auto permission mode.** Every step that runs the Claude Code action (`uses: anthropics/claude-code-action`, or this repository's own `./` and `./base-action`) passes `--permission-mode auto` in `claude_args`. A tool call that needs permission and that the allowed tools do not cover then runs only if Claude Code's safety review passes it. Allow only the tools the job needs, and keep any `--disallowedTools` list a step has. Use `claude-opus-4-6` or a newer model: on an older one Claude Code falls back to its default permission mode.
The action starts sessions in auto mode by default (`src/modes/permission-mode.ts`), except with `allowed_non_write_users`, `allowed_bots`, a custom `ANTHROPIC_BASE_URL`, `path_to_claude_code_executable`, or a `settings` input that sets a base URL or another mode or is a file path. Workflows here still pass the flag, so that the check below can see it and a change to the default can't weaken them.
`.github/workflows/workflow-hardening.yml` fails when a job that runs the Claude Code action or mentions `ANTHROPIC_FEDERATION_RULE_ID` breaks protection 1 or 3, or when the allow list is missing, empty, not `mode: enforce`, or names a host with `*`. It cannot see a job that calls Claude another way, so check new workflows by hand too. If a job cannot meet protection 1 or 3, add it with the reason to the matching exemption table in `.github/scripts/check_workflow_hardening.py`. A job in `EXEMPT_FROM_AUTO_MODE` must set no permission mode at all. Do not skip or weaken the check.
Keep each workflow's `permissions:` block minimal, and never print tokens or environment variables in workflow logs.
+17 -1
View File
@@ -79,10 +79,26 @@ Thank you for your interest in contributing to Claude Code Action! This document
git push origin feature/your-feature-name
```
5. Ensure all CI checks pass
5. Ensure all CI checks pass. On a pull request from a fork, checks usually start only after a maintainer approves them, and jobs that call Claude are skipped.
6. Request review from maintainers
### Breaking changes
Add the `breaking-change` label if a workflow that works today could fail or behave differently after upgrading, with no change on the workflow's side. Workflows on `@v1` get each release straight away, so this is their only warning.
Labeled pull requests go in their own section of the release notes, and the release title says `(breaking changes)`. The notes show only the pull request's title, so make it say what changes.
### For Maintainers: Pull Requests from Outside Contributors
These steps are for pull requests from forks whose author does not have write access.
1. Review the changes, including any to workflow files and test scripts.
2. Select **Approve and run**, which starts every waiting workflow. `ci.yml` runs the tests, type check and format check. GitHub gives these runs a read-only token and no secrets. New commits can need a new approval.
3. Check the result. Jobs that call Claude, such as `claude-review.yml` and the ones in `test-*.yml`, show as skipped, which doesn't mean they passed. The `test-*.yml` jobs run on `main` after the merge.
Mentioning `@claude` on a pull request from a fork does nothing in this repository: `claude.yml` skips its Claude job. Don't work around that by pushing the fork's commits to a branch here, because Claude then trusts them as yours. See [Running code from pull requests](./docs/security.md#running-code-from-pull-requests).
## Action Development
### Testing Your Changes
+68 -15
View File
@@ -17,7 +17,7 @@ inputs:
required: false
default: "claude"
base_branch:
description: "The branch to use as the base/source when creating new branches (defaults to repository default branch)"
description: "The branch to use as the base/source when creating new branches (defaults to repository default branch). On pull requests from authors without write access that don't target the default branch, Claude's config also comes from it, so don't set it to a value from the pull request"
required: false
branch_prefix:
description: "The prefix to use for Claude branches (defaults to 'claude/', use 'claude-' for dash format)"
@@ -44,6 +44,14 @@ inputs:
only use for workflows with very limited permissions and validate all outputs.
required: false
default: ""
allow_fork_prs:
description: "Whether to run on pull requests from forks, and on workflow_run events for a fork's run. With any value other than 'true', the action skips them, whoever opened them, before it checks anything out. See docs/security.md."
required: false
default: "true"
restrict_untrusted_prs:
description: "On pull requests whose author does not have write access, deny commands that commonly run code from the checkout (package managers, test runners, interpreters, shells), even if claude_args allows them. Set to 'false' to turn this off. See docs/security.md."
required: false
default: "true"
include_comments_by_actor:
description: "Comma-separated list of actor usernames to INCLUDE in comments. Supports wildcards: '*[bot]' matches all bots, 'dependabot[bot]' matches specific bot. Empty (default) includes all actors."
required: false
@@ -134,7 +142,7 @@ inputs:
required: false
default: "claude[bot]"
track_progress:
description: "Force tag mode with tracking comments for pull_request and issue events. Only applicable to pull_request (opened, synchronize, ready_for_review, reopened) and issue (opened, edited, labeled, assigned) events."
description: "Force tag mode with tracking comments for pull_request and issue events. Only applicable to pull_request (opened, synchronize, ready_for_review, reopened, labeled) and issue (opened, edited, labeled, assigned) events."
required: false
default: "false"
include_fix_links:
@@ -158,7 +166,7 @@ inputs:
required: false
default: "false"
plugins:
description: "Newline-separated list of Claude Code plugin names to install (e.g., 'code-review@claude-code-plugins\nfeature-dev@claude-code-plugins')"
description: "Newline-separated list of Claude Code plugin names to install (e.g., 'code-review@claude-plugins-official\nfeature-dev@claude-plugins-official')"
required: false
default: ""
plugin_marketplaces:
@@ -167,6 +175,9 @@ inputs:
default: ""
outputs:
conclusion:
description: "Execution status of Claude Code ('success' or 'failure')"
value: ${{ steps.run.outputs.conclusion }}
execution_file:
description: "Path to the Claude Code execution output file"
value: ${{ steps.run.outputs.execution_file }}
@@ -187,11 +198,20 @@ runs:
using: "composite"
steps:
- name: Install Bun
id: setup-bun
if: inputs.path_to_bun_executable == ''
uses: oven-sh/setup-bun@0c5077e51419868618aeaa5fe8019c62421857d6 # https://github.com/oven-sh/setup-bun/releases/tag/v2.2.0
with:
bun-version: 1.3.14
bun-version: 1.4.4
token: ${{ inputs.github_token || github.token }}
# Disable setup-bun's cache. The upstream save step uses a deterministic
# key (Bun version) and isn't ref-aware: on every second-and-subsequent
# run against the same PR ref the GitHub cache API rejects the duplicate
# key+ref with a 409 (HTML body), and @actions/cache treats the unparsable
# response as transient and burns ~20-30s on 5 retries before warning.
# The 35 MB Bun binary downloads in 2-3s, so disabling the cache is a net
# wallclock win and removes the noisy warning. See issue #1252.
no-cache: true
- name: Setup Custom Bun Path
if: inputs.path_to_bun_executable != ''
@@ -238,11 +258,20 @@ runs:
if: ${{ inputs.allowed_non_write_users != '' }}
continue-on-error: true
shell: bash
env:
PATH_TO_BUN_EXECUTABLE: ${{ inputs.path_to_bun_executable }}
SETUP_BUN_PATH: ${{ steps.setup-bun.outputs.bun-path }}
run: |
# Keep a copy of the bun binary alongside the action's own files so
# post-steps use the same version that was on PATH at action start.
# post-steps use the same version the action installed or was given.
mkdir -p "$GITHUB_ACTION_PATH/bin"
cp "$(command -v bun)" "$GITHUB_ACTION_PATH/bin/bun"
for bun_path in "$PATH_TO_BUN_EXECUTABLE" "$SETUP_BUN_PATH" "$(command -v bun || true)"; do
if [ -n "$bun_path" ] && [ -x "$bun_path" ]; then
cp "$bun_path" "$GITHUB_ACTION_PATH/bin/bun"
break
fi
done
test -x "$GITHUB_ACTION_PATH/bin/bun"
- name: Prepend system bin dirs to PATH
if: ${{ inputs.allowed_non_write_users != '' && runner.os != 'Windows' }}
@@ -256,11 +285,25 @@ runs:
id: run
shell: bash
run: |
bun --no-env-file \
--config="${GITHUB_ACTION_PATH}/bunfig.toml" \
--tsconfig-override="${GITHUB_ACTION_PATH}/tsconfig.json" \
run ${GITHUB_ACTION_PATH}/src/entrypoints/run.ts
# Keep these options wherever the action runs its own scripts, so that Bun
# takes its settings from the action's directory. Bun before 1.4.3 logs
# "directory mismatch" for --tsconfig-override (oven-sh/bun#25730).
#
# Bash reads this script as it runs. The braces and exit make it read the
# rest first and stop there. With exec instead, cancelling the run would
# interrupt Bun at once, before it can clean up.
{
bun --no-env-file \
--config="${GITHUB_ACTION_PATH}/bunfig.toml" \
--tsconfig-override="${GITHUB_ACTION_PATH}/tsconfig.json" \
run ${GITHUB_ACTION_PATH}/src/entrypoints/run.ts
exit
}
env:
# Turns off Bun's cache of transpiled files, which is outside the action's
# directory. Set it wherever the action runs its own scripts.
BUN_RUNTIME_TRANSPILER_CACHE_PATH: "0"
# Prepare inputs
MODE: ${{ inputs.mode }}
PROMPT: ${{ inputs.prompt }}
@@ -273,6 +316,8 @@ runs:
OVERRIDE_GITHUB_TOKEN: ${{ inputs.github_token }}
ALLOWED_BOTS: ${{ inputs.allowed_bots }}
ALLOWED_NON_WRITE_USERS: ${{ inputs.allowed_non_write_users }}
ALLOW_FORK_PRS: ${{ inputs.allow_fork_prs }}
RESTRICT_UNTRUSTED_PRS: ${{ inputs.restrict_untrusted_prs }}
CLAUDE_CODE_SUBPROCESS_ENV_SCRUB: ${{ env.CLAUDE_CODE_SUBPROCESS_ENV_SCRUB || (inputs.allowed_non_write_users != '' && '1') || '' }}
CLAUDE_CODE_SCRIPT_CAPS: ${{ env.CLAUDE_CODE_SCRIPT_CAPS || '' }}
INCLUDE_COMMENTS_BY_ACTOR: ${{ inputs.include_comments_by_actor }}
@@ -307,8 +352,8 @@ runs:
NODE_VERSION: ${{ env.NODE_VERSION }}
# Provider configuration
ANTHROPIC_API_KEY: ${{ inputs.anthropic_api_key }}
CLAUDE_CODE_OAUTH_TOKEN: ${{ inputs.claude_code_oauth_token }}
ANTHROPIC_API_KEY: ${{ inputs.anthropic_api_key || env.ANTHROPIC_API_KEY }}
CLAUDE_CODE_OAUTH_TOKEN: ${{ inputs.claude_code_oauth_token || env.CLAUDE_CODE_OAUTH_TOKEN }}
ANTHROPIC_FEDERATION_RULE_ID: ${{ inputs.anthropic_federation_rule_id }}
ANTHROPIC_ORGANIZATION_ID: ${{ inputs.anthropic_organization_id }}
ANTHROPIC_SERVICE_ACCOUNT_ID: ${{ inputs.anthropic_service_account_id }}
@@ -395,6 +440,8 @@ runs:
- name: Cleanup SSH signing key
if: always() && inputs.ssh_signing_key != ''
shell: bash
env:
BUN_RUNTIME_TRANSPILER_CACHE_PATH: "0"
run: |
BUN_BIN="${GITHUB_ACTION_PATH}/bin/bun"
[ -x "$BUN_BIN" ] || BUN_BIN="bun"
@@ -407,6 +454,7 @@ runs:
if: always() && inputs.classify_inline_comments != 'false'
shell: bash
env:
BUN_RUNTIME_TRANSPILER_CACHE_PATH: "0"
GITHUB_TOKEN: ${{ steps.run.outputs.github_token || inputs.github_token || github.token }}
REPO_OWNER: ${{ github.event.repository.owner.login }}
REPO_NAME: ${{ github.event.repository.name }}
@@ -423,10 +471,15 @@ runs:
- name: Revoke app token
if: always() && inputs.github_token == '' && steps.run.outputs.github_token != '' && steps.run.outputs.skipped_due_to_workflow_validation_mismatch != 'true'
shell: bash
env:
APP_TOKEN: ${{ steps.run.outputs.github_token }}
run: |
curl -L \
# -q comes first and skips ~/.curlrc, which a session can write.
curl -q -L \
--connect-timeout 5 \
--max-time 10 \
-X DELETE \
-H "Accept: application/vnd.github+json" \
-H "Authorization: Bearer ${{ steps.run.outputs.github_token }}" \
-H "Authorization: Bearer $APP_TOKEN" \
-H "X-GitHub-Api-Version: 2022-11-28" \
${GITHUB_API_URL:-https://api.github.com}/installation/token
${GITHUB_API_URL:-https://api.github.com}/installation/token || true
+123
View File
@@ -0,0 +1,123 @@
# Agent Approval Check
Require **N human approvals** on any pull request that contains commits
authored by an AI agent (Claude, Claude Code, or any bot identity you
configure). PRs without agent activity are unaffected.
This is the same gate Anthropic runs internally on every agent-authored PR.
## What it does
When a PR is opened, pushed to, or commented on, this action:
1. Scans the PR's commits, author, and reviews for the configured agent
identities (committer email, bot login, or an `APPROVED` review from a
bot). If none are found it posts `success: No agent activity` and stops.
2. Counts distinct human approvals: the latest `APPROVED` review per login,
plus any `/approve <head-sha>` comment whose SHA matches the current
head. Only users with write access to the repo count (verified per-user
via the collaborators permission API); agent and excluded-bot logins
never count.
3. Posts an `agent-approval-check` commit status (`success` once the count
reaches `required_approvals`, otherwise `pending`) and a sticky PR
comment explaining what's still needed.
4. Re-evaluates on every new push or comment. A push moves the head SHA,
so earlier `/approve <old-sha>` comments are flagged stale. Approving
reviews still count toward the threshold — they're picked up the next
time the workflow runs (on push or `/approve`); they just don't trigger
a run on their own.
Mark `agent-approval-check` as a **required status check** on your protected
branches and GitHub will refuse to merge until it's green.
## Setup
Copy [`examples/agent-approval-check.yml`](../examples/agent-approval-check.yml)
into `.github/workflows/` in your repo, then add `agent-approval-check` to the
required status checks on your protected branch.
This action is designed to run **alongside** GitHub's native branch
protection, not replace it. On the same protected branch you should also:
1. Require at least 1 approving review from someone with write access.
2. Enable **Dismiss stale pull request approvals when new commits are pushed**.
```yaml
name: agent-approval-check
on:
pull_request_target:
types: [opened, synchronize, reopened, ready_for_review]
issue_comment:
types: [created]
permissions:
contents: read
pull-requests: write
statuses: write
jobs:
check:
if: github.event_name != 'issue_comment' || github.event.issue.pull_request
runs-on: ubuntu-latest
steps:
- uses: anthropics/claude-code-action/agent-approval-check@main
with:
required_approvals: 2
agent_emails: [email protected]
agent_logins: claude[bot],claude-code[bot]
```
## Inputs
| Input | Default | Meaning |
| ---------------------- | ------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------- |
| `required_approvals` | `2` | Distinct human approvals needed. |
| `agent_emails` | `[email protected]` | Committer emails that mark a commit agent-authored. |
| `agent_logins` | `claude[bot],claude-code[bot]` | Logins treated as agents (PR author or approving reviewer). |
| `excluded_approvers` | _(empty)_ | Logins whose approvals never count. |
| `exempt_head_branches` | _(empty)_ | Head-branch globs that auto-pass. ⚠️ Leave empty — branch names are attacker-controlled, so this is not a safe place to encode trust. |
| `exempt_path_prefixes` | _(empty)_ | PRs touching only these prefixes auto-pass. |
| `protected_bases` | _(default branch)_ | Base branches this check gates (see threat model). |
| `config_file` | _(empty)_ | Path to an [agent-identities YAML](./agent-identities.example.yaml) replacing the inline inputs. See the warning below. |
| `docs_url` | this README | Link in the PR comment footer. |
| `github_token` | `${{ github.token }}` | Needs `statuses:write` + `pull-requests:write`. |
> ⚠️ **`config_file` and checkout:** if you set `config_file`, your workflow
> must check out the **base** branch to read it (the default behaviour of
> `actions/checkout` under `pull_request_target`). Never check out the PR
> head ref — doing so would let the PR author control the config and bypass
> this check.
## Approving
A human counts as an approver by either:
- submitting a normal GitHub **Approve** review, or
- commenting `/approve <sha>` where `<sha>` is the current head commit
(12–40 hex chars). This path lets the PR author — who can't approve their
own PR in GitHub's UI — vouch for commits an agent pushed on their behalf.
The author's `/approve` is subject to the same write-access verification
as any other approver, so a fork-PR author without write access on the
base repository cannot self-count. The author counts as **one** approval;
the remaining approvals must come from other reviewers with write access.
## Threat model
- **Tamper-proof triggers.** `pull_request_target` and `issue_comment` run
the workflow file from the base/default branch, so the PR under review
cannot edit this check. `pull_request_review` does **not** share this
property — it runs from the merge ref — so the example workflow omits it;
native Approve reviews are picked up on the next synchronize or
`/approve` comment. This tamper-resistance assumes the workflow file
itself is protected: an actor who can push workflow changes to the
default branch can spoof any required status check, including this one,
so protect `.github/workflows/` via branch protection or CODEOWNERS.
- **Fail-closed.** Any unhandled error exits non-zero; the required status
stays non-success and the PR stays blocked. PRs with >100 commits are
treated as agent-authored because the full commit list can't be verified.
- **Sibling-PR guard.** Commit statuses attach to a SHA, not a PR. The
action refuses to post a status on a PR whose base isn't in
`protected_bases`, and withholds `success` while another open PR to a
protected base shares the same head commit — otherwise a green status on
one PR would also unblock the other.
- **No checkout of PR code.** The action never checks out the PR's branch;
it reads PR metadata via the GitHub API, so the usual
`pull_request_target` code-execution risk does not apply.
+72
View File
@@ -0,0 +1,72 @@
name: Agent Approval Check
description: |
Require N human approvals on PRs that contain agent-authored commits
(Claude, Claude Code, or any configured bot identity). Posts an
`agent-approval-check` commit status — mark it as a required check on
protected branches to gate merges.
inputs:
github_token:
description: Token with statuses:write and pull-requests:write on this repo.
default: ${{ github.token }}
required_approvals:
description: Number of distinct human approvals required. Must be >= 1.
default: "2"
agent_emails:
description: Comma-separated committer emails treated as agent-authored.
default: [email protected]
agent_logins:
description: |
Comma-separated GitHub logins treated as agents — a PR opened by, or an
APPROVED review from, one of these triggers the check.
default: claude[bot],claude-code[bot]
excluded_approvers:
description: Comma-separated logins whose approvals never count (e.g. rubber-stamp bots).
default: ""
exempt_head_branches:
description: |
Comma-separated glob patterns; PRs from matching head branches auto-pass.
WARNING: leave empty — branch names are attacker-controlled, so this is
not a safe place to encode trust.
default: ""
exempt_path_prefixes:
description: Comma-separated path prefixes; PRs touching only these auto-pass.
default: ""
protected_bases:
description: |
Comma-separated base branches this check gates. Empty = the repo's
default branch only. PRs targeting any other base are refused (no
status posted) so a sibling PR sharing the head SHA can't get the
shared commit stamped green.
default: ""
config_file:
description: Optional path to an agent-identities YAML file (overrides the inline inputs).
default: ""
docs_url:
description: Link shown in the PR comment footer.
default: "https://github.com/anthropics/claude-code-action/tree/main/agent-approval-check"
runs:
using: composite
steps:
- uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5
with:
python-version: "3.12"
- run: pip install 'httpx==0.28.1' 'pyyaml==6.0.3' 'tenacity==9.1.4'
shell: bash
- run: python "${{ github.action_path }}/agent_approval_check.py"
shell: bash
env:
GH_TOKEN: ${{ inputs.github_token }}
GH_REPOSITORY: ${{ github.repository }}
GH_EVENT_NAME: ${{ github.event_name }}
GH_EVENT_PATH: ${{ github.event_path }}
REQUIRED_APPROVALS: ${{ inputs.required_approvals }}
AGENT_EMAILS: ${{ inputs.agent_emails }}
AGENT_LOGINS: ${{ inputs.agent_logins }}
EXCLUDED_APPROVERS: ${{ inputs.excluded_approvers }}
EXEMPT_HEAD_BRANCHES: ${{ inputs.exempt_head_branches }}
EXEMPT_PATH_PREFIXES: ${{ inputs.exempt_path_prefixes }}
PROTECTED_BASES: ${{ inputs.protected_bases }}
CONFIG_FILE: ${{ inputs.config_file }}
DOCS_URL: ${{ inputs.docs_url }}
@@ -0,0 +1,33 @@
---
# Optional config-file form of the agent-approval-check inputs.
# Pass via `with: { config_file: .github/agent-identities.yaml }` instead of
# the inline `agent_emails` / `agent_logins` / … inputs.
# Committer emails that mark a commit as agent-authored.
agent_emails:
- [email protected]
# GitHub logins treated as agents — a PR opened by, or an APPROVED review
# from, one of these triggers the check.
agent_app_logins:
- claude[bot]
- claude-code[bot]
# Logins whose approvals never count toward the required total.
excluded_approver_logins: []
# Head-branch glob patterns that auto-pass. Leave empty: branch names are
# attacker-controlled, so this is not a safe place to encode trust.
exempt_head_branches: []
# Per-repo path prefixes whose PRs auto-pass when ONLY those paths change.
exempt_path_prefixes:
owner/repo:
- docs/
# Per-repo base branches this check gates. A repo with no entry defaults to
# its default branch only. Listing a repo here REPLACES that default.
protected_bases:
owner/repo:
exact: [main]
prefixes: [release/]
File diff suppressed because it is too large Load Diff
+117 -119
View File
@@ -22,7 +22,7 @@ Add the following to your workflow file:
uses: anthropics/claude-code-base-action@beta
with:
prompt: "Your prompt here"
allowed_tools: "Bash(git:*),View,GlobTool,GrepTool,BatchTool"
claude_args: '--allowedTools "Bash(git:*),Read,Glob,Grep"'
anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
# Or using a prompt from a file
@@ -30,7 +30,7 @@ Add the following to your workflow file:
uses: anthropics/claude-code-base-action@beta
with:
prompt_file: "/path/to/prompt.txt"
allowed_tools: "Bash(git:*),View,GlobTool,GrepTool,BatchTool"
claude_args: '--allowedTools "Bash(git:*),Read,Glob,Grep"'
anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
# Or limiting the conversation turns
@@ -38,8 +38,9 @@ Add the following to your workflow file:
uses: anthropics/claude-code-base-action@beta
with:
prompt: "Your prompt here"
allowed_tools: "Bash(git:*),View,GlobTool,GrepTool,BatchTool"
max_turns: "5" # Limit conversation to 5 turns
claude_args: |
--allowedTools "Bash(git:*),Read,Glob,Grep"
--max-turns 5
anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
# Using custom system prompts
@@ -47,8 +48,9 @@ Add the following to your workflow file:
uses: anthropics/claude-code-base-action@beta
with:
prompt: "Build a REST API"
system_prompt: "You are a senior backend engineer. Focus on security, performance, and maintainability."
allowed_tools: "Bash(git:*),View,GlobTool,GrepTool,BatchTool"
claude_args: |
--system-prompt "You are a senior backend engineer. Focus on security, performance, and maintainability."
--allowedTools "Bash(git:*),Read,Glob,Grep"
anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
# Or appending to the default system prompt
@@ -56,8 +58,9 @@ Add the following to your workflow file:
uses: anthropics/claude-code-base-action@beta
with:
prompt: "Create a database schema"
append_system_prompt: "After writing code, be sure to code review yourself."
allowed_tools: "Bash(git:*),View,GlobTool,GrepTool,BatchTool"
claude_args: |
--append-system-prompt "After writing code, be sure to code review yourself."
--allowedTools "Bash(git:*),Read,Glob,Grep"
anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
# Using custom environment variables
@@ -65,11 +68,15 @@ Add the following to your workflow file:
uses: anthropics/claude-code-base-action@beta
with:
prompt: "Deploy to staging environment"
claude_env: |
ENVIRONMENT: staging
API_URL: https://api-staging.example.com
DEBUG: true
allowed_tools: "Bash(git:*),View,GlobTool,GrepTool,BatchTool"
settings: |
{
"env": {
"ENVIRONMENT": "staging",
"API_URL": "https://api-staging.example.com",
"DEBUG": "true"
}
}
claude_args: '--allowedTools "Bash(git:*),Read,Glob,Grep"'
anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
# Using fallback model for handling API errors
@@ -77,9 +84,10 @@ Add the following to your workflow file:
uses: anthropics/claude-code-base-action@beta
with:
prompt: "Review and fix TypeScript errors"
model: "claude-opus-4-1-20250805"
fallback_model: "claude-sonnet-4-20250514"
allowed_tools: "Bash(git:*),View,GlobTool,GrepTool,BatchTool"
claude_args: |
--model "claude-opus-4-1-20250805"
--fallback-model "claude-sonnet-4-20250514"
--allowedTools "Bash(git:*),Read,Glob,Grep"
anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
# Using OAuth token instead of API key
@@ -87,7 +95,7 @@ Add the following to your workflow file:
uses: anthropics/claude-code-base-action@beta
with:
prompt: "Update dependencies"
allowed_tools: "Bash(git:*),View,GlobTool,GrepTool,BatchTool"
claude_args: '--allowedTools "Bash(git:*),Read,Glob,Grep"'
claude_code_oauth_token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}
```
@@ -114,32 +122,28 @@ Do not set `anthropic_api_key` or `claude_code_oauth_token` alongside the federa
## Inputs
| Input | Description | Required | Default |
| ------------------------------ | ----------------------------------------------------------------------------------------------------------------------- | -------- | ---------------------------- |
| `prompt` | The prompt to send to Claude Code | No\* | '' |
| `prompt_file` | Path to a file containing the prompt to send to Claude Code | No\* | '' |
| `allowed_tools` | Comma-separated list of allowed tools for Claude Code to use | No | '' |
| `disallowed_tools` | Comma-separated list of disallowed tools that Claude Code cannot use | No | '' |
| `max_turns` | Maximum number of conversation turns (default: no limit) | No | '' |
| `mcp_config` | Path to the MCP configuration JSON file, or MCP configuration JSON string | No | '' |
| `settings` | Path to Claude Code settings JSON file, or settings JSON string | No | '' |
| `system_prompt` | Override system prompt | No | '' |
| `append_system_prompt` | Append to system prompt | No | '' |
| `claude_env` | Custom environment variables to pass to Claude Code execution (YAML multiline format) | No | '' |
| `model` | Model to use (provider-specific format required for Bedrock/Vertex) | No | 'claude-4-0-sonnet-20250219' |
| `anthropic_model` | DEPRECATED: Use 'model' instead | No | 'claude-4-0-sonnet-20250219' |
| `fallback_model` | Enable automatic fallback to specified model when default model is overloaded | No | '' |
| `anthropic_api_key` | Anthropic API key (required for direct Anthropic API) | No | '' |
| `claude_code_oauth_token` | Claude Code OAuth token (alternative to anthropic_api_key) | No | '' |
| `anthropic_federation_rule_id` | Workload identity federation rule ID (fdrl\_...). Requires `id-token: write` permission | No | '' |
| `anthropic_organization_id` | Anthropic organization UUID used for workload identity federation | No | '' |
| `anthropic_service_account_id` | Service account ID (svac\_...) the federated token acts as (optional) | No | '' |
| `anthropic_workspace_id` | Workspace ID (wrkspc\_...) for federation. Optional when the rule targets a single workspace | No | '' |
| `anthropic_oidc_audience` | Audience to request on the GitHub OIDC token. Defaults to https://api.anthropic.com | No | '' |
| `use_bedrock` | Use Amazon Bedrock with OIDC authentication instead of direct Anthropic API | No | 'false' |
| `use_vertex` | Use Google Vertex AI with OIDC authentication instead of direct Anthropic API | No | 'false' |
| `use_node_cache` | Whether to use Node.js dependency caching (set to true only for Node.js projects with lock files) | No | 'false' |
| `show_full_output` | Show full JSON output (⚠️ May expose secrets - see [security docs](../docs/security.md#️-full-output-security-warning)) | No | 'false'\*\* |
| Input | Description | Required | Default |
| -------------------------------- | ----------------------------------------------------------------------------------------------------------------------- | -------- | ------------- |
| `prompt` | The prompt to send to Claude Code | No\* | `''` |
| `prompt_file` | Path to a file containing the prompt to send to Claude Code | No\* | `''` |
| `settings` | Claude Code settings as a JSON string or path to a settings JSON file | No | `''` |
| `claude_args` | Additional arguments to pass directly to the Claude CLI | No | `''` |
| `anthropic_api_key` | Anthropic API key for direct Anthropic API authentication | No | `''` |
| `claude_code_oauth_token` | Claude Code OAuth token as an alternative to an Anthropic API key | No | `''` |
| `anthropic_federation_rule_id` | Workload identity federation rule ID (fdrl\_...). Requires `id-token: write` permission | No | `''` |
| `anthropic_organization_id` | Anthropic organization UUID used for workload identity federation | No | `''` |
| `anthropic_service_account_id` | Service account ID (svac\_...) the federated token acts as | No | `''` |
| `anthropic_workspace_id` | Workspace ID (wrkspc\_...) for federation | No | `''` |
| `anthropic_oidc_audience` | Audience for the GitHub OIDC token request | No | `''` |
| `use_bedrock` | Use Amazon Bedrock with OIDC authentication | No | `'false'` |
| `use_vertex` | Use Google Vertex AI with OIDC authentication | No | `'false'` |
| `use_foundry` | Use Microsoft Foundry with OIDC authentication | No | `'false'` |
| `use_node_cache` | Enable Node.js dependency caching for projects with lock files | No | `'false'` |
| `path_to_claude_code_executable` | Path to a custom Claude Code executable | No | `''` |
| `path_to_bun_executable` | Path to a custom Bun executable | No | `''` |
| `show_full_output` | Show full JSON output (⚠️ May expose secrets - see [security docs](../docs/security.md#️-full-output-security-warning)) | No | `'false'`\*\* |
| `plugins` | Newline-separated Claude Code plugin names to install | No | `''` |
| `plugin_marketplaces` | Newline-separated plugin marketplace Git URLs to install | No | `''` |
\*Either `prompt` or `prompt_file` must be provided, but not both.
@@ -147,10 +151,12 @@ Do not set `anthropic_api_key` or `claude_code_oauth_token` alongside the federa
## Outputs
| Output | Description |
| ---------------- | ---------------------------------------------------------- |
| `conclusion` | Execution status of Claude Code ('success' or 'failure') |
| `execution_file` | Path to the JSON file containing Claude Code execution log |
| Output | Description |
| ------------------- | ------------------------------------------------------------------------------------------------- |
| `conclusion` | Execution status of Claude Code ('success' or 'failure') |
| `execution_file` | Path to the JSON file containing Claude Code execution log |
| `structured_output` | JSON string containing structured output fields when `--json-schema` is provided in `claude_args` |
| `session_id` | The Claude Code session ID that can be used with `--resume` to continue this conversation |
## Environment Variables
@@ -174,55 +180,28 @@ Example usage:
## Custom Environment Variables
You can pass custom environment variables to Claude Code execution using the `claude_env` input. This allows Claude to access environment-specific configuration during its execution.
The `claude_env` input accepts YAML multiline format with key-value pairs:
You can pass custom environment variables to Claude Code through the `env` object in `settings`:
```yaml
- name: Deploy with custom environment
uses: anthropics/claude-code-base-action@beta
with:
prompt: "Deploy the application to the staging environment"
claude_env: |
ENVIRONMENT: staging
API_BASE_URL: https://api-staging.example.com
DATABASE_URL: ${{ secrets.STAGING_DB_URL }}
DEBUG: true
LOG_LEVEL: debug
allowed_tools: "Bash(git:*),View,GlobTool,GrepTool,BatchTool"
settings: |
{
"env": {
"ENVIRONMENT": "staging",
"API_BASE_URL": "https://api-staging.example.com",
"DATABASE_URL": "${{ secrets.STAGING_DB_URL }}",
"DEBUG": "true",
"LOG_LEVEL": "debug"
}
}
claude_args: '--allowedTools "Bash(git:*),Read,Glob,Grep"'
anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
```
### Features:
- **YAML Format**: Use standard YAML key-value syntax (`KEY: value`)
- **Multiline Support**: Define multiple environment variables in a single input
- **Comments**: Lines starting with `#` are ignored
- **GitHub Secrets**: Can reference GitHub secrets using `${{ secrets.SECRET_NAME }}`
- **Runtime Access**: Environment variables are available to Claude during execution
### Example Use Cases:
```yaml
# Development configuration
claude_env: |
NODE_ENV: development
API_URL: http://localhost:3000
DEBUG: true
# Production deployment
claude_env: |
NODE_ENV: production
API_URL: https://api.example.com
DATABASE_URL: ${{ secrets.PROD_DB_URL }}
REDIS_URL: ${{ secrets.REDIS_URL }}
# Feature flags and configuration
claude_env: |
FEATURE_NEW_UI: enabled
MAX_RETRIES: 3
TIMEOUT_MS: 5000
```
The `settings` input accepts either inline JSON or a path to a settings JSON file. Values in the `env` object are available during the Claude Code session and can reference GitHub secrets.
## Using Settings Configuration
@@ -238,7 +217,7 @@ Provide a path to a JSON file containing Claude Code settings:
with:
prompt: "Your prompt here"
settings: "path/to/settings.json"
allowed_tools: "Bash(git:*),View,GlobTool,GrepTool,BatchTool"
claude_args: '--allowedTools "Bash(git:*),Read,Glob,Grep"'
anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
```
@@ -272,7 +251,7 @@ Provide the settings configuration directly as a JSON string:
}]
}
}
allowed_tools: "Bash(git:*),View,GlobTool,GrepTool,BatchTool"
claude_args: '--allowedTools "Bash(git:*),Read,Glob,Grep"'
anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
```
@@ -300,8 +279,9 @@ Provide a path to a JSON file containing MCP configuration:
uses: anthropics/claude-code-base-action@beta
with:
prompt: "Your prompt here"
mcp_config: "path/to/mcp-config.json"
allowed_tools: "Bash(git:*),View,GlobTool,GrepTool,BatchTool"
claude_args: |
--mcp-config "path/to/mcp-config.json"
--allowedTools "Bash(git:*),Read,Glob,Grep"
anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
```
@@ -314,19 +294,9 @@ Provide the MCP configuration directly as a JSON string:
uses: anthropics/claude-code-base-action@beta
with:
prompt: "Your prompt here"
mcp_config: |
{
"mcpServers": {
"server-name": {
"command": "node",
"args": ["./server.js"],
"env": {
"API_KEY": "your-api-key"
}
}
}
}
allowed_tools: "Bash(git:*),View,GlobTool,GrepTool,BatchTool"
claude_args: >-
--mcp-config '{"mcpServers":{"server-name":{"command":"node","args":["./server.js"],"env":{"API_KEY":"your-api-key"}}}}'
--allowedTools "Bash(git:*),Read,Glob,Grep"
anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
```
@@ -354,8 +324,9 @@ You can combine MCP config with other inputs like allowed tools:
uses: anthropics/claude-code-base-action@beta
with:
prompt: "Access the custom MCP server and use its tools"
mcp_config: "mcp-config.json"
allowed_tools: "Bash(git:*),View,mcp__server-name__custom_tool"
claude_args: |
--mcp-config "mcp-config.json"
--allowedTools "Bash(git:*),Read,mcp__server-name__custom_tool"
anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
```
@@ -382,7 +353,7 @@ jobs:
uses: anthropics/claude-code-base-action@beta
with:
prompt: "Review the PR changes. Focus on code quality, potential bugs, and performance issues. Suggest improvements where appropriate. Write your review as markdown text."
allowed_tools: "Bash(git diff --name-only HEAD~1),Bash(git diff HEAD~1),View,GlobTool,GrepTool,Write"
claude_args: '--allowedTools "Bash(git diff --name-only HEAD~1),Bash(git diff HEAD~1),Read,Glob,Grep,Write"'
anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
- name: Extract and Comment PR Review
@@ -395,18 +366,39 @@ jobs:
const executionFile = '${{ steps.code-review.outputs.execution_file }}';
const executionLog = JSON.parse(fs.readFileSync(executionFile, 'utf8'));
// Extract the review content from the execution log
// The execution log contains the full conversation including Claude's responses
// Extract the review content from the execution log.
// The SDK writes top-level events with `type`; assistant text is nested
// under `message.content`.
let review = '';
// Find the last assistant message which should contain the review
// Prefer the final result event when it is available.
for (let i = executionLog.length - 1; i >= 0; i--) {
if (executionLog[i].role === 'assistant') {
review = executionLog[i].content;
const entry = executionLog[i];
if (entry?.type === 'result' && typeof entry.result === 'string') {
review = entry.result;
break;
}
}
// Fallback to the last assistant text block if no result event was written.
if (!review) {
for (let i = executionLog.length - 1; i >= 0; i--) {
const entry = executionLog[i];
if (entry?.type !== 'assistant' || !Array.isArray(entry.message?.content)) {
continue;
}
review = entry.message.content
.filter((block) => block?.type === 'text' && typeof block.text === 'string')
.map((block) => block.text)
.join('\n');
if (review) {
break;
}
}
}
if (review) {
github.rest.issues.createComment({
issue_number: context.issue.number,
@@ -417,6 +409,10 @@ jobs:
}
```
For typed automation output, prefer passing `--json-schema` in `claude_args`
and reading `steps.<id>.outputs.structured_output` instead of parsing the full
execution log.
Check out additional examples in [`./examples`](./examples).
## Using Cloud Providers
@@ -445,7 +441,7 @@ Use provider-specific model names based on your chosen provider:
uses: anthropics/claude-code-base-action@beta
with:
prompt: "Your prompt here"
model: "claude-3-7-sonnet-20250219"
claude_args: "--model claude-3-7-sonnet-20250219"
anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
# For Amazon Bedrock (requires OIDC authentication)
@@ -459,7 +455,7 @@ Use provider-specific model names based on your chosen provider:
uses: anthropics/claude-code-base-action@beta
with:
prompt: "Your prompt here"
model: "anthropic.claude-3-7-sonnet-20250219-v1:0"
claude_args: "--model anthropic.claude-3-7-sonnet-20250219-v1:0"
use_bedrock: "true"
# For Google Vertex AI (requires OIDC authentication)
@@ -473,7 +469,7 @@ Use provider-specific model names based on your chosen provider:
uses: anthropics/claude-code-base-action@beta
with:
prompt: "Your prompt here"
model: "claude-3-7-sonnet@20250219"
claude_args: "--model claude-3-7-sonnet@20250219"
use_vertex: "true"
```
@@ -493,8 +489,9 @@ This example shows how to use OIDC authentication with AWS Bedrock:
with:
prompt: "Your prompt here"
use_bedrock: "true"
model: "anthropic.claude-3-7-sonnet-20250219-v1:0"
allowed_tools: "Bash(git:*),View,GlobTool,GrepTool,BatchTool"
claude_args: |
--model "anthropic.claude-3-7-sonnet-20250219-v1:0"
--allowedTools "Bash(git:*),Read,Glob,Grep"
```
## Example: Using OIDC Authentication for GCP Vertex AI
@@ -513,8 +510,9 @@ This example shows how to use OIDC authentication with GCP Vertex AI:
with:
prompt: "Your prompt here"
use_vertex: "true"
model: "claude-3-7-sonnet@20250219"
allowed_tools: "Bash(git:*),View,GlobTool,GrepTool,BatchTool"
claude_args: |
--model "claude-3-7-sonnet@20250219"
--allowedTools "Bash(git:*),Read,Glob,Grep"
```
## Security Best Practices
+18 -4
View File
@@ -84,7 +84,7 @@ inputs:
required: false
default: "false"
plugins:
description: "Newline-separated list of Claude Code plugin names to install (e.g., 'code-review@claude-code-plugins\nfeature-dev@claude-code-plugins')"
description: "Newline-separated list of Claude Code plugin names to install (e.g., 'code-review@claude-plugins-official\nfeature-dev@claude-plugins-official')"
required: false
default: ""
plugin_marketplaces:
@@ -120,7 +120,9 @@ runs:
if: inputs.path_to_bun_executable == ''
uses: oven-sh/setup-bun@0c5077e51419868618aeaa5fe8019c62421857d6 # https://github.com/oven-sh/setup-bun/releases/tag/v2.2.0
with:
bun-version: 1.3.14
bun-version: 1.4.4
# Disable setup-bun's cache. See action.yml for details and issue #1252.
no-cache: true
- name: Setup Custom Bun Path
if: inputs.path_to_bun_executable != ''
@@ -145,7 +147,7 @@ runs:
PATH_TO_CLAUDE_CODE_EXECUTABLE: ${{ inputs.path_to_claude_code_executable }}
run: |
if [ -z "$PATH_TO_CLAUDE_CODE_EXECUTABLE" ]; then
CLAUDE_CODE_VERSION="2.1.161"
CLAUDE_CODE_VERSION="2.1.330"
echo "Installing Claude Code v${CLAUDE_CODE_VERSION}..."
for attempt in 1 2 3; do
echo "Installation attempt $attempt..."
@@ -163,6 +165,8 @@ runs:
sleep 5
done
echo "Claude Code installed successfully"
# Add ~/.local/bin to PATH so the claude executable is available in subsequent steps
echo "$HOME/.local/bin" >> "$GITHUB_PATH"
else
echo "Using custom Claude Code executable: $PATH_TO_CLAUDE_CODE_EXECUTABLE"
# Add the directory containing the custom executable to PATH
@@ -179,8 +183,18 @@ runs:
echo "Changing directory to CLAUDE_WORKING_DIR: $CLAUDE_WORKING_DIR"
cd "$CLAUDE_WORKING_DIR"
fi
bun run ${GITHUB_ACTION_PATH}/src/index.ts
# Bash reads this script as it runs. The braces and exit make it read the
# rest first and stop there. With exec instead, cancelling the run would
# interrupt Bun at once, before it can clean up.
{
bun run ${GITHUB_ACTION_PATH}/src/index.ts
exit
}
env:
# Turns off Bun's cache of transpiled files, which is outside the action's
# directory
BUN_RUNTIME_TRANSPILER_CACHE_PATH: "0"
# Model configuration
CLAUDE_CODE_ACTION: "1"
INPUT_PROMPT: ${{ inputs.prompt }}
+19 -21
View File
@@ -6,11 +6,11 @@
"name": "@anthropic-ai/claude-code-base-action",
"dependencies": {
"@actions/core": "^1.10.1",
"@anthropic-ai/claude-agent-sdk": "^0.3.161",
"shell-quote": "^1.8.3",
"@anthropic-ai/claude-agent-sdk": "^0.3.330",
"shell-quote": "^1.8.4",
},
"devDependencies": {
"@types/bun": "^1.2.12",
"@types/bun": "^1.4.3",
"@types/node": "^20.0.0",
"@types/shell-quote": "^1.7.5",
"prettier": "3.5.3",
@@ -27,23 +27,23 @@
"@actions/io": ["@actions/[email protected]", "", {}, "sha512-wi9JjgKLYS7U/z8PPbco+PvTb/nRWjeoFlJ1Qer83k/3C5PHQi28hiVdeE2kHXmIL99mQFawx8qt/JPjZilJ8Q=="],
"@anthropic-ai/claude-agent-sdk": ["@anthropic-ai/[email protected].161", "", { "optionalDependencies": { "@anthropic-ai/claude-agent-sdk-darwin-arm64": "0.3.161", "@anthropic-ai/claude-agent-sdk-darwin-x64": "0.3.161", "@anthropic-ai/claude-agent-sdk-linux-arm64": "0.3.161", "@anthropic-ai/claude-agent-sdk-linux-arm64-musl": "0.3.161", "@anthropic-ai/claude-agent-sdk-linux-x64": "0.3.161", "@anthropic-ai/claude-agent-sdk-linux-x64-musl": "0.3.161", "@anthropic-ai/claude-agent-sdk-win32-arm64": "0.3.161", "@anthropic-ai/claude-agent-sdk-win32-x64": "0.3.161" }, "peerDependencies": { "@anthropic-ai/sdk": ">=0.93.0", "@modelcontextprotocol/sdk": "^1.29.0", "zod": "^4.0.0" } }, "sha512-TWVFmPsGePXPNnNuWAAuKagQai9kNj99tuu2KLo8FN3oWF5r82OnWTNpQ3IubpWYPyO1vC0+kQu2GfqoEme/hA=="],
"@anthropic-ai/claude-agent-sdk": ["@anthropic-ai/[email protected].330", "", { "optionalDependencies": { "@anthropic-ai/claude-agent-sdk-darwin-arm64": "0.3.330", "@anthropic-ai/claude-agent-sdk-darwin-x64": "0.3.330", "@anthropic-ai/claude-agent-sdk-linux-arm64": "0.3.330", "@anthropic-ai/claude-agent-sdk-linux-arm64-musl": "0.3.330", "@anthropic-ai/claude-agent-sdk-linux-x64": "0.3.330", "@anthropic-ai/claude-agent-sdk-linux-x64-musl": "0.3.330", "@anthropic-ai/claude-agent-sdk-win32-arm64": "0.3.330", "@anthropic-ai/claude-agent-sdk-win32-x64": "0.3.330" }, "peerDependencies": { "@anthropic-ai/sdk": ">=0.93.0", "@modelcontextprotocol/sdk": "^1.29.0", "zod": "^4.0.0" } }, "sha512-7om9Xus6tfkeCUCfzg6ilJuN6Z7okkEvJ0mx+ukXa/Id5U/6BGtPkOYni+3ebedwa3b9SVF7A05+ln+/9EoU/w=="],
"@anthropic-ai/claude-agent-sdk-darwin-arm64": ["@anthropic-ai/[email protected].161", "", { "os": "darwin", "cpu": "arm64" }, "sha512-KwDpi1O2P++uAskFt454K/sShPmhHSYhSbBxaap2KLFoVFti1pttDcEqDPfmKCk2XzKTwlpnVSI1IwTiawPJTw=="],
"@anthropic-ai/claude-agent-sdk-darwin-arm64": ["@anthropic-ai/[email protected].330", "", { "os": "darwin", "cpu": "arm64" }, "sha512-J9nqvcfD2V90O5uoR8At66R51VYtSMKnxSmPc0S4A4F5oX1VbokcpJLXU1yVCOuTtzr/OIHjRr9nn2vKi/S6rQ=="],
"@anthropic-ai/claude-agent-sdk-darwin-x64": ["@anthropic-ai/[email protected].161", "", { "os": "darwin", "cpu": "x64" }, "sha512-VoLz1BfKDc6UOYL+UhRKNSFA1k4l6kvOUrWqkHJ34V5zI5/7VzQilaGl7hU5JECE1TleXMzrWpDrLZO7akApJA=="],
"@anthropic-ai/claude-agent-sdk-darwin-x64": ["@anthropic-ai/[email protected].330", "", { "os": "darwin", "cpu": "x64" }, "sha512-woe+f9qeZYWML5fACibNyth0xMDf22riNRNDryIYuhdteh7IoWrK6zC31cC2MrZRoqDZcPVtfOLND79+iZbrLA=="],
"@anthropic-ai/claude-agent-sdk-linux-arm64": ["@anthropic-ai/[email protected].161", "", { "os": "linux", "cpu": "arm64" }, "sha512-QQbGTJ+2AFeP6+vsZQtafqeGZGX72jaHOjOyrhEcMk1T2DSrJV05J0xnkZT4yIOsByGosntTcY6963EfkuwuTg=="],
"@anthropic-ai/claude-agent-sdk-linux-arm64": ["@anthropic-ai/[email protected].330", "", { "os": "linux", "cpu": "arm64", "libc": "glibc" }, "sha512-2Q+kfzaoTvJv1rsyu+iR4UHDZ3nxiOfdXISk+2ydQShU9DqXNlgQAt3BdJMh0Yx4MrxB66TMphVvVQeRslwwhA=="],
"@anthropic-ai/claude-agent-sdk-linux-arm64-musl": ["@anthropic-ai/[email protected].161", "", { "os": "linux", "cpu": "arm64" }, "sha512-y+RPkQ6ZFje6LP0WyGvM4yHp7IQAaxAbcKSP9QVzKtT9yWiDeqWSnPzdP2Rp3B+6Kd12MuhwIMAbiKS4onAusA=="],
"@anthropic-ai/claude-agent-sdk-linux-arm64-musl": ["@anthropic-ai/[email protected].330", "", { "os": "linux", "cpu": "arm64", "libc": "musl" }, "sha512-fXmuAEegw31v+kiBd51yn7RVhjs7Rz2lnhfx/+XJXqt5lyqR2v7Ul/U/9r1CQZKuIw8M3ms9jQqGpmKbgKndqQ=="],
"@anthropic-ai/claude-agent-sdk-linux-x64": ["@anthropic-ai/[email protected].161", "", { "os": "linux", "cpu": "x64" }, "sha512-gq4tvNRRcfSf8o35mLWKN351I6FowQa2n28YiZJBvZU1glQ3HKQM4srrXzjwLwv1VcndyWBfLxoaDB1I9BS++w=="],
"@anthropic-ai/claude-agent-sdk-linux-x64": ["@anthropic-ai/[email protected].330", "", { "os": "linux", "cpu": "x64", "libc": "glibc" }, "sha512-9w4uMdA3vNZXocwOUhivOGotXoRDxN/C/xPMkZMY9WPpKLAlj8upMZskNgPjQKlVRXhxPfNagcBijGPUH4WepA=="],
"@anthropic-ai/claude-agent-sdk-linux-x64-musl": ["@anthropic-ai/[email protected].161", "", { "os": "linux", "cpu": "x64" }, "sha512-ev0eJhypbyatvSL+Cl2LqOos5+XZbgBMV6MJk7jIh/9dirMLwDj2JFlLlOnHdmIpFcznESZ/Z2wgMtqmOMUCHQ=="],
"@anthropic-ai/claude-agent-sdk-linux-x64-musl": ["@anthropic-ai/[email protected].330", "", { "os": "linux", "cpu": "x64", "libc": "musl" }, "sha512-kooZMdp6tFwNFkDXbGoNTZMgFcPQfJVfG+gZ2qAdkk11S6MFts9ybHRLQmzUK3VrqlahH7T9fmI2mNfzYgOXkw=="],
"@anthropic-ai/claude-agent-sdk-win32-arm64": ["@anthropic-ai/[email protected].161", "", { "os": "win32", "cpu": "arm64" }, "sha512-toJhVo/7RKyPc/IEGvgIcN/mPmGs38gnuDOrdtdxXoPR7WRkHcHfwprD5S9zV/pZAhuLhj/cznhno3G7b4Pnkg=="],
"@anthropic-ai/claude-agent-sdk-win32-arm64": ["@anthropic-ai/[email protected].330", "", { "os": "win32", "cpu": "arm64" }, "sha512-6jscch7vk/XBSn7mm1i6zwTeiz05q98nKrQ3HdTavQEDGvXycMcB/QxIdMrs3vrdNEjWuoxwrL1qlT1lOyAWLg=="],
"@anthropic-ai/claude-agent-sdk-win32-x64": ["@anthropic-ai/[email protected].161", "", { "os": "win32", "cpu": "x64" }, "sha512-QmYFeX/P7bt4diCZhXgazkmzLxL+uK1zwPaCSL2sIQyYGWAXiUulZ41TVLQbHglCBlKzUzVP9kwEXaAZye7LPA=="],
"@anthropic-ai/claude-agent-sdk-win32-x64": ["@anthropic-ai/[email protected].330", "", { "os": "win32", "cpu": "x64" }, "sha512-ooThWbUXlXKTDgj2VgWBXQ/6vSOTaAbBDWPvvCGNHi20haoSkw4uPX9gczkX+HJuVTslvvLiYIfe8zleVPzi7A=="],
"@anthropic-ai/sdk": ["@anthropic-ai/[email protected]", "", { "dependencies": { "json-schema-to-ts": "^3.1.1" }, "peerDependencies": { "zod": "^3.25.0 || ^4.0.0" }, "optionalPeers": ["zod"], "bin": { "anthropic-ai-sdk": "bin/cli" } }, "sha512-q9vaSZQVFx6B/gPxetGYfLXSJD5v0sOmh0OpZDq7yCrTSA+Rscvrtyol7JJTW40wEpQB4U1B4JXzxQitbQ3CAA=="],
@@ -55,12 +55,10 @@
"@modelcontextprotocol/sdk": ["@modelcontextprotocol/[email protected]", "", { "dependencies": { "@hono/node-server": "^1.19.9", "ajv": "^8.17.1", "ajv-formats": "^3.0.1", "content-type": "^1.0.5", "cors": "^2.8.5", "cross-spawn": "^7.0.5", "eventsource": "^3.0.2", "eventsource-parser": "^3.0.0", "express": "^5.2.1", "express-rate-limit": "^8.2.1", "hono": "^4.11.4", "jose": "^6.1.3", "json-schema-typed": "^8.0.2", "pkce-challenge": "^5.0.0", "raw-body": "^3.0.0", "zod": "^3.25 || ^4.0", "zod-to-json-schema": "^3.25.1" }, "peerDependencies": { "@cfworker/json-schema": "^4.1.1" }, "optionalPeers": ["@cfworker/json-schema"] }, "sha512-zo37mZA9hJWpULgkRpowewez1y6ML5GsXJPY8FI0tBBCd77HEvza4jDqRKOXgHNn867PVGCyTdzqpz0izu5ZjQ=="],
"@types/bun": ["@types/bun@1.2.19", "", { "dependencies": { "bun-types": "1.2.19" } }, "sha512-d9ZCmrH3CJ2uYKXQIUuZ/pUnTqIvLDS0SK7pFmbx8ma+ziH/FRMoAq5bYpRG7y+w1gl+HgyNZbtqgMq4W4e2Lg=="],
"@types/bun": ["@types/bun@1.4.3", "", { "dependencies": { "bun-types": "1.4.3" } }, "sha512-N67Pqy8JXbAY/j0tNraS3V5mubGvc6rrrNDmDCtZLlFzccbUJPDb0tvgUBC5q4m8ybkP93Jcn1QB+RYQ6PvBhw=="],
"@types/node": ["@types/[email protected]", "", { "dependencies": { "undici-types": "~6.21.0" } }, "sha512-cuVNgarYWZqxRJDQHEB58GEONhOK79QVR/qYx4S7kcUObQvUwvFnYxJuuHUKm2aieN9X3yZB4LZsuYNU1Qphsw=="],
"@types/react": ["@types/[email protected]", "", { "dependencies": { "csstype": "^3.0.2" } }, "sha512-AwAfQ2Wa5bCx9WP8nZL2uMZWod7J7/JSplxbTmBQ5ms6QpqNYm672H0Vu9ZVKVngQ+ii4R/byguVEUZQyeg44g=="],
"@types/shell-quote": ["@types/[email protected]", "", {}, "sha512-+UE8GAGRPbJVQDdxi16dgadcBfQ+KG2vgZhV1+3A1XmHbmwcdwhCUwIdy+d3pAGrbvgRoVSjeI9vOWyq376Yzw=="],
"accepts": ["[email protected]", "", { "dependencies": { "mime-types": "^3.0.0", "negotiator": "^1.0.0" } }, "sha512-5cvg6CtKwfgdmVqY1WIiXKc3Q1bkRqGLi+2W/6ao+6Y7gu/RCwRuAhGEzh5B4KlszSuTLgZYuqFqo5bImjNKng=="],
@@ -71,7 +69,7 @@
"body-parser": ["[email protected]", "", { "dependencies": { "bytes": "^3.1.2", "content-type": "^1.0.5", "debug": "^4.4.3", "http-errors": "^2.0.0", "iconv-lite": "^0.7.0", "on-finished": "^2.4.1", "qs": "^6.14.1", "raw-body": "^3.0.1", "type-is": "^2.0.1" } }, "sha512-oP5VkATKlNwcgvxi0vM0p/D3n2C3EReYVX+DNYs5TjZFn/oQt2j+4sVJtSMr18pdRr8wjTcBl6LoV+FUwzPmNA=="],
"bun-types": ["bun-types@1.2.19", "", { "dependencies": { "@types/node": "*" }, "peerDependencies": { "@types/react": "^19" } }, "sha512-uAOTaZSPuYsWIXRpj7o56Let0g/wjihKCkeRqUBhlLVM/Bt+Fj9xTo+LhC1OV1XDaGkz4hNC80et5xgy+9KTHQ=="],
"bun-types": ["bun-types@1.4.3", "", { "dependencies": { "@types/node": "*", "undici-types": "*" } }, "sha512-tinRJNGOmUN+tEAsn2y2e942n9JHRHR42iUYzDH75NmhJqVBokGSQkkfSFdNWwsANIdrK6ACH9t01M452YyTwg=="],
"bytes": ["[email protected]", "", {}, "sha512-/Nf7TyzTx6S3yRJObOAV7956r8cr2+Oj8AC5dt8wSP3BQAoeX58NoHyCU8P8zGkNXStjTSi6fzO6F0pBdcYbEg=="],
@@ -91,8 +89,6 @@
"cross-spawn": ["[email protected]", "", { "dependencies": { "path-key": "^3.1.0", "shebang-command": "^2.0.0", "which": "^2.0.1" } }, "sha512-uV2QOWP2nWzsy2aMp8aRibhi9dlzF5Hgh5SHaB9OiTGEyDTiJJyx0uy51QXdyWbtAHNua4XJzUKca3OzKUd3vA=="],
"csstype": ["[email protected]", "", {}, "sha512-M1uQkMl8rQK/szD0LNhtqxIPLpimGm8sOBwU7lLnCpSbTyY3yeU1Vc7l4KT5zT4s/yOxHH5O7tIuuLOCnLADRw=="],
"debug": ["[email protected]", "", { "dependencies": { "ms": "^2.1.3" } }, "sha512-RGwwWnwQvkVfavKVt22FGLw+xYSdzARwm0ru6DhTVA3umU5hZc28V3kO4stgYryrTlLpuvgI9GiijltAjNbcqA=="],
"depd": ["[email protected]", "", {}, "sha512-g7nH6P6dyDioJogAAGprGpCtVImJhpPk/roCzdb3fIh61/s/nPsfR6onyMwkCAR/OlC3yBC0lESvUoQEAssIrw=="],
@@ -223,7 +219,7 @@
"shebang-regex": ["[email protected]", "", {}, "sha512-7++dFhtcx3353uBaq8DDR4NuxBetBzC7ZQOhmTQInHEd6bSrXdiEyzCvG07Z44UYdLShWUyXt5M/yhz8ekcb1A=="],
"shell-quote": ["[email protected].3", "", {}, "sha512-ObmnIF4hXNg1BqhnHmgbDETF8dLPCggZWBjkQfhZpbszZnYur5DUljTcCHii5LC3J5E0yeO/1LIMyH+UvHQgyw=="],
"shell-quote": ["[email protected].4", "", {}, "sha512-VsC6n6vz1ihYYyZZwX7YZSF5l5x36ca17OC+a69h94YqB7X6XLwf+5MOgynYir2SLFUbl8gIYvBo8K8RoNQ6bQ=="],
"side-channel": ["[email protected]", "", { "dependencies": { "es-errors": "^1.3.0", "object-inspect": "^1.13.3", "side-channel-list": "^1.0.0", "side-channel-map": "^1.0.1", "side-channel-weakmap": "^1.0.2" } }, "sha512-ZX99e6tRweoUXqR+VBrslhda51Nh5MTQwou5tnUDgbtyM0dBgmhEDtWGP/xbKn6hqfPRHujUNwz5fy/wbbhnpw=="],
@@ -257,8 +253,10 @@
"wrappy": ["[email protected]", "", {}, "sha512-l4Sp/DRseor9wL6EvV2+TuQn63dMkPjZ/sp9XkghTEbV9KlPS1xUsZ3u7/IQO4wxtcFB4bgpQPRcR3QCvezPcQ=="],
"zod": ["zod@3.25.76", "", {}, "sha512-gzUt/qt81nXsFGKIFcC3YnfEAx5NkunCfnDlvuBSSFS02bcXu4Lmea0AFIUwbLWxWPx3d9p8S5QoaujKcNQxcQ=="],
"zod": ["zod@4.6.7", "", {}, "sha512-g8W2D2kU6a8untc053fX1uVBKzM7yqAxW/3zwk83+6RyF2+SKOuV+c0wBDhUwzV1PaGapmZH+rmcUzPf+3CGwg=="],
"zod-to-json-schema": ["zod-to-json-schema@3.25.2", "", { "peerDependencies": { "zod": "^3.25.28 || ^4" } }, "sha512-O/PgfnpT1xKSDeQYSCfRI5Gy3hPf91mKVDuYLUHZJMiDFptvP41MSnWofm8dnCm0256ZNfZIM7DSzuSMAFnjHA=="],
"@modelcontextprotocol/sdk/zod": ["zod@3.25.76", "", {}, "sha512-gzUt/qt81nXsFGKIFcC3YnfEAx5NkunCfnDlvuBSSFS02bcXu4Lmea0AFIUwbLWxWPx3d9p8S5QoaujKcNQxcQ=="],
"@modelcontextprotocol/sdk/zod-to-json-schema": ["[email protected]", "", { "peerDependencies": { "zod": "^3.25.28 || ^4" } }, "sha512-O/PgfnpT1xKSDeQYSCfRI5Gy3hPf91mKVDuYLUHZJMiDFptvP41MSnWofm8dnCm0256ZNfZIM7DSzuSMAFnjHA=="],
}
}
+4 -4
View File
@@ -9,7 +9,7 @@
"version": "1.0.0",
"dependencies": {
"@actions/core": "^1.10.1",
"shell-quote": "^1.8.3"
"shell-quote": "^1.8.4"
},
"devDependencies": {
"@types/bun": "^1.2.12",
@@ -139,9 +139,9 @@
}
},
"node_modules/shell-quote": {
"version": "1.8.3",
"resolved": "https://registry.npmjs.org/shell-quote/-/shell-quote-1.8.3.tgz",
"integrity": "sha512-ObmnIF4hXNg1BqhnHmgbDETF8dLPCggZWBjkQfhZpbszZnYur5DUljTcCHii5LC3J5E0yeO/1LIMyH+UvHQgyw==",
"version": "1.8.4",
"resolved": "https://registry.npmjs.org/shell-quote/-/shell-quote-1.8.4.tgz",
"integrity": "sha512-VsC6n6vz1ihYYyZZwX7YZSF5l5x36ca17OC+a69h94YqB7X6XLwf+5MOgynYir2SLFUbl8gIYvBo8K8RoNQ6bQ==",
"license": "MIT",
"engines": {
"node": ">= 0.4"
+3 -3
View File
@@ -11,11 +11,11 @@
},
"dependencies": {
"@actions/core": "^1.10.1",
"@anthropic-ai/claude-agent-sdk": "^0.3.161",
"shell-quote": "^1.8.3"
"@anthropic-ai/claude-agent-sdk": "^0.3.330",
"shell-quote": "^1.8.4"
},
"devDependencies": {
"@types/bun": "^1.2.12",
"@types/bun": "^1.4.3",
"@types/node": "^20.0.0",
"@types/shell-quote": "^1.7.5",
"prettier": "3.5.3",
+5 -1
View File
@@ -11,6 +11,9 @@ import { setupWorkloadIdentity } from "./workload-identity";
import type { WorkloadIdentityHandle } from "./workload-identity";
async function run() {
// Bun has read this by now, and it's only for the action's own script
delete process.env.BUN_RUNTIME_TRANSPILER_CACHE_PATH;
let workloadIdentity: WorkloadIdentityHandle | undefined;
try {
// When workload identity federation is configured, fetch the GitHub OIDC
@@ -75,7 +78,8 @@ async function run() {
core.setOutput("conclusion", "failure");
process.exit(1);
} finally {
// Stop refreshing the workload identity token file so the process can exit
// Stop refreshing the workload identity token file (so the process can
// exit) and delete the token material so it doesn't outlive this step
workloadIdentity?.stop();
}
}
+96 -8
View File
@@ -9,6 +9,7 @@ export type ParsedSdkOptions = {
sdkOptions: SdkOptions;
showFullOutput: boolean;
hasJsonSchema: boolean;
fallbackPermissionMode?: SdkOptions["permissionMode"];
};
// Flags that should accumulate multiple values instead of overwriting
@@ -19,10 +20,40 @@ const ACCUMULATING_FLAGS = new Set([
"disallowedTools",
"disallowed-tools",
"mcp-config",
"add-dir",
]);
// Delimiter used to join accumulated flag values
const ACCUMULATE_DELIMITER = "\x00";
export const ACCUMULATE_DELIMITER = "\x00";
// shell-quote treats ()|&;<> as control operators and splits adjacent text
// around them into separate tokens (returned as `{op}` objects, which we then
// dropped). For CLI args these must be literal characters — e.g. unquoted
// `--allowedTools Bash(gh:*)` was being mangled into bare `Bash`, silently
// widening a scoped permission rule to Bash(*). We escape each metachar to a
// Unicode private-use codepoint before parsing and restore it afterward,
// keeping shell-quote's quote/whitespace handling intact.
const SHELL_META_PAIRS: [string, string][] = [
["(", ""],
[")", ""],
["|", ""],
["&", ""],
[";", ""],
["<", ""],
[">", ""],
];
const SHELL_META_ESCAPE = new Map(SHELL_META_PAIRS);
const SHELL_META_UNESCAPE = new Map(SHELL_META_PAIRS.map(([k, v]) => [v, k]));
const SHELL_META_ESCAPE_RE = /[()|&;<>]/g;
const SHELL_META_UNESCAPE_RE = /[-]/g;
export function escapeShellMeta(s: string): string {
return s.replace(SHELL_META_ESCAPE_RE, (c) => SHELL_META_ESCAPE.get(c)!);
}
export function unescapeShellMeta(s: string): string {
return s.replace(SHELL_META_UNESCAPE_RE, (c) => SHELL_META_UNESCAPE.get(c)!);
}
type McpConfig = {
mcpServers?: Record<string, unknown>;
@@ -100,15 +131,25 @@ function stripShellComments(input: string): string {
* Accumulating flags also consume all consecutive non-flag values
* (e.g., --allowed-tools "Tool1" "Tool2" "Tool3" captures all three).
*/
function parseClaudeArgsToExtraArgs(
export function parseClaudeArgsToExtraArgs(
claudeArgs?: string,
): Record<string, string | null> {
if (!claudeArgs?.trim()) return {};
const result: Record<string, string | null> = {};
const args = parseShellArgs(stripShellComments(claudeArgs)).filter(
(arg): arg is string => typeof arg === "string",
);
const args = parseShellArgs(escapeShellMeta(stripShellComments(claudeArgs)))
.map((arg) => {
if (typeof arg === "string") return unescapeShellMeta(arg);
// With control metachars escaped above, the only non-string shell-quote
// can still emit is a glob op (bareword containing *, ?, or [). Its
// `pattern` field is the verbatim token text — use it as-is so values
// like `Bash(cmd:*)` and `Read(path/**)` round-trip intact.
if (typeof arg === "object" && arg !== null && "pattern" in arg) {
return unescapeShellMeta((arg as { pattern: string }).pattern);
}
return undefined;
})
.filter((arg): arg is string => typeof arg === "string");
for (let i = 0; i < args.length; i++) {
const arg = args[i];
@@ -146,6 +187,21 @@ function parseClaudeArgsToExtraArgs(
return result;
}
/**
* The permission mode the CLI ends up with: the last one given. `--permission-mode=x`
* is kept under its own key, with the value in the key's name.
*/
function lastPermissionMode(
extraArgs: Record<string, string | null>,
): string | null | undefined {
const modes = Object.entries(extraArgs).flatMap(([flag, value]) => {
if (flag === "permission-mode") return [value];
const [name, inlineValue] = flag.split(/=(.*)/s);
return name === "permission-mode" ? [inlineValue ?? null] : [];
});
return modes.at(-1);
}
/**
* Parse ClaudeOptions into SDK-compatible options
* Uses extraArgs for CLI pass-through instead of duplicating option parsing
@@ -161,6 +217,20 @@ export function parseSdkOptions(options: ClaudeOptions): ParsedSdkOptions {
// Detect if --json-schema is present (for hasJsonSchema flag)
const hasJsonSchema = "json-schema" in extraArgs;
const modelFromClaudeArgs = extraArgs["model"] || undefined;
delete extraArgs["model"];
const maxTurnsFromClaudeArgs = extraArgs["max-turns"] || undefined;
delete extraArgs["max-turns"];
const additionalDirectories = extraArgs["add-dir"]
? extraArgs["add-dir"]
.split(ACCUMULATE_DELIMITER)
.map((dir) => dir.trim())
.filter(Boolean)
: [];
delete extraArgs["add-dir"];
// Extract and merge allowedTools from all sources:
// 1. From extraArgs (parsed from claudeArgs - contains tag mode's tools)
// - Check both camelCase (--allowedTools) and hyphenated (--allowed-tools) variants
@@ -205,8 +275,11 @@ export function parseSdkOptions(options: ClaudeOptions): ParsedSdkOptions {
const directDisallowedTools = options.disallowedTools
? options.disallowedTools.split(",").map((t) => t.trim())
: [];
// Not deduplicated: a rule can hold a comma, so these can be pieces of rules, and the
// SDK joins them again. Dropping a repeated piece would break a rule and the ones after it.
const mergedDisallowedTools = [
...new Set([...extraArgsDisallowedTools, ...directDisallowedTools]),
...extraArgsDisallowedTools,
...directDisallowedTools,
];
delete extraArgs["disallowedTools"];
delete extraArgs["disallowed-tools"];
@@ -235,6 +308,10 @@ export function parseSdkOptions(options: ClaudeOptions): ParsedSdkOptions {
delete env.ACTIONS_ID_TOKEN_REQUEST_URL;
delete env.ACTIONS_ID_TOKEN_REQUEST_TOKEN;
// Remove ALL_INPUTS as it is only needed during initial setup to determine
// input presence (collectActionInputsPresence) and contains serialized workflow inputs.
delete env.ALL_INPUTS;
// Build system prompt option - default to claude_code preset
let systemPrompt: SdkOptions["systemPrompt"];
if (options.systemPrompt) {
@@ -256,8 +333,12 @@ export function parseSdkOptions(options: ClaudeOptions): ParsedSdkOptions {
// Build SDK options - use merged tools from both direct options and claudeArgs
const sdkOptions: SdkOptions = {
// Direct options from ClaudeOptions inputs
model: options.model,
maxTurns: options.maxTurns ? parseInt(options.maxTurns, 10) : undefined,
model: options.model || modelFromClaudeArgs,
maxTurns: options.maxTurns
? parseInt(options.maxTurns, 10)
: maxTurnsFromClaudeArgs
? parseInt(maxTurnsFromClaudeArgs, 10)
: undefined,
allowedTools:
mergedAllowedTools.length > 0 ? mergedAllowedTools : undefined,
disallowedTools:
@@ -265,6 +346,8 @@ export function parseSdkOptions(options: ClaudeOptions): ParsedSdkOptions {
systemPrompt,
fallbackModel: options.fallbackModel,
pathToClaudeCodeExecutable: options.pathToClaudeCodeExecutable,
additionalDirectories:
additionalDirectories.length > 0 ? additionalDirectories : undefined,
// Pass through claudeArgs as extraArgs - CLI handles --mcp-config, --json-schema, etc.
// Note: allowedTools and disallowedTools have been removed from extraArgs to prevent duplicates
@@ -287,5 +370,10 @@ export function parseSdkOptions(options: ClaudeOptions): ParsedSdkOptions {
sdkOptions,
showFullOutput,
hasJsonSchema,
// Only when auto mode is what the session ends up asking for: a mode someone
// chose in claude_args must stay as they set it.
...(lastPermissionMode(extraArgs) === "auto" && {
fallbackPermissionMode: options.fallbackPermissionMode,
}),
};
}
+119 -9
View File
@@ -36,11 +36,14 @@ async function fileExists(path: string): Promise<boolean> {
* Creates a prompt configuration for the SDK.
* If a user request file exists alongside the prompt file, returns a multi-block
* SDKUserMessage that enables slash command processing in the CLI.
* Otherwise, returns the prompt as a simple string.
* Otherwise, returns the prompt as a simple string, or as a single-block message
* when the session has to take control requests. Those need its input open, and
* the SDK closes it once the prompt ends, so the prompt then waits for holdOpenUntil.
*/
async function createPromptConfig(
promptPath: string,
showFullOutput: boolean,
holdOpenUntil?: Promise<void>,
): Promise<string | AsyncIterable<SDKUserMessage>> {
const promptContent = await readFile(promptPath, "utf-8");
@@ -50,7 +53,19 @@ async function createPromptConfig(
if (!hasUserRequest) {
// No user request file - use simple string prompt
return promptContent;
if (!holdOpenUntil) return promptContent;
return (async function* (): AsyncGenerator<SDKUserMessage> {
yield {
type: "user",
session_id: "",
message: {
role: "user",
content: [{ type: "text", text: promptContent }],
},
parent_tool_use_id: null,
};
await holdOpenUntil;
})();
}
// User request file exists - create multi-block message
@@ -77,11 +92,41 @@ async function createPromptConfig(
},
parent_tool_use_id: null,
};
await holdOpenUntil;
}
return createMultiBlockMessage();
}
type ModelUsageSummary = Record<
string,
{
contextWindow: number;
maxOutputTokens: number;
}
>;
/**
* Keep resolved model limits visible without exposing token usage or cost details.
*/
function sanitizeModelUsage(
modelUsage: SDKResultMessage["modelUsage"] | undefined,
): ModelUsageSummary | undefined {
if (!modelUsage) {
return undefined;
}
return Object.fromEntries(
Object.entries(modelUsage).map(([model, usage]) => [
model,
{
contextWindow: usage.contextWindow,
maxOutputTokens: usage.maxOutputTokens,
},
]),
);
}
/**
* Sanitizes SDK output to match CLI sanitization behavior
*/
@@ -119,6 +164,7 @@ function sanitizeSdkOutput(
num_turns: resultMsg.num_turns,
total_cost_usd: resultMsg.total_cost_usd,
permission_denials_count: resultMsg.permission_denials?.length ?? 0,
modelUsage: sanitizeModelUsage(resultMsg.modelUsage),
},
null,
2,
@@ -134,10 +180,23 @@ function sanitizeSdkOutput(
*/
export async function runClaudeWithSdk(
promptPath: string,
{ sdkOptions, showFullOutput, hasJsonSchema }: ParsedSdkOptions,
{
sdkOptions,
showFullOutput,
hasJsonSchema,
fallbackPermissionMode,
}: ParsedSdkOptions,
): Promise<ClaudeRunResult> {
// Create prompt configuration - may be a string or multi-block message
const prompt = await createPromptConfig(promptPath, showFullOutput);
let releasePrompt = () => {};
const promptHeldOpen = new Promise<void>((resolve) => {
releasePrompt = resolve;
});
const prompt = await createPromptConfig(
promptPath,
showFullOutput,
fallbackPermissionMode ? promptHeldOpen : undefined,
);
if (!showFullOutput) {
console.log(
@@ -157,9 +216,27 @@ export async function runClaudeWithSdk(
let resultMessage: SDKResultMessage | undefined;
try {
for await (const message of query({ prompt, options: sdkOptions })) {
const session = query({ prompt, options: sdkOptions });
for await (const message of session) {
messages.push(message);
if (message.type === "system" && message.subtype === "init") {
// A session that can't use auto mode starts in Manual without an error.
if (fallbackPermissionMode && message.permissionMode === "default") {
core.warning(
`Auto mode is not available for this run, for example because of its model. Using ${fallbackPermissionMode} instead.`,
);
try {
await session.setPermissionMode(fallbackPermissionMode);
} catch {
core.warning(
`Could not switch to ${fallbackPermissionMode}, so tool calls that need permission will be denied.`,
);
}
}
releasePrompt();
}
const sanitized = sanitizeSdkOutput(message, showFullOutput);
if (sanitized) {
console.log(sanitized);
@@ -167,12 +244,24 @@ export async function runClaudeWithSdk(
if (message.type === "result") {
resultMessage = message as SDKResultMessage;
// The SDK's query() iterator should close itself after the
// result message, but in some workflow contexts (notably
// pull_request-triggered runs) it stays open indefinitely and
// the for-await hangs until the workflow's timeout-minutes
// kills the job. This causes the action to "succeed" inside
// Claude (verdict posted, $cost recorded) but be reported as
// cancelled with no execution-output.json written. Break
// explicitly: by SDK contract no further messages follow a
// result, so the break is safe.
break;
}
}
} catch (error) {
console.error("SDK execution error:", error);
await writeExecutionFile(messages);
throw new Error(`SDK execution error: ${error}`);
} finally {
releasePrompt();
}
const result: ClaudeRunResult = {
@@ -198,7 +287,21 @@ export async function runClaudeWithSdk(
throw new Error("No result message received from Claude");
}
const isSuccess = resultMessage.subtype === "success";
if (
resultMessage.subtype === "success" &&
!resultMessage.is_error &&
sdkOptions.maxTurns !== undefined &&
resultMessage.num_turns > sdkOptions.maxTurns
) {
const message = `Claude reported a successful result after ${resultMessage.num_turns} turns, exceeding the configured maximum of ${sdkOptions.maxTurns}`;
core.error(message);
throw new Error(message);
}
// subtype "success" with is_error:true means the run errored without producing
// a real result — treat it as failure so CI does not show a misleading green check.
const isSuccess =
resultMessage.subtype === "success" && !resultMessage.is_error;
result.conclusion = isSuccess ? "success" : "failure";
// Handle structured output
@@ -224,14 +327,21 @@ export async function runClaudeWithSdk(
}
if (!isSuccess) {
if (resultMessage.subtype === "success" && resultMessage.is_error) {
core.error(
"Claude result reported subtype success with is_error:true (run did not complete successfully)",
);
}
if ("errors" in resultMessage && resultMessage.errors) {
core.error(`Execution failed: ${resultMessage.errors.join(", ")}`);
}
throw new Error(
`Claude execution failed: ${
"errors" in resultMessage && resultMessage.errors
? resultMessage.errors.join(", ")
: "unknown error"
resultMessage.subtype === "success" && resultMessage.is_error
? "result is_error:true"
: "errors" in resultMessage && resultMessage.errors
? resultMessage.errors.join(", ")
: "unknown error"
}`,
);
}
+4
View File
@@ -2,6 +2,8 @@ import { runClaudeWithSdk } from "./run-claude-sdk";
import type { ClaudeRunResult } from "./run-claude-sdk";
import { parseSdkOptions } from "./parse-sdk-options";
import type { PermissionMode } from "@anthropic-ai/claude-agent-sdk";
export type ClaudeOptions = {
claudeArgs?: string;
model?: string;
@@ -14,6 +16,8 @@ export type ClaudeOptions = {
appendSystemPrompt?: string;
fallbackModel?: string;
showFullOutput?: string;
/** Mode to switch to when auto mode was asked for but isn't available. */
fallbackPermissionMode?: PermissionMode;
};
export async function runClaude(
+79 -3
View File
@@ -15,7 +15,8 @@
*/
import * as core from "@actions/core";
import { mkdirSync, writeFileSync } from "fs";
import { createHash } from "crypto";
import { mkdirSync, rmSync, writeFileSync } from "fs";
import { join } from "path";
import { retryWithBackoff } from "./retry";
@@ -50,6 +51,63 @@ async function fetchIdentityToken(audience: string) {
return retryWithBackoff(() => core.getIDToken(audience));
}
/**
* Writes a profile config that switches federation resolution to the
* file-backed path. Resolving federation through a profile (rather than bare
* env vars) enables the SDK's on-disk credentials cache, so the several
* `claude` processes the action spawns (plugin installs, main query) share
* one exchanged access token instead of each re-exchanging the single-use
* GitHub OIDC token, which fails with 401 (`jti_reused`).
*
* The profile is intentionally minimal: the SDK gap-fills the federation
* fields (rule, organization, identity-token file, service account, base URL)
* from the ANTHROPIC_* env vars the action already exports, so the file only
* needs to exist to turn the cache on.
*
* The config dir name embeds a fingerprint of the federation inputs. The
* SDK's cache reuses a token on `expires_at` alone, with no record of the
* config that minted it, and the token's scope is bound at mint time — so a
* later action step in the same job (RUNNER_TEMP is per-job) with different
* federation inputs must land in a different dir or it would silently reuse
* the first step's token.
*
* Sharing the cache is only safe while the action spawns its `claude`
* subprocesses sequentially: the SDK cache is not cross-process serialized,
* and concurrent cache misses would each re-exchange the same single-use
* identity token. Parallelizing the plugin installs would reintroduce the
* `jti_reused` failures.
*/
function writeFederationProfile(baseDir: string): string {
// Every input that changes which credential the exchange mints must be in
// here; service_account_id and scope are sent in the exchange request body.
const fingerprint = createHash("sha256")
.update(
JSON.stringify([
process.env.ANTHROPIC_FEDERATION_RULE_ID?.trim() ?? "",
process.env.ANTHROPIC_ORGANIZATION_ID?.trim() ?? "",
process.env.ANTHROPIC_SERVICE_ACCOUNT_ID?.trim() ?? "",
process.env.ANTHROPIC_WORKSPACE_ID?.trim() ?? "",
process.env.ANTHROPIC_BASE_URL?.trim() ?? "",
process.env.ANTHROPIC_SCOPE?.trim() ?? "",
]),
)
.digest("hex")
.slice(0, 16);
const configDir = join(baseDir, `config-${fingerprint}`);
mkdirSync(join(configDir, "configs"), { recursive: true, mode: 0o700 });
writeFileSync(
join(configDir, "configs", "default.json"),
JSON.stringify(
{ version: "1.0", authentication: { type: "oidc_federation" } },
null,
2,
),
{ mode: 0o600 },
);
return configDir;
}
/**
* Fetches a GitHub Actions OIDC token, writes it to a file in RUNNER_TEMP,
* exports ANTHROPIC_IDENTITY_TOKEN_FILE, and starts a background refresh so
@@ -57,7 +115,8 @@ async function fetchIdentityToken(audience: string) {
*
* Returns undefined when federation is not configured or is shadowed by a
* higher-precedence credential. Callers must invoke stop() when execution
* finishes.
* finishes; it also deletes the identity token and any cached exchanged
* credential.
*/
export async function setupWorkloadIdentity(): Promise<
WorkloadIdentityHandle | undefined
@@ -101,6 +160,17 @@ export async function setupWorkloadIdentity(): Promise<
}
process.env.ANTHROPIC_IDENTITY_TOKEN_FILE = tokenFile;
if (
process.env.ANTHROPIC_CONFIG_DIR?.trim() ||
process.env.ANTHROPIC_PROFILE?.trim()
) {
core.warning(
"ANTHROPIC_CONFIG_DIR or ANTHROPIC_PROFILE is already set, so the action will not write its own federation profile. Credential caching across the spawned Claude processes follows the existing profile configuration.",
);
} else {
process.env.ANTHROPIC_CONFIG_DIR = writeFederationProfile(tokenDir);
process.env.ANTHROPIC_PROFILE = "default";
}
console.log(
`Workload identity federation configured (rule: ${process.env.ANTHROPIC_FEDERATION_RULE_ID}, identity token file: ${tokenFile})`,
);
@@ -115,6 +185,12 @@ export async function setupWorkloadIdentity(): Promise<
return {
tokenFile,
stop: () => clearInterval(refreshInterval),
stop: () => {
clearInterval(refreshInterval);
// RUNNER_TEMP is per-job, not per-step: remove the identity token, the
// profile, and the cached exchanged credential so they don't outlive
// this step.
rmSync(tokenDir, { recursive: true, force: true });
},
};
}
+278 -5
View File
@@ -106,7 +106,8 @@ describe("parseSdkOptions", () => {
const result = parseSdkOptions(options);
expect(result.sdkOptions.extraArgs?.["allowedTools"]).toBeUndefined();
expect(result.sdkOptions.extraArgs?.["model"]).toBe("claude-3-5-sonnet");
expect(result.sdkOptions.extraArgs?.["model"]).toBeUndefined();
expect(result.sdkOptions.model).toBe("claude-3-5-sonnet");
});
test("should handle hyphenated --allowed-tools flag", () => {
@@ -137,6 +138,110 @@ describe("parseSdkOptions", () => {
]);
});
test("should preserve unquoted Bash(cmd:*) rules instead of collapsing to bare Bash", () => {
// Regression: shell-quote tokenizes unquoted `(`/`)` as control ops and
// `*` as a glob, which were filtered out — collapsing scoped rules like
// `Bash(gh:*)` into bare `Bash` (= Bash(*), unrestricted shell).
const options: ClaudeOptions = {
claudeArgs: "--allowedTools View,Bash(gh:*),Bash(cat:*)",
};
const result = parseSdkOptions(options);
expect(result.sdkOptions.allowedTools).toEqual([
"View",
"Bash(gh:*)",
"Bash(cat:*)",
]);
expect(result.sdkOptions.allowedTools).not.toContain("Bash");
});
test("should preserve unquoted space-separated Bash(cmd:*) rules", () => {
const options: ClaudeOptions = {
claudeArgs: "--allowed-tools Bash(gh:*) Bash(cat:*) Read(//tmp/**)",
};
const result = parseSdkOptions(options);
expect(result.sdkOptions.allowedTools).toEqual([
"Bash(gh:*)",
"Bash(cat:*)",
"Read(//tmp/**)",
]);
expect(result.sdkOptions.allowedTools).not.toContain("Bash");
});
test("should preserve unquoted Tool(content) rules without glob chars", () => {
const options: ClaudeOptions = {
claudeArgs:
"--allowedTools Read(~/file),WebFetch(domain:example.com),Edit",
};
const result = parseSdkOptions(options);
expect(result.sdkOptions.allowedTools).toEqual([
"Read(~/file)",
"WebFetch(domain:example.com)",
"Edit",
]);
});
test("should still preserve quoted Bash(cmd:*) rules (no regression)", () => {
const options: ClaudeOptions = {
claudeArgs: '--allowedTools "Bash(gh:*),Bash(cat:*)"',
};
const result = parseSdkOptions(options);
expect(result.sdkOptions.allowedTools).toEqual([
"Bash(gh:*)",
"Bash(cat:*)",
]);
});
test("should merge quoted tag-mode tools with unquoted user tools without widening", () => {
// Real-world shape: the action's tag mode wraps its own --allowedTools in
// double quotes, then appends the user's claude_args (typically unquoted
// in workflow YAML). Both halves must round-trip.
const options: ClaudeOptions = {
claudeArgs:
'--permission-mode acceptEdits --allowedTools "Glob,Grep,Read,Bash(git add:*),Bash(git commit:*)" ' +
"--model claude-opus-4-7\n" +
"--allowedTools View,Bash(gh:*),Bash(printf:*),Bash(cat:*)",
};
const result = parseSdkOptions(options);
expect(result.sdkOptions.allowedTools).toEqual([
"Glob",
"Grep",
"Read",
"Bash(git add:*)",
"Bash(git commit:*)",
"View",
"Bash(gh:*)",
"Bash(printf:*)",
"Bash(cat:*)",
]);
expect(result.sdkOptions.allowedTools).not.toContain("Bash");
});
test("should preserve unquoted disallowedTools rules without widening", () => {
// Same bug class on the deny side: a scoped deny collapsing to bare
// `Bash` would block all shell instead of the intended prefix.
const options: ClaudeOptions = {
claudeArgs: "--disallowedTools Bash(rm:*),Bash(sudo:*)",
};
const result = parseSdkOptions(options);
expect(result.sdkOptions.disallowedTools).toEqual([
"Bash(rm:*)",
"Bash(sudo:*)",
]);
expect(result.sdkOptions.disallowedTools).not.toContain("Bash");
});
test("should handle mixed camelCase and hyphenated allowedTools flags", () => {
const options: ClaudeOptions = {
claudeArgs: '--allowedTools "Edit,Read" --allowed-tools "Write,Glob"',
@@ -164,6 +269,17 @@ describe("parseSdkOptions", () => {
expect(result.sdkOptions.extraArgs?.["disallowedTools"]).toBeUndefined();
});
test("keeps rules that share text after a comma", () => {
const result = parseSdkOptions({
claudeArgs: `--disallowedTools 'Bash(/x.sh {a,b}/*)' --disallowedTools 'Bash(git add {a,b}/*)'`,
disallowedTools: "Bash(bun:*)",
});
expect(result.sdkOptions.disallowedTools?.join(",")).toBe(
"Bash(/x.sh {a,b}/*),Bash(git add {a,b}/*),Bash(bun:*)",
);
});
test("should merge disallowedTools from both sources", () => {
const options: ClaudeOptions = {
claudeArgs: '--disallowedTools "Bash"',
@@ -262,7 +378,8 @@ describe("parseSdkOptions", () => {
);
expect(mcpConfig.mcpServers).toHaveProperty("server1");
expect(mcpConfig.mcpServers).toHaveProperty("server2");
expect(result.sdkOptions.extraArgs?.["model"]).toBe("claude-3-5-sonnet");
expect(result.sdkOptions.extraArgs?.["model"]).toBeUndefined();
expect(result.sdkOptions.model).toBe("claude-3-5-sonnet");
});
test("should handle real-world scenario: action config + user config", () => {
@@ -298,6 +415,46 @@ describe("parseSdkOptions", () => {
});
});
describe("add-dir handling", () => {
test("should accumulate multiple add-dir flags into additionalDirectories", () => {
const options: ClaudeOptions = {
claudeArgs: '--add-dir "/path/to/dir-a"\n--add-dir "/path/to/dir-b"',
};
const result = parseSdkOptions(options);
expect(result.sdkOptions.additionalDirectories).toEqual([
"/path/to/dir-a",
"/path/to/dir-b",
]);
expect(result.sdkOptions.extraArgs?.["add-dir"]).toBeUndefined();
});
test("should map a single add-dir flag to additionalDirectories", () => {
const options: ClaudeOptions = {
claudeArgs: '--add-dir "/path/to/dir"',
};
const result = parseSdkOptions(options);
expect(result.sdkOptions.additionalDirectories).toEqual(["/path/to/dir"]);
expect(result.sdkOptions.extraArgs?.["add-dir"]).toBeUndefined();
});
test("should preserve other extraArgs when extracting add-dir", () => {
const options: ClaudeOptions = {
claudeArgs: '--model "claude-3-5-sonnet" --add-dir "/path/to/dir"',
};
const result = parseSdkOptions(options);
expect(result.sdkOptions.additionalDirectories).toEqual(["/path/to/dir"]);
expect(result.sdkOptions.extraArgs?.["model"]).toBeUndefined();
expect(result.sdkOptions.model).toBe("claude-3-5-sonnet");
expect(result.sdkOptions.extraArgs?.["add-dir"]).toBeUndefined();
});
});
describe("other extraArgs passthrough", () => {
test("should pass through json-schema in extraArgs", () => {
const options: ClaudeOptions = {
@@ -321,7 +478,8 @@ describe("parseSdkOptions", () => {
const result = parseSdkOptions(options);
expect(result.sdkOptions.extraArgs?.["model"]).toBe("claude-haiku");
expect(result.sdkOptions.extraArgs?.["model"]).toBeUndefined();
expect(result.sdkOptions.model).toBe("claude-haiku");
expect(result.sdkOptions.allowedTools).toEqual(["Edit"]);
});
@@ -332,7 +490,8 @@ describe("parseSdkOptions", () => {
const result = parseSdkOptions(options);
expect(result.sdkOptions.extraArgs?.["model"]).toBe("claude-haiku");
expect(result.sdkOptions.extraArgs?.["model"]).toBeUndefined();
expect(result.sdkOptions.model).toBe("claude-haiku");
});
test("should not strip inline # that appears inside a quoted value", () => {
@@ -342,11 +501,62 @@ describe("parseSdkOptions", () => {
const result = parseSdkOptions(options);
expect(result.sdkOptions.extraArgs?.["model"]).toBe("claude-haiku");
expect(result.sdkOptions.extraArgs?.["model"]).toBeUndefined();
expect(result.sdkOptions.model).toBe("claude-haiku");
expect(result.sdkOptions.extraArgs?.["prompt"]).toBe("use color #ff0000");
});
});
describe("model handling", () => {
test("should map --model from claudeArgs to sdkOptions.model", () => {
const options: ClaudeOptions = {
claudeArgs: "--model claude-haiku-4-5-20251001",
};
const result = parseSdkOptions(options);
expect(result.sdkOptions.model).toBe("claude-haiku-4-5-20251001");
expect(result.sdkOptions.extraArgs?.["model"]).toBeUndefined();
});
test("should prefer direct model option over --model from claudeArgs", () => {
const options: ClaudeOptions = {
model: "claude-sonnet-4-6",
claudeArgs: "--model claude-haiku-4-5-20251001",
};
const result = parseSdkOptions(options);
expect(result.sdkOptions.model).toBe("claude-sonnet-4-6");
expect(result.sdkOptions.extraArgs?.["model"]).toBeUndefined();
});
});
describe("max turns handling", () => {
test("should map --max-turns from claudeArgs to sdkOptions.maxTurns", () => {
const options: ClaudeOptions = {
claudeArgs: "--max-turns 60",
};
const result = parseSdkOptions(options);
expect(result.sdkOptions.maxTurns).toBe(60);
expect(result.sdkOptions.extraArgs?.["max-turns"]).toBeUndefined();
});
test("should prefer the direct maxTurns option", () => {
const options: ClaudeOptions = {
maxTurns: "25",
claudeArgs: "--max-turns 60",
};
const result = parseSdkOptions(options);
expect(result.sdkOptions.maxTurns).toBe(25);
expect(result.sdkOptions.extraArgs?.["max-turns"]).toBeUndefined();
});
});
describe("environment variables passthrough", () => {
test("should include OTEL environment variables in sdkOptions.env", () => {
// Set up test environment variables
@@ -421,5 +631,68 @@ describe("parseSdkOptions", () => {
process.env = originalEnv;
}
});
test("should strip ALL_INPUTS from env", () => {
const originalEnv = { ...process.env };
process.env.ALL_INPUTS = JSON.stringify({
anthropic_api_key: "sk-ant-test-key",
github_token: "ghp_test_token",
});
try {
const options: ClaudeOptions = {};
const result = parseSdkOptions(options);
expect(result.sdkOptions.env?.ALL_INPUTS).toBeUndefined();
} finally {
process.env = originalEnv;
}
});
});
});
describe("fallbackPermissionMode", () => {
const fallback = (claudeArgs: string) =>
parseSdkOptions({ claudeArgs, fallbackPermissionMode: "acceptEdits" })
.fallbackPermissionMode;
test("applies when auto mode is the mode asked for", () => {
expect(fallback("--permission-mode auto")).toBe("acceptEdits");
});
test("applies when a later flag asks for auto mode", () => {
expect(fallback("--permission-mode default --permission-mode auto")).toBe(
"acceptEdits",
);
});
test("does not override a mode set after auto", () => {
expect(
fallback("--permission-mode auto --permission-mode default"),
).toBeUndefined();
});
test("does not override a mode set with an equals sign", () => {
expect(
fallback("--permission-mode auto --permission-mode=default"),
).toBeUndefined();
});
test("applies when auto mode is set with an equals sign", () => {
expect(fallback("--permission-mode=auto")).toBe("acceptEdits");
expect(fallback("--permission-mode default --permission-mode=auto")).toBe(
"acceptEdits",
);
});
test("does not apply with no mode", () => {
expect(fallback("--model x")).toBeUndefined();
});
test("is unset unless the caller asks for it", () => {
expect(
parseSdkOptions({ claudeArgs: "--permission-mode auto" })
.fallbackPermissionMode,
).toBeUndefined();
});
});
+50
View File
@@ -0,0 +1,50 @@
import { readFileSync } from "node:fs";
import { describe, expect, test } from "bun:test";
const actionMetadata = readFileSync(
new URL("../action.yml", import.meta.url),
"utf8",
);
const readme = readFileSync(new URL("../README.md", import.meta.url), "utf8");
describe("base action README", () => {
test("should document every input declared in the action metadata", () => {
const inputMetadata = actionMetadata.match(
/^inputs:\n([\s\S]*?)^outputs:/m,
)?.[1];
const inputReference = readme.match(
/^## Inputs\n([\s\S]*?)^## Outputs/m,
)?.[1];
expect(inputMetadata).toBeDefined();
expect(inputReference).toBeDefined();
const declaredInputs = [
...(inputMetadata?.matchAll(/^ ([a-z0-9_]+):$/gm) ?? []),
].map((match) => match[1]);
const documentedInputs = [
...(inputReference?.matchAll(/^\| `([^`]+)`/gm) ?? []),
].map((match) => match[1]);
expect(documentedInputs).toEqual(declaredInputs);
});
test("should not use removed legacy inputs in workflow examples", () => {
const removedInputs = [
"allowed_tools",
"disallowed_tools",
"max_turns",
"mcp_config",
"system_prompt",
"append_system_prompt",
"claude_env",
"model",
"anthropic_model",
"fallback_model",
];
for (const input of removedInputs) {
expect(readme).not.toMatch(new RegExp(`^\\s+${input}:`, "m"));
}
});
});
+221
View File
@@ -63,4 +63,225 @@ describe("runClaudeWithSdk", () => {
consoleLogSpy.mockRestore();
}
});
test("logs resolved model limits without exposing token usage", async () => {
const consoleLogSpy = spyOn(console, "log").mockImplementation(() => {});
tempDir = await mkdtemp(join(tmpdir(), "claude-sdk-"));
process.env.RUNNER_TEMP = tempDir;
const promptPath = join(tempDir, "prompt.txt");
await writeFile(promptPath, "test prompt");
const initMessage = {
type: "system",
subtype: "init",
session_id: "session-123",
model: "claude-opus-5",
};
const resultMessage = {
type: "result",
subtype: "success",
is_error: false,
duration_ms: 434,
num_turns: 1,
total_cost_usd: 1.23,
permission_denials: [],
modelUsage: {
"claude-opus-5": {
inputTokens: 96209,
outputTokens: 55324,
cacheReadInputTokens: 1135701,
cacheCreationInputTokens: 149043,
webSearchRequests: 0,
costUSD: 1.23,
contextWindow: 200000,
maxOutputTokens: 64000,
},
},
};
mock.module("@anthropic-ai/claude-agent-sdk", () => ({
query: async function* () {
yield initMessage;
yield resultMessage;
},
}));
try {
const { runClaudeWithSdk } = await import("../src/run-claude-sdk");
await expect(
runClaudeWithSdk(promptPath, {
sdkOptions: {},
showFullOutput: false,
hasJsonSchema: false,
}),
).resolves.toMatchObject({ conclusion: "success" });
const sanitizedResult = consoleLogSpy.mock.calls
.map(([message]) => message)
.find(
(message) =>
typeof message === "string" && message.includes('"type": "result"'),
);
expect(sanitizedResult).toBeDefined();
if (typeof sanitizedResult !== "string") {
throw new Error("Sanitized result output was not logged");
}
expect(JSON.parse(sanitizedResult)).toEqual({
type: "result",
subtype: "success",
is_error: false,
duration_ms: 434,
num_turns: 1,
total_cost_usd: 1.23,
permission_denials_count: 0,
modelUsage: {
"claude-opus-5": {
contextWindow: 200000,
maxOutputTokens: 64000,
},
},
});
expect(sanitizedResult).not.toContain("inputTokens");
expect(sanitizedResult).not.toContain("costUSD");
} finally {
consoleLogSpy.mockRestore();
}
});
test("fails when result subtype is success but is_error is true", async () => {
const consoleErrorSpy = spyOn(console, "error").mockImplementation(
() => {},
);
const consoleLogSpy = spyOn(console, "log").mockImplementation(() => {});
const coreErrorSpy = spyOn(
await import("@actions/core"),
"error",
).mockImplementation(() => {});
tempDir = await mkdtemp(join(tmpdir(), "claude-sdk-"));
process.env.RUNNER_TEMP = tempDir;
const promptPath = join(tempDir, "prompt.txt");
await writeFile(promptPath, "test prompt");
const initMessage = {
type: "system",
subtype: "init",
session_id: "session-123",
model: "claude-sonnet-5",
};
const errorResultMessage = {
type: "result",
subtype: "success",
is_error: true,
duration_ms: 434,
num_turns: 1,
total_cost_usd: 0,
permission_denials: [],
};
mock.module("@anthropic-ai/claude-agent-sdk", () => ({
query: async function* () {
yield initMessage;
yield errorResultMessage;
},
}));
try {
const { runClaudeWithSdk } = await import("../src/run-claude-sdk");
await expect(
runClaudeWithSdk(promptPath, {
sdkOptions: {},
showFullOutput: false,
hasJsonSchema: false,
}),
).rejects.toThrow("result is_error:true");
const executionFile = join(tempDir, "claude-execution-output.json");
await expect(readFile(executionFile, "utf-8")).resolves.toBe(
JSON.stringify([initMessage, errorResultMessage], null, 2),
);
expect(coreErrorSpy).toHaveBeenCalledWith(
"Claude result reported subtype success with is_error:true (run did not complete successfully)",
);
} finally {
consoleErrorSpy.mockRestore();
consoleLogSpy.mockRestore();
coreErrorSpy.mockRestore();
}
});
test("fails closed when a successful result exceeds maxTurns", async () => {
const consoleErrorSpy = spyOn(console, "error").mockImplementation(
() => {},
);
const consoleLogSpy = spyOn(console, "log").mockImplementation(() => {});
const coreErrorSpy = spyOn(
await import("@actions/core"),
"error",
).mockImplementation(() => {});
tempDir = await mkdtemp(join(tmpdir(), "claude-sdk-"));
process.env.RUNNER_TEMP = tempDir;
const promptPath = join(tempDir, "prompt.txt");
await writeFile(promptPath, "test prompt");
const initMessage = {
type: "system",
subtype: "init",
session_id: "session-123",
model: "claude-opus-4-7",
};
const successResultMessage = {
type: "result",
subtype: "success",
is_error: false,
duration_ms: 960000,
num_turns: 73,
total_cost_usd: 0,
permission_denials: [],
};
mock.module("@anthropic-ai/claude-agent-sdk", () => ({
query: async function* () {
yield initMessage;
yield successResultMessage;
},
}));
try {
const { runClaudeWithSdk } = await import("../src/run-claude-sdk");
await expect(
runClaudeWithSdk(promptPath, {
sdkOptions: { maxTurns: 60 },
showFullOutput: false,
hasJsonSchema: false,
}),
).rejects.toThrow(
"Claude reported a successful result after 73 turns, exceeding the configured maximum of 60",
);
const executionFile = join(tempDir, "claude-execution-output.json");
await expect(readFile(executionFile, "utf-8")).resolves.toBe(
JSON.stringify([initMessage, successResultMessage], null, 2),
);
expect(coreErrorSpy).toHaveBeenCalledWith(
"Claude reported a successful result after 73 turns, exceeding the configured maximum of 60",
);
} finally {
consoleErrorSpy.mockRestore();
consoleLogSpy.mockRestore();
coreErrorSpy.mockRestore();
}
});
});
+132 -1
View File
@@ -2,7 +2,14 @@
import { describe, test, expect, beforeEach, afterEach, spyOn } from "bun:test";
import * as core from "@actions/core";
import { existsSync, mkdtempSync, readFileSync, rmSync, statSync } from "fs";
import {
existsSync,
mkdtempSync,
readdirSync,
readFileSync,
rmSync,
statSync,
} from "fs";
import { tmpdir } from "os";
import { join } from "path";
import {
@@ -27,6 +34,12 @@ describe("workload identity federation", () => {
delete process.env.ANTHROPIC_ORGANIZATION_ID;
delete process.env.ANTHROPIC_OIDC_AUDIENCE;
delete process.env.ANTHROPIC_IDENTITY_TOKEN_FILE;
delete process.env.ANTHROPIC_SERVICE_ACCOUNT_ID;
delete process.env.ANTHROPIC_WORKSPACE_ID;
delete process.env.ANTHROPIC_BASE_URL;
delete process.env.ANTHROPIC_SCOPE;
delete process.env.ANTHROPIC_CONFIG_DIR;
delete process.env.ANTHROPIC_PROFILE;
getIDTokenSpy = spyOn(core, "getIDToken").mockResolvedValue(
"test-identity-token",
@@ -123,5 +136,123 @@ describe("workload identity federation", () => {
handle?.stop();
}
});
test("writes a minimal federation profile and selects it", async () => {
process.env.ANTHROPIC_FEDERATION_RULE_ID = "fdrl_test";
process.env.ANTHROPIC_ORGANIZATION_ID =
"00000000-0000-0000-0000-000000000000";
process.env.ANTHROPIC_SERVICE_ACCOUNT_ID = "svac_test";
process.env.ANTHROPIC_WORKSPACE_ID = "wrkspc_test";
const handle = await setupWorkloadIdentity();
try {
const configDir = process.env.ANTHROPIC_CONFIG_DIR;
expect(configDir).toBeDefined();
expect(
configDir!.startsWith(
join(tempDir, "claude-workload-identity", "config-"),
),
).toBe(true);
expect(process.env.ANTHROPIC_PROFILE).toBe("default");
const profilePath = join(configDir!, "configs", "default.json");
expect(statSync(profilePath).mode & 0o777).toBe(0o600);
// Minimal on purpose: the SDK gap-fills the federation fields from
// the ANTHROPIC_* env vars the action exports.
expect(JSON.parse(readFileSync(profilePath, "utf-8"))).toEqual({
version: "1.0",
authentication: { type: "oidc_federation" },
});
} finally {
handle?.stop();
}
});
test("derives the config dir from the federation inputs", async () => {
process.env.ANTHROPIC_FEDERATION_RULE_ID = "fdrl_test";
process.env.ANTHROPIC_ORGANIZATION_ID =
"00000000-0000-0000-0000-000000000000";
process.env.ANTHROPIC_WORKSPACE_ID = "wrkspc_a";
(await setupWorkloadIdentity())?.stop();
const firstConfigDir = process.env.ANTHROPIC_CONFIG_DIR;
expect(firstConfigDir).toBeDefined();
// A later step in the same job with a different workspace must not
// share the first step's credentials cache.
delete process.env.ANTHROPIC_CONFIG_DIR;
delete process.env.ANTHROPIC_PROFILE;
process.env.ANTHROPIC_WORKSPACE_ID = "wrkspc_b";
(await setupWorkloadIdentity())?.stop();
const secondConfigDir = process.env.ANTHROPIC_CONFIG_DIR;
expect(secondConfigDir).toBeDefined();
expect(secondConfigDir).not.toBe(firstConfigDir);
// Same inputs land in the same dir, so an unchanged config can still
// reuse a cached token.
delete process.env.ANTHROPIC_CONFIG_DIR;
delete process.env.ANTHROPIC_PROFILE;
(await setupWorkloadIdentity())?.stop();
expect(process.env.ANTHROPIC_CONFIG_DIR).toBe(secondConfigDir!);
});
test("does not overwrite an operator-set ANTHROPIC_PROFILE", async () => {
process.env.ANTHROPIC_FEDERATION_RULE_ID = "fdrl_test";
process.env.ANTHROPIC_ORGANIZATION_ID =
"00000000-0000-0000-0000-000000000000";
process.env.ANTHROPIC_PROFILE = "operator";
const handle = await setupWorkloadIdentity();
try {
expect(process.env.ANTHROPIC_PROFILE).toBe("operator");
expect(process.env.ANTHROPIC_CONFIG_DIR).toBeUndefined();
expect(warningSpy).toHaveBeenCalled();
const entries = readdirSync(join(tempDir, "claude-workload-identity"));
expect(entries.filter((e) => e.startsWith("config-"))).toEqual([]);
// The identity token file is still provisioned for the operator's
// profile (or the env-var fallback) to consume.
expect(process.env.ANTHROPIC_IDENTITY_TOKEN_FILE).toBe(
handle!.tokenFile,
);
} finally {
handle?.stop();
}
});
test("does not overwrite an operator-set ANTHROPIC_CONFIG_DIR", async () => {
process.env.ANTHROPIC_FEDERATION_RULE_ID = "fdrl_test";
process.env.ANTHROPIC_ORGANIZATION_ID =
"00000000-0000-0000-0000-000000000000";
const operatorConfigDir = join(tempDir, "operator-config");
process.env.ANTHROPIC_CONFIG_DIR = operatorConfigDir;
const handle = await setupWorkloadIdentity();
try {
expect(process.env.ANTHROPIC_CONFIG_DIR).toBe(operatorConfigDir);
expect(process.env.ANTHROPIC_PROFILE).toBeUndefined();
expect(warningSpy).toHaveBeenCalled();
} finally {
handle?.stop();
}
});
test("stop removes the identity token and credential cache", async () => {
process.env.ANTHROPIC_FEDERATION_RULE_ID = "fdrl_test";
process.env.ANTHROPIC_ORGANIZATION_ID =
"00000000-0000-0000-0000-000000000000";
const handle = await setupWorkloadIdentity();
const tokenDir = join(tempDir, "claude-workload-identity");
expect(existsSync(handle!.tokenFile)).toBe(true);
expect(existsSync(process.env.ANTHROPIC_CONFIG_DIR!)).toBe(true);
handle!.stop();
expect(existsSync(tokenDir)).toBe(false);
});
});
});
+17 -29
View File
@@ -7,17 +7,17 @@
"dependencies": {
"@actions/core": "^1.10.1",
"@actions/github": "^6.0.1",
"@anthropic-ai/claude-agent-sdk": "^0.3.161",
"@anthropic-ai/claude-agent-sdk": "^0.3.330",
"@modelcontextprotocol/sdk": "^1.11.0",
"@octokit/graphql": "^8.2.2",
"@octokit/rest": "^21.1.1",
"@octokit/webhooks-types": "^7.6.1",
"node-fetch": "^3.3.2",
"shell-quote": "^1.8.3",
"shell-quote": "^1.8.4",
"zod": "^3.24.4",
},
"devDependencies": {
"@types/bun": "1.2.11",
"@types/bun": "1.4.3",
"@types/node": "^20.0.0",
"@types/node-fetch": "^2.6.12",
"@types/shell-quote": "^1.7.5",
@@ -37,23 +37,23 @@
"@actions/io": ["@actions/[email protected]", "", {}, "sha512-wi9JjgKLYS7U/z8PPbco+PvTb/nRWjeoFlJ1Qer83k/3C5PHQi28hiVdeE2kHXmIL99mQFawx8qt/JPjZilJ8Q=="],
"@anthropic-ai/claude-agent-sdk": ["@anthropic-ai/[email protected].161", "", { "optionalDependencies": { "@anthropic-ai/claude-agent-sdk-darwin-arm64": "0.3.161", "@anthropic-ai/claude-agent-sdk-darwin-x64": "0.3.161", "@anthropic-ai/claude-agent-sdk-linux-arm64": "0.3.161", "@anthropic-ai/claude-agent-sdk-linux-arm64-musl": "0.3.161", "@anthropic-ai/claude-agent-sdk-linux-x64": "0.3.161", "@anthropic-ai/claude-agent-sdk-linux-x64-musl": "0.3.161", "@anthropic-ai/claude-agent-sdk-win32-arm64": "0.3.161", "@anthropic-ai/claude-agent-sdk-win32-x64": "0.3.161" }, "peerDependencies": { "@anthropic-ai/sdk": ">=0.93.0", "@modelcontextprotocol/sdk": "^1.29.0", "zod": "^4.0.0" } }, "sha512-TWVFmPsGePXPNnNuWAAuKagQai9kNj99tuu2KLo8FN3oWF5r82OnWTNpQ3IubpWYPyO1vC0+kQu2GfqoEme/hA=="],
"@anthropic-ai/claude-agent-sdk": ["@anthropic-ai/[email protected].330", "", { "optionalDependencies": { "@anthropic-ai/claude-agent-sdk-darwin-arm64": "0.3.330", "@anthropic-ai/claude-agent-sdk-darwin-x64": "0.3.330", "@anthropic-ai/claude-agent-sdk-linux-arm64": "0.3.330", "@anthropic-ai/claude-agent-sdk-linux-arm64-musl": "0.3.330", "@anthropic-ai/claude-agent-sdk-linux-x64": "0.3.330", "@anthropic-ai/claude-agent-sdk-linux-x64-musl": "0.3.330", "@anthropic-ai/claude-agent-sdk-win32-arm64": "0.3.330", "@anthropic-ai/claude-agent-sdk-win32-x64": "0.3.330" }, "peerDependencies": { "@anthropic-ai/sdk": ">=0.93.0", "@modelcontextprotocol/sdk": "^1.29.0", "zod": "^4.0.0" } }, "sha512-7om9Xus6tfkeCUCfzg6ilJuN6Z7okkEvJ0mx+ukXa/Id5U/6BGtPkOYni+3ebedwa3b9SVF7A05+ln+/9EoU/w=="],
"@anthropic-ai/claude-agent-sdk-darwin-arm64": ["@anthropic-ai/[email protected].161", "", { "os": "darwin", "cpu": "arm64" }, "sha512-KwDpi1O2P++uAskFt454K/sShPmhHSYhSbBxaap2KLFoVFti1pttDcEqDPfmKCk2XzKTwlpnVSI1IwTiawPJTw=="],
"@anthropic-ai/claude-agent-sdk-darwin-arm64": ["@anthropic-ai/[email protected].330", "", { "os": "darwin", "cpu": "arm64" }, "sha512-J9nqvcfD2V90O5uoR8At66R51VYtSMKnxSmPc0S4A4F5oX1VbokcpJLXU1yVCOuTtzr/OIHjRr9nn2vKi/S6rQ=="],
"@anthropic-ai/claude-agent-sdk-darwin-x64": ["@anthropic-ai/[email protected].161", "", { "os": "darwin", "cpu": "x64" }, "sha512-VoLz1BfKDc6UOYL+UhRKNSFA1k4l6kvOUrWqkHJ34V5zI5/7VzQilaGl7hU5JECE1TleXMzrWpDrLZO7akApJA=="],
"@anthropic-ai/claude-agent-sdk-darwin-x64": ["@anthropic-ai/[email protected].330", "", { "os": "darwin", "cpu": "x64" }, "sha512-woe+f9qeZYWML5fACibNyth0xMDf22riNRNDryIYuhdteh7IoWrK6zC31cC2MrZRoqDZcPVtfOLND79+iZbrLA=="],
"@anthropic-ai/claude-agent-sdk-linux-arm64": ["@anthropic-ai/[email protected].161", "", { "os": "linux", "cpu": "arm64" }, "sha512-QQbGTJ+2AFeP6+vsZQtafqeGZGX72jaHOjOyrhEcMk1T2DSrJV05J0xnkZT4yIOsByGosntTcY6963EfkuwuTg=="],
"@anthropic-ai/claude-agent-sdk-linux-arm64": ["@anthropic-ai/[email protected].330", "", { "os": "linux", "cpu": "arm64", "libc": "glibc" }, "sha512-2Q+kfzaoTvJv1rsyu+iR4UHDZ3nxiOfdXISk+2ydQShU9DqXNlgQAt3BdJMh0Yx4MrxB66TMphVvVQeRslwwhA=="],
"@anthropic-ai/claude-agent-sdk-linux-arm64-musl": ["@anthropic-ai/[email protected].161", "", { "os": "linux", "cpu": "arm64" }, "sha512-y+RPkQ6ZFje6LP0WyGvM4yHp7IQAaxAbcKSP9QVzKtT9yWiDeqWSnPzdP2Rp3B+6Kd12MuhwIMAbiKS4onAusA=="],
"@anthropic-ai/claude-agent-sdk-linux-arm64-musl": ["@anthropic-ai/[email protected].330", "", { "os": "linux", "cpu": "arm64", "libc": "musl" }, "sha512-fXmuAEegw31v+kiBd51yn7RVhjs7Rz2lnhfx/+XJXqt5lyqR2v7Ul/U/9r1CQZKuIw8M3ms9jQqGpmKbgKndqQ=="],
"@anthropic-ai/claude-agent-sdk-linux-x64": ["@anthropic-ai/[email protected].161", "", { "os": "linux", "cpu": "x64" }, "sha512-gq4tvNRRcfSf8o35mLWKN351I6FowQa2n28YiZJBvZU1glQ3HKQM4srrXzjwLwv1VcndyWBfLxoaDB1I9BS++w=="],
"@anthropic-ai/claude-agent-sdk-linux-x64": ["@anthropic-ai/[email protected].330", "", { "os": "linux", "cpu": "x64", "libc": "glibc" }, "sha512-9w4uMdA3vNZXocwOUhivOGotXoRDxN/C/xPMkZMY9WPpKLAlj8upMZskNgPjQKlVRXhxPfNagcBijGPUH4WepA=="],
"@anthropic-ai/claude-agent-sdk-linux-x64-musl": ["@anthropic-ai/[email protected].161", "", { "os": "linux", "cpu": "x64" }, "sha512-ev0eJhypbyatvSL+Cl2LqOos5+XZbgBMV6MJk7jIh/9dirMLwDj2JFlLlOnHdmIpFcznESZ/Z2wgMtqmOMUCHQ=="],
"@anthropic-ai/claude-agent-sdk-linux-x64-musl": ["@anthropic-ai/[email protected].330", "", { "os": "linux", "cpu": "x64", "libc": "musl" }, "sha512-kooZMdp6tFwNFkDXbGoNTZMgFcPQfJVfG+gZ2qAdkk11S6MFts9ybHRLQmzUK3VrqlahH7T9fmI2mNfzYgOXkw=="],
"@anthropic-ai/claude-agent-sdk-win32-arm64": ["@anthropic-ai/[email protected].161", "", { "os": "win32", "cpu": "arm64" }, "sha512-toJhVo/7RKyPc/IEGvgIcN/mPmGs38gnuDOrdtdxXoPR7WRkHcHfwprD5S9zV/pZAhuLhj/cznhno3G7b4Pnkg=="],
"@anthropic-ai/claude-agent-sdk-win32-arm64": ["@anthropic-ai/[email protected].330", "", { "os": "win32", "cpu": "arm64" }, "sha512-6jscch7vk/XBSn7mm1i6zwTeiz05q98nKrQ3HdTavQEDGvXycMcB/QxIdMrs3vrdNEjWuoxwrL1qlT1lOyAWLg=="],
"@anthropic-ai/claude-agent-sdk-win32-x64": ["@anthropic-ai/[email protected].161", "", { "os": "win32", "cpu": "x64" }, "sha512-QmYFeX/P7bt4diCZhXgazkmzLxL+uK1zwPaCSL2sIQyYGWAXiUulZ41TVLQbHglCBlKzUzVP9kwEXaAZye7LPA=="],
"@anthropic-ai/claude-agent-sdk-win32-x64": ["@anthropic-ai/[email protected].330", "", { "os": "win32", "cpu": "x64" }, "sha512-ooThWbUXlXKTDgj2VgWBXQ/6vSOTaAbBDWPvvCGNHi20haoSkw4uPX9gczkX+HJuVTslvvLiYIfe8zleVPzi7A=="],
"@anthropic-ai/sdk": ["@anthropic-ai/[email protected]", "", { "dependencies": { "json-schema-to-ts": "^3.1.1" }, "peerDependencies": { "zod": "^3.25.0 || ^4.0.0" }, "optionalPeers": ["zod"], "bin": { "anthropic-ai-sdk": "bin/cli" } }, "sha512-q9vaSZQVFx6B/gPxetGYfLXSJD5v0sOmh0OpZDq7yCrTSA+Rscvrtyol7JJTW40wEpQB4U1B4JXzxQitbQ3CAA=="],
@@ -75,8 +75,6 @@
"@octokit/plugin-paginate-rest": ["@octokit/[email protected]", "", { "dependencies": { "@octokit/types": "^12.6.0" }, "peerDependencies": { "@octokit/core": "5" } }, "sha512-u3KYkGF7GcZnSD/3UP0S7K5XUFT2FkOQdcfXZGZQPGv3lm4F2Xbf71lvjldr8c1H3nNbF+33cLEkWYbokGWqiQ=="],
"@octokit/plugin-request-log": ["@octokit/[email protected]", "", { "peerDependencies": { "@octokit/core": ">=6" } }, "sha512-n/lNeCtq+9ofhC15xzmJCNKP2BWTv8Ih2TTy+jatNCCq/gQP/V7rK3fjIfuz0pDWDALO/o/4QY4hyOF6TQQFUw=="],
"@octokit/plugin-rest-endpoint-methods": ["@octokit/[email protected]", "", { "dependencies": { "@octokit/types": "^12.6.0" }, "peerDependencies": { "@octokit/core": "5" } }, "sha512-xV1b+ceKV9KytQe3zCVqjg+8GTGfDYwaT1ATU5isiUyVtlVAO3HNdzpS4sr4GBx4hxQ46s7ITtZrAsxG22+rVg=="],
"@octokit/request": ["@octokit/[email protected]", "", { "dependencies": { "@octokit/endpoint": "^9.0.6", "@octokit/request-error": "^5.1.1", "@octokit/types": "^13.1.0", "universal-user-agent": "^6.0.0" } }, "sha512-qnB2+SY3hkCmBxZsR/MPCybNmbJe4KAlfWErXq+rBKkQJlbjdJeS85VI9r8UqeLYLvnAenU8Q1okM/0MBsAGXw=="],
@@ -89,7 +87,7 @@
"@octokit/webhooks-types": ["@octokit/[email protected]", "", {}, "sha512-S8u2cJzklBC0FgTwWVLaM8tMrDuDMVE4xiTK4EYXM9GntyvrdbSoxqDQa+Fh57CCNApyIpyeqPhhFEmHPfrXgw=="],
"@types/bun": ["@types/bun@1.2.11", "", { "dependencies": { "bun-types": "1.2.11" } }, "sha512-ZLbbI91EmmGwlWTRWuV6J19IUiUC5YQ3TCEuSHI3usIP75kuoA8/0PVF+LTrbEnVc8JIhpElWOxv1ocI1fJBbw=="],
"@types/bun": ["@types/bun@1.4.3", "", { "dependencies": { "bun-types": "1.4.3" } }, "sha512-N67Pqy8JXbAY/j0tNraS3V5mubGvc6rrrNDmDCtZLlFzccbUJPDb0tvgUBC5q4m8ybkP93Jcn1QB+RYQ6PvBhw=="],
"@types/node": ["@types/[email protected]", "", { "dependencies": { "undici-types": "~6.21.0" } }, "sha512-cuVNgarYWZqxRJDQHEB58GEONhOK79QVR/qYx4S7kcUObQvUwvFnYxJuuHUKm2aieN9X3yZB4LZsuYNU1Qphsw=="],
@@ -107,7 +105,7 @@
"body-parser": ["[email protected]", "", { "dependencies": { "bytes": "^3.1.2", "content-type": "^1.0.5", "debug": "^4.4.0", "http-errors": "^2.0.0", "iconv-lite": "^0.6.3", "on-finished": "^2.4.1", "qs": "^6.14.0", "raw-body": "^3.0.0", "type-is": "^2.0.0" } }, "sha512-02qvAaxv8tp7fBa/mw1ga98OGm+eCbqzJOKoRt70sLmfEEi+jyBYVTDGfCL/k06/4EMk/z01gCe7HoCH/f2LTg=="],
"bun-types": ["bun-types@1.2.11", "", { "dependencies": { "@types/node": "*" } }, "sha512-dbkp5Lo8HDrXkLrONm6bk+yiiYQSntvFUzQp0v3pzTAsXk6FtgVMjdQ+lzFNVAmQFUkPQZ3WMZqH5tTo+Dp/IA=="],
"bun-types": ["bun-types@1.4.3", "", { "dependencies": { "@types/node": "*", "undici-types": "*" } }, "sha512-tinRJNGOmUN+tEAsn2y2e942n9JHRHR42iUYzDH75NmhJqVBokGSQkkfSFdNWwsANIdrK6ACH9t01M452YyTwg=="],
"bytes": ["[email protected]", "", {}, "sha512-/Nf7TyzTx6S3yRJObOAV7956r8cr2+Oj8AC5dt8wSP3BQAoeX58NoHyCU8P8zGkNXStjTSi6fzO6F0pBdcYbEg=="],
@@ -275,7 +273,7 @@
"shebang-regex": ["[email protected]", "", {}, "sha512-7++dFhtcx3353uBaq8DDR4NuxBetBzC7ZQOhmTQInHEd6bSrXdiEyzCvG07Z44UYdLShWUyXt5M/yhz8ekcb1A=="],
"shell-quote": ["[email protected].3", "", {}, "sha512-ObmnIF4hXNg1BqhnHmgbDETF8dLPCggZWBjkQfhZpbszZnYur5DUljTcCHii5LC3J5E0yeO/1LIMyH+UvHQgyw=="],
"shell-quote": ["[email protected].4", "", {}, "sha512-VsC6n6vz1ihYYyZZwX7YZSF5l5x36ca17OC+a69h94YqB7X6XLwf+5MOgynYir2SLFUbl8gIYvBo8K8RoNQ6bQ=="],
"side-channel": ["[email protected]", "", { "dependencies": { "es-errors": "^1.3.0", "object-inspect": "^1.13.3", "side-channel-list": "^1.0.0", "side-channel-map": "^1.0.1", "side-channel-weakmap": "^1.0.2" } }, "sha512-ZX99e6tRweoUXqR+VBrslhda51Nh5MTQwou5tnUDgbtyM0dBgmhEDtWGP/xbKn6hqfPRHujUNwz5fy/wbbhnpw=="],
@@ -333,8 +331,6 @@
"@octokit/plugin-paginate-rest/@octokit/types": ["@octokit/[email protected]", "", { "dependencies": { "@octokit/openapi-types": "^20.0.0" } }, "sha512-1rhSOfRa6H9w4YwK0yrf5faDaDTb+yLyBUKOCV4xtCDB5VmIPqd/v9yr9o6SAzOAlRxMiRiCic6JVM1/kunVkw=="],
"@octokit/plugin-request-log/@octokit/core": ["@octokit/[email protected]", "", { "dependencies": { "@octokit/auth-token": "^5.0.0", "@octokit/graphql": "^8.2.2", "@octokit/request": "^9.2.3", "@octokit/request-error": "^6.1.8", "@octokit/types": "^14.0.0", "before-after-hook": "^3.0.2", "universal-user-agent": "^7.0.0" } }, "sha512-kIU8SLQkYWGp3pVKiYzA5OSaNF5EE03P/R8zEmmrG6XwOg5oBjXyQVVIauQ0dgau4zYhpZEhJrvIYt6oM+zZZA=="],
"@octokit/plugin-rest-endpoint-methods/@octokit/types": ["@octokit/[email protected]", "", { "dependencies": { "@octokit/openapi-types": "^20.0.0" } }, "sha512-1rhSOfRa6H9w4YwK0yrf5faDaDTb+yLyBUKOCV4xtCDB5VmIPqd/v9yr9o6SAzOAlRxMiRiCic6JVM1/kunVkw=="],
"@octokit/request/@octokit/types": ["@octokit/[email protected]", "", { "dependencies": { "@octokit/openapi-types": "^24.2.0" } }, "sha512-ifLaO34EbbPj0Xgro4G5lP5asESjwHracYJvVaPIyXMuiuXLlhic3S47cBdTb+jfODkTE5YtGCLt3Ay3+J97sA=="],
@@ -347,6 +343,8 @@
"@octokit/rest/@octokit/plugin-paginate-rest": ["@octokit/[email protected]", "", { "dependencies": { "@octokit/types": "^13.10.0" }, "peerDependencies": { "@octokit/core": ">=6" } }, "sha512-n5KPteiF7pWKgBIBJSk8qzoZWcUkza2O6A0za97pMGVrGfPdltxrfmfF5GucHYvHGZD8BdaZmmHGz5cX/3gdpw=="],
"@octokit/rest/@octokit/plugin-request-log": ["@octokit/[email protected]", "", { "peerDependencies": { "@octokit/core": ">=6" } }, "sha512-n/lNeCtq+9ofhC15xzmJCNKP2BWTv8Ih2TTy+jatNCCq/gQP/V7rK3fjIfuz0pDWDALO/o/4QY4hyOF6TQQFUw=="],
"@octokit/rest/@octokit/plugin-rest-endpoint-methods": ["@octokit/[email protected]", "", { "dependencies": { "@octokit/types": "^13.10.0" }, "peerDependencies": { "@octokit/core": ">=6" } }, "sha512-9Pas60Iv9ejO3WlAX3maE1+38c5nqbJXV5GrncEfkndIpZrJ/WPMRd2xYDcPPEt5yzpxcjw9fWNoPhsSGzqKqw=="],
"accepts/mime-types": ["[email protected]", "", { "dependencies": { "mime-db": "^1.54.0" } }, "sha512-xRc4oEhT6eaBpU1XF7AjpOFD+xQmXNB5OVKwp4tqCuBpHLS/ZbBDrc07mYTDqVMg6PfxUjjNp85O6Cd2Z/5HWA=="],
@@ -367,14 +365,6 @@
"@octokit/plugin-paginate-rest/@octokit/types/@octokit/openapi-types": ["@octokit/[email protected]", "", {}, "sha512-EtqRBEjp1dL/15V7WiX5LJMIxxkdiGJnabzYx5Apx4FkQIFgAfKumXeYAqqJCj1s+BMX4cPFIFC4OLCR6stlnA=="],
"@octokit/plugin-request-log/@octokit/core/@octokit/auth-token": ["@octokit/[email protected]", "", {}, "sha512-JcQDsBdg49Yky2w2ld20IHAlwr8d/d8N6NiOXbtuoPCqzbsiJgF633mVUw3x4mo0H5ypataQIX7SFu3yy44Mpw=="],
"@octokit/plugin-request-log/@octokit/core/@octokit/request": ["@octokit/[email protected]", "", { "dependencies": { "@octokit/endpoint": "^10.1.4", "@octokit/request-error": "^6.1.8", "@octokit/types": "^14.0.0", "fast-content-type-parse": "^2.0.0", "universal-user-agent": "^7.0.2" } }, "sha512-q8ybdytBmxa6KogWlNa818r0k1wlqzNC+yNkcQDECHvQo8Vmstrg18JwqJHdJdUiHD2sjlwBgSm9kHkOKe2iyA=="],
"@octokit/plugin-request-log/@octokit/core/@octokit/request-error": ["@octokit/[email protected]", "", { "dependencies": { "@octokit/types": "^14.0.0" } }, "sha512-WEi/R0Jmq+IJKydWlKDmryPcmdYSVjL3ekaiEL1L9eo1sUnqMJ+grqmC9cjk7CA7+b2/T397tO5d8YLOH3qYpQ=="],
"@octokit/plugin-request-log/@octokit/core/before-after-hook": ["[email protected]", "", {}, "sha512-Nik3Sc0ncrMK4UUdXQmAnRtzmNQTAAXmXIopizwZ1W1t8QmfJj+zL4OA2I7XPTPW5z5TDqv4hRo/JzouDJnX3A=="],
"@octokit/plugin-rest-endpoint-methods/@octokit/types/@octokit/openapi-types": ["@octokit/[email protected]", "", {}, "sha512-EtqRBEjp1dL/15V7WiX5LJMIxxkdiGJnabzYx5Apx4FkQIFgAfKumXeYAqqJCj1s+BMX4cPFIFC4OLCR6stlnA=="],
"@octokit/request-error/@octokit/types/@octokit/openapi-types": ["@octokit/[email protected]", "", {}, "sha512-9sIH3nSUttelJSXUrmGzl7QUBFul0/mB8HRYl3fOlgHbIWG+WnYDXU3v/2zMtAvuzZ/ed00Ei6on975FhBfzrg=="],
@@ -401,8 +391,6 @@
"type-is/mime-types/mime-db": ["[email protected]", "", {}, "sha512-aU5EJuIN2WDemCcAp2vFBfp/m4EAhWJnUNSSw0ixs7/kXbd6Pg64EmwJkNdFhB8aWt1sH2CTXrLxo/iAGV3oPQ=="],
"@octokit/plugin-request-log/@octokit/core/@octokit/request/@octokit/endpoint": ["@octokit/[email protected]", "", { "dependencies": { "@octokit/types": "^14.0.0", "universal-user-agent": "^7.0.2" } }, "sha512-OlYOlZIsfEVZm5HCSR8aSg02T2lbUWOsCQoPKfTXJwDzcHQBrVBGdGXb89dv2Kw2ToZaRtudp8O3ZIYoaOjKlA=="],
"@octokit/rest/@octokit/core/@octokit/request/@octokit/endpoint": ["@octokit/[email protected]", "", { "dependencies": { "@octokit/types": "^14.0.0", "universal-user-agent": "^7.0.2" } }, "sha512-OlYOlZIsfEVZm5HCSR8aSg02T2lbUWOsCQoPKfTXJwDzcHQBrVBGdGXb89dv2Kw2ToZaRtudp8O3ZIYoaOjKlA=="],
"@octokit/rest/@octokit/plugin-paginate-rest/@octokit/types/@octokit/openapi-types": ["@octokit/[email protected]", "", {}, "sha512-9sIH3nSUttelJSXUrmGzl7QUBFul0/mB8HRYl3fOlgHbIWG+WnYDXU3v/2zMtAvuzZ/ed00Ei6on975FhBfzrg=="],
+5 -2
View File
@@ -15,12 +15,15 @@
## What Claude Cannot Do
The action doesn't give Claude tools for these. In [auto mode](./usage.md#permission-mode), the default, a tool call outside your allow rules can still run if the safety review passes it. For a strict list, use `--permission-mode acceptEdits`.
- **Submit PR Reviews**: Claude cannot submit formal GitHub PR reviews
- **Approve PRs**: For security reasons, Claude cannot approve pull requests
- **Push to Other Branches**: By default, Claude can only push to the branch it was set up on: a new branch for an issue or closed PR, or the branch of an open PR. For a PR from a fork, that is a `pr-N` branch in your repository, and the fork's branch is not updated. [Keep that branch out of CI that holds secrets](./security.md#running-code-from-pull-requests)
- **Post Multiple Comments**: Claude only acts by updating its initial comment
- **Execute Commands Outside Its Context**: Claude only has access to the repository and PR/issue context it's triggered in
- **Run Arbitrary Bash Commands**: By default, Claude cannot execute Bash commands unless explicitly allowed using the `allowed_tools` configuration
- **Perform Branch Operations**: Cannot merge branches, rebase, or perform other git operations beyond pushing commits
- **Run Bash Commands Unchecked**: A Bash command runs only if an allow rule covers it or the safety review passes it
- **Perform Branch Operations**: Claude is instructed not to merge branches, rebase, or perform other git operations beyond pushing commits
## How It Works
+1 -1
View File
@@ -7,7 +7,7 @@ You can authenticate with Claude using any of these four methods:
3. Google Vertex AI with OIDC authentication
4. Microsoft Foundry with OIDC authentication
For detailed setup instructions for AWS Bedrock and Google Vertex AI, see the [official documentation](https://code.claude.com/docs/en/github-actions#for-aws-bedrock:).
For detailed setup instructions for AWS Bedrock and Google Vertex AI, see the [official documentation](https://code.claude.com/docs/en/github-actions#using-with-amazon-bedrock-and-google-cloud).
**Note**:
+40 -12
View File
@@ -140,10 +140,13 @@ To allow Claude to view workflow run results, job logs, and CI status:
3. **Claude will automatically get access to CI/CD tools**:
When you enable `actions: read`, Claude can use the following MCP tools:
- `mcp__github_ci__get_ci_status` - View workflow run statuses
- `mcp__github_ci__get_workflow_run_details` - Get detailed workflow information
- `mcp__github_ci__download_job_log` - Download and analyze job logs
On pull requests from outside contributors, the contributor can control the text of job logs. See [Security](./security.md#running-code-from-pull-requests) for how to turn off the tools that read them.
### Example: Debugging Failed CI Runs
```yaml
@@ -228,7 +231,7 @@ By default, Claude only has access to:
- Comment management (creating/updating comments)
- Basic GitHub operations
Claude does **not** have access to execute arbitrary Bash commands by default. If you want Claude to run specific commands (e.g., npm install, npm test), you must explicitly allow them using the `claude_args` configuration:
In [auto mode](./usage.md#permission-mode), the default, a safety review decides any Bash command that no allow rule covers. To let specific commands (e.g., npm install, npm test) run without review, or to allow them under `--permission-mode acceptEdits`, add them with `--allowedTools`:
**Note**: If your repository has a `.mcp.json` file in the root directory, Claude will automatically detect and use the MCP server tools defined there. However, these tools still need to be explicitly allowed.
@@ -241,6 +244,8 @@ Claude does **not** have access to execute arbitrary Bash commands by default. I
# ... other inputs
```
**Note**: On pull requests whose author does not have write access, commands that commonly run repository code, such as `npm`, are denied even if you allow them here. See [Security](./security.md#running-code-from-pull-requests).
**Note**: The base GitHub tools are always included. Use `--allowedTools` to add additional tools (including specific Bash commands), and `--disallowedTools` to prevent specific tools from being used.
## Custom Model
@@ -275,6 +280,29 @@ For provider-specific models:
# ... other inputs
```
### 1M context models through an API gateway
When `ANTHROPIC_BASE_URL` points to an Anthropic-compatible API gateway,
Claude Code may not be able to verify that the gateway supports a model's native
1M context window and can budget the session at 200K instead. Append the
`[1m]` selector to explicitly use the 1M context window for supported models,
including Claude Opus 5 and Claude Sonnet 5:
```yaml
- uses: anthropics/claude-code-action@v1
with:
claude_args: |
--model "claude-opus-5[1m]"
# ... other inputs
```
Use the same selector when setting a model through `ANTHROPIC_MODEL` or another
Claude Code model environment variable. The selector is resolved by Claude Code
before requests are sent to the provider. The action's sanitized result output
includes each model's resolved
`contextWindow` and `maxOutputTokens` under `modelUsage`, so these limits are
visible without enabling `show_full_output`.
## Claude Code Settings
You can provide Claude Code settings to customize behavior such as model selection, environment variables, permissions, and hooks. Settings can be provided either as a JSON string or a path to a settings file.
@@ -337,17 +365,17 @@ For a complete list of available settings and their descriptions, see the [Claud
Many individual input parameters have been consolidated into `claude_args` or `settings`. Here's how to migrate:
| Old Input | New Approach |
| --------------------- | -------------------------------------------------------- |
| `allowed_tools` | Use `claude_args: "--allowedTools Tool1,Tool2"` |
| `disallowed_tools` | Use `claude_args: "--disallowedTools Tool1,Tool2"` |
| `max_turns` | Use `claude_args: "--max-turns 10"` |
| `model` | Use `claude_args: "--model claude-4-0-sonnet-20250805"` |
| `claude_env` | Use `settings` with `"env"` object |
| `custom_instructions` | Use `claude_args: "--system-prompt 'Your instructions'"` |
| `mcp_config` | Use `claude_args: "--mcp-config '{...}'"` |
| `direct_prompt` | Use `prompt` input instead |
| `override_prompt` | Use `prompt` with GitHub context variables |
| Old Input | New Approach |
| --------------------- | --------------------------------------------------------------- |
| `allowed_tools` | Use `claude_args: "--allowedTools Tool1,Tool2"` |
| `disallowed_tools` | Use `claude_args: "--disallowedTools Tool1,Tool2"` |
| `max_turns` | Use `claude_args: "--max-turns 10"` |
| `model` | Use `claude_args: "--model claude-4-0-sonnet-20250805"` |
| `claude_env` | Use `settings` with `"env"` object |
| `custom_instructions` | Use `claude_args: "--append-system-prompt 'Your instructions'"` |
| `mcp_config` | Use `claude_args: "--mcp-config '{...}'"` |
| `direct_prompt` | Use `prompt` input instead |
| `override_prompt` | Use `prompt` with GitHub context variables |
## Custom Executables for Specialized Environments
+2 -2
View File
@@ -26,7 +26,7 @@ This action supports the following GitHub events ([learn more GitHub event trigg
## Automated Documentation Updates
Automatically update documentation when specific files change (see [`examples/claude-pr-path-specific.yml`](../examples/claude-pr-path-specific.yml)):
Automatically update documentation when specific files change (see [`examples/pr-review-filtered-paths.yml`](../examples/pr-review-filtered-paths.yml)):
```yaml
on:
@@ -47,7 +47,7 @@ When API files are modified, the action automatically detects that a `prompt` is
## Author-Specific Code Reviews
Automatically review PRs from specific authors or external contributors (see [`examples/claude-review-from-author.yml`](../examples/claude-review-from-author.yml)):
Automatically review PRs from specific authors or external contributors (see [`examples/pr-review-filtered-authors.yml`](../examples/pr-review-filtered-authors.yml)):
```yaml
on:
+8 -11
View File
@@ -63,17 +63,14 @@ The GitHub App for Claude doesn't have workflow write access for security reason
### Why won't Claude rebase my branch?
By default, Claude only uses commit tools for non-destructive changes to the branch. Claude is configured to:
Claude only creates and pushes commits. It does not merge branches, rebase, force push, or perform other destructive git operations. Specifically, Claude is configured to:
- Never push to branches other than where it was invoked (either its own branch or the PR branch)
- Never force push or perform destructive operations
You can grant additional tools via the `claude_args` input if needed:
This restriction is enforced in Claude's system prompt, so it applies even if you grant the underlying git tools (for example `--allowedTools "Bash(git rebase:*)"`). In that case Claude will still decline rebase requests and explain the limitation rather than running the command.
```yaml
claude_args: |
--allowedTools "Bash(git rebase:*)" # Use with caution
```
If you need to rebase, do it yourself locally — or with the Claude Code CLI outside of this action — and push the result.
### Why won't Claude create a pull request?
@@ -101,7 +98,7 @@ Yes! Claude can access GitHub Actions workflow runs, job logs, and test results
actions: read
```
Claude will then be able to analyze CI failures and help debug workflow issues. For running tests locally before commits, you can still instruct Claude to do so in your request.
Claude will then be able to analyze CI failures and help debug workflow issues. For running tests locally before commits, you can still instruct Claude to do so in your request, except on pull requests whose author does not have write access. For pull requests from forks, see [Security](./security.md#running-code-from-pull-requests) for another way to get test results.
### Why does Claude only update one comment instead of creating new ones?
@@ -156,7 +153,7 @@ prompt: "Review this PR for security vulnerabilities"
**These inputs are deprecated in v1.0:**
- **`direct_prompt`** → Use `prompt` instead
- **`custom_instructions`** → Use `claude_args` with `--system-prompt`
- **`custom_instructions`** → Use `claude_args` with `--append-system-prompt` (appends to the default system prompt, matching v0 behavior; `--system-prompt` replaces it entirely)
Migration examples:
@@ -168,16 +165,16 @@ custom_instructions: "Focus on security"
# New (v1.0)
prompt: "Review this PR"
claude_args: |
--system-prompt "Focus on security"
--append-system-prompt "Focus on security"
```
### Why doesn't Claude execute my bash commands?
The Bash tool is **disabled by default** for security. To enable individual bash commands using `claude_args`:
In [auto mode](./usage.md#permission-mode), the default, a safety review decides any Bash command that no allow rule covers, and it can block yours. With `--permission-mode acceptEdits` or `default`, most Bash commands are denied unless allowed. To allow individual bash commands using `claude_args`:
```yaml
claude_args: |
--allowedTools "Bash(npm:*),Bash(git:*)" # Allows only npm and git commands
--allowedTools "Bash(npm:*),Bash(git:*)" # Allows npm and git commands without asking
```
### Can Claude work across multiple repositories?
+26 -23
View File
@@ -14,19 +14,19 @@ This guide helps you migrate from Claude Code Action v0.x to v1.0. The new versi
The following inputs have been deprecated and replaced:
| Deprecated Input | Replacement | Notes |
| --------------------- | ------------------------------------ | --------------------------------------------- |
| `mode` | Auto-detected | Action automatically chooses based on context |
| `direct_prompt` | `prompt` | Direct drop-in replacement |
| `override_prompt` | `prompt` | Use GitHub context variables instead |
| `custom_instructions` | `claude_args: --system-prompt` | Move to CLI arguments |
| `max_turns` | `claude_args: --max-turns` | Use CLI format |
| `model` | `claude_args: --model` | Specify via CLI |
| `allowed_tools` | `claude_args: --allowedTools` | Use CLI format |
| `disallowed_tools` | `claude_args: --disallowedTools` | Use CLI format |
| `claude_env` | `settings` with env object | Use settings JSON |
| `mcp_config` | `claude_args: --mcp-config` | Pass MCP config via CLI arguments |
| `timeout_minutes` | Use GitHub Actions `timeout-minutes` | Configure at job level instead of input level |
| Deprecated Input | Replacement | Notes |
| --------------------- | ------------------------------------- | ----------------------------------------------------------------------------------- |
| `mode` | Auto-detected | Action automatically chooses based on context |
| `direct_prompt` | `prompt` | Direct drop-in replacement |
| `override_prompt` | `prompt` | Use GitHub context variables instead |
| `custom_instructions` | `claude_args: --append-system-prompt` | Appends to the default prompt (v0 behavior); `--system-prompt` replaces it entirely |
| `max_turns` | `claude_args: --max-turns` | Use CLI format |
| `model` | `claude_args: --model` | Specify via CLI |
| `allowed_tools` | `claude_args: --allowedTools` | Use CLI format |
| `disallowed_tools` | `claude_args: --disallowedTools` | Use CLI format |
| `claude_env` | `settings` with env object | Use settings JSON |
| `mcp_config` | `claude_args: --mcp-config` | Pass MCP config via CLI arguments |
| `timeout_minutes` | Use GitHub Actions `timeout-minutes` | Configure at job level instead of input level |
## Migration Examples
@@ -52,7 +52,7 @@ The following inputs have been deprecated and replaced:
anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
claude_args: |
--max-turns 10
--system-prompt "Follow our coding standards"
--append-system-prompt "Follow our coding standards"
--allowedTools Edit,Read,Write
```
@@ -142,6 +142,8 @@ The `track_progress` input only works with these GitHub events:
- `labeled` - Label added to issue
- `assigned` - Issue assigned to user
> **Note**: On `labeled` and `assigned` issue events with `track_progress` and a `prompt`, the label must match `label_trigger` (default `claude`), and the assignee must match `assignee_trigger` if you set one. To use another label, set `label_trigger` to it. To filter labels in your workflow's `if:`, set `label_trigger: ""`, which accepts any label when a `prompt` is set.
> **Note**: Using `track_progress: true` with unsupported events will cause an error.
### Custom Template with Variables
@@ -255,14 +257,15 @@ claude_args: |
### Common claude_args Options
| Option | Description | Example |
| ------------------- | ------------------------ | -------------------------------------- |
| `--max-turns` | Limit conversation turns | `--max-turns 10` |
| `--model` | Specify Claude model | `--model claude-4-0-sonnet-20250805` |
| `--allowedTools` | Enable specific tools | `--allowedTools Edit,Read,Write` |
| `--disallowedTools` | Disable specific tools | `--disallowedTools WebSearch` |
| `--system-prompt` | Add system instructions | `--system-prompt "Focus on security"` |
| `--mcp-config` | Add MCP server config | `--mcp-config '{"mcpServers": {...}}'` |
| Option | Description | Example |
| ------------------------ | ------------------------------------------------------------------------- | ------------------------------------------------------ |
| `--max-turns` | Limit conversation turns | `--max-turns 10` |
| `--model` | Specify Claude model | `--model claude-4-0-sonnet-20250805` |
| `--allowedTools` | Enable specific tools | `--allowedTools Edit,Read,Write` |
| `--disallowedTools` | Disable specific tools | `--disallowedTools WebSearch` |
| `--system-prompt` | Replace the entire default system prompt | `--system-prompt "Focus on security"` |
| `--append-system-prompt` | Append to the default system prompt (keeps Claude Code's built-in prompt) | `--append-system-prompt "Follow our coding standards"` |
| `--mcp-config` | Add MCP server config | `--mcp-config '{"mcpServers": {...}}'` |
## Provider-Specific Updates
@@ -330,7 +333,7 @@ You can also pass MCP configuration from a file:
- [ ] Remove `mode` input (auto-detected now)
- [ ] Replace `direct_prompt` with `prompt`
- [ ] Replace `override_prompt` with `prompt` using GitHub context
- [ ] Move `custom_instructions` to `claude_args` with `--system-prompt`
- [ ] Move `custom_instructions` to `claude_args` with `--append-system-prompt`
- [ ] Convert `max_turns` to `claude_args` with `--max-turns`
- [ ] Convert `model` to `claude_args` with `--model`
- [ ] Convert `allowed_tools` to `claude_args` with `--allowedTools`
+105 -3
View File
@@ -2,9 +2,12 @@
## Access Control
- **Repository Access**: The action can only be triggered by users with write access to the repository
- **Text From Other People**: Triggering a run does not make someone else's text a request. When a maintainer labels or assigns another person's issue, slash commands in its body do not run, but Claude still reads the body to understand the task, so review it first. A pull request's body can also start a run when a maintainer pushes to, reopens or labels it, where `track_progress` is set. Editing or dismissing another person's comment or review, or deleting any comment, does not start a run, even with a `prompt`.
- **Repository Access**: The action can only be triggered by users with write access to the repository. This is checked for issue, pull request, comment, and review events, and for `workflow_run` events, where both the workflow actor and the actor that started the upstream run are checked. `workflow_dispatch`, `repository_dispatch`, and `schedule` events are not checked separately — GitHub itself requires write access to dispatch a workflow, and scheduled runs have no external actor.
- **Bot User Control**: By default, GitHub Apps and bots cannot trigger this action for security reasons. Use the `allowed_bots` parameter to enable specific bots or all bots
- **⚠️ Allowed bots are not checked for repository permissions.** A bot that matches an entry does **not** need to be installed on your repository or have write access. On a **public repository**, external parties — including GitHub Apps created by anyone — may be able to trigger workflow events such as opening issues, commenting, or reviewing pull requests. If your workflow listens on those events and `allowed_bots` is set to `'*'`, any such App can invoke this action with a prompt it controls.
- On `pull_request` events, a bot's activity does not vouch for the pull request. Before a run starts, its author must also have write access or be an allowed bot.
- When `allowed_bots` is set, the token stays out of `.git/config` and git gets it from a credential helper that reads `GH_TOKEN`. A later step that uses the `origin` remote to push, or to fetch, pull or update submodules on a private repository, then needs its own credentials. `origin` names the username `x-access-token`, so those must come from `GH_TOKEN`, `GIT_ASKPASS`, a credential saved under that username, or a remote URL the step sets itself. The token is still in Claude's environment.
- Prefer an explicit list over `'*'`
- Only list App names you trust
- If you need `'*'`, scope workflow `permissions:` to the minimum required
@@ -15,6 +18,7 @@
- Is designed for automation workflows where user permissions are already restricted by the workflow's permission scope
- When set, Claude does a best-effort scrub of Anthropic, cloud, and GitHub Actions secrets from subprocess environments. On Linux runners with bubblewrap available, subprocesses additionally run with PID-namespace isolation. This reduces but does not eliminate prompt injection risk — keep workflow permissions minimal and validate all outputs. Set `CLAUDE_CODE_SUBPROCESS_ENV_SCRUB: 0` in your workflow or job `env:` block to opt out.
- Optionally set `CLAUDE_CODE_SCRIPT_CAPS` in your workflow `env:` block to limit how many times Claude can call specific scripts per run. Value is JSON: `{"script-name.sh": maxCalls}`. Example: `CLAUDE_CODE_SCRIPT_CAPS: '{"edit-issue-labels.sh":2}'` allows at most 2 calls to `edit-issue-labels.sh`. Useful for write-capable helper scripts.
- Runs with this input or `allowed_bots` keep the earlier permission defaults instead of auto mode, so the safety review can't pass a call those defaults deny, unless `claude_args` sets another [permission mode](./usage.md#permission-mode)
- When using `allowed_non_write_users`, always pass `github_token: ${{ secrets.GITHUB_TOKEN }}`. The auto-generated workflow token is scoped to the job's declared permissions and expires when the job completes. **Do not use a personal access token** — a static token does not rotate between runs and could be partially or fully recovered over time via prompt injection. Restricting allowed tools via `claude_args` reduces the rate of recovery but may not eliminate the risk. We recommend restricting allowed tools (e.g. `claude_args: '--allowedTools "Bash(gh issue view:*)"'`) to the minimum required when using `allowed_non_write_users`.
- **Token Permissions**: The GitHub app receives only a short-lived token scoped specifically to the repository it's operating in
- **No Cross-Repository Access**: Each action invocation is limited to the repository where it was triggered
@@ -22,6 +26,8 @@
## Using this action with `pull_request_target` or `workflow_run`
For `workflow_run` events, the action checks the repository access of the actor that started the upstream run (for example, the author of the fork pull request that triggered your CI workflow) in addition to the workflow actor. When the run came from a fork, it also checks the fork's owner, because GitHub credits a run to whoever caused it, such as a maintainer reopening the pull request. If any of these accounts does not have write access, the action stops before running Claude. To run on `workflow_run` events downstream of pull requests from contributors without write access, add those users to `allowed_non_write_users` and pass `github_token: ${{ secrets.GITHUB_TOKEN }}` — see the notes on that input above and keep the workflow's permissions minimal.
`pull_request_target` and `workflow_run` execute with the **base repository's secrets**. If your workflow checks out the PR head (`ref: ${{ github.event.pull_request.head.sha }}` for `pull_request_target`, `ref: ${{ github.event.workflow_run.head_sha }}` for `workflow_run`) into `$GITHUB_WORKSPACE` before this action, the action and Claude run with that checkout as the working directory.
**Do not check out an untrusted ref into the workspace root before this action.** Use one of these patterns instead:
@@ -49,15 +55,105 @@
This is general guidance for these event types — see [GitHub's documentation](https://securitylab.github.com/research/github-actions-preventing-pwn-requests/).
### Which files come from the base branch on pull requests
When the action runs against a pull request, it restores a fixed list of Claude configuration paths from the PR base branch before starting Claude: `.claude/`, `.mcp.json`, `.claude.json`, `.gitmodules`, `.ripgreprc`, `CLAUDE.md`, `CLAUDE.local.md`, `AGENTS.md`, and `.husky/`. Paths in that list that do not exist on the base branch are removed, and the PR-authored versions are kept under `.claude-pr/` for reference only.
On a pull request whose author does not have write access, everything this section takes from the base branch comes from your default branch instead, even if the pull request targets another branch such as a release branch. To use a different branch for pull requests that don't target the default branch, set the `base_branch` input to a fixed branch name, never to a value from the pull request. On pull request and review events, the default branch is used if the input equals the pull request's head branch name. Changes to these files are left out when Claude commits with its commit command. A bot's pull request from a branch in your repository still uses its base, unless that is a `pr-N` branch or a comment in the pull request's conversation started the run and the pull request is closed or the workflow has a `prompt` without `track_progress`. `restrict_untrusted_prs: false` turns this off.
Everything else in the working tree — including `package.json`, lockfiles, `Makefile`, `node_modules/`, and formatter/linter config files — stays at the PR head, apart from the files described below. If a hook, `apiKeyHelper`, or `statusLine` command in your base-branch `.claude/settings.json` runs a package-manager script (`bun run …`, `npm run …`, `yarn …`, `pnpm run …`), a `make` target, a repo-relative script, or a tool that loads executable project config, that command resolves through files the pull request supplies. Keep such commands self-contained: invoke the tool directly with a pinned version and pass its configuration on the command line (for example `bunx [email protected] --no-config --write .` rather than `bun run format`).
On pull requests from authors without write access, the action also replaces what npm and Bun read from the repository root with the base branch's version, wherever the two differ: `.npmrc`, `bunfig.toml` and tracked files in `node_modules/`. In a `package.json` whose `name`, `bin`, `directories.bin` or `workspaces` change the commands it provides, only `name`, `bin`, `directories` and `workspaces` are replaced, or the whole file if it can't be understood, including in other directories if the base branch uses workspaces. The pull request's versions are kept under `.claude-pr/` for reading, except for `node_modules/`. Claude can't edit `.npmrc`, `bunfig.toml`, `node_modules/` or, where one was replaced, `package.json` on these pull requests, in any directory. Its Bash commands start in the repository root there, because a hook runs in the session's current directory and only the root's files are replaced. The replaced files stay in the working tree for later steps in the job, and if Claude stages everything, its commit can include them, as with the paths above. `restrict_untrusted_prs: false` turns this off.
On these pull requests, files and directories in the checkout that your workflow's inputs name also come from the base branch where the pull request changes them or a `.gitattributes` that applies to them: `settings`, `plugin_marketplaces`, and in `claude_args` `--settings`, `--mcp-config`, `--plugin-dir`, `--plugin-dir-no-mcp`, `--agents`, `--project-config-root`, `--system-prompt-file`, `--append-system-prompt-file` and `--append-subagent-system-prompt-file`. The pull request's versions are kept under `.claude-pr/` for reading, and Claude can't edit what these inputs name, apart from the paths listed above. The replaced files stay in the working tree here too. A file that only the pull request has is removed. The run stops if an input names the whole repository as a plugin or marketplace, except that a marketplace at the root works if all its plugins are in subdirectories, if it names a path with `..` after another directory, or if it names files in a submodule that the pull request has at another commit. What symbolic links lead to, files that other config names, for example in a `settings.json`, and executables are not covered. Git LFS files are written as pointers.
This doesn't cover other package managers, the `.env` files and `tsconfig.json` that Bun loads, other files that a hook's own command reads, or a git command that resets the working tree during the session, so keep hooks self-contained as described above. Files that git doesn't track, such as packages an earlier step installed, stay as they are. A file that is the same link on both branches is left as a link.
Note that the runtime executing the tool also reads project config. `bunx <tool>` runs the tool's script under `node` when `node` is on `PATH` (as it is on GitHub-hosted runners); when only Bun is available, Bun executes the script itself and reads `bunfig.toml` from the checkout — including `preload` entries — which comes from the PR head. On such runners, make sure `node` is on `PATH` for the hook, and treat `bunfig.toml` and `.npmrc` in the checkout as PR-controlled runtime config on pull requests from authors with write access.
### Refusing pull requests from forks
By default, someone with write access can have the action work on a pull request from a fork. If you don't need that, turn it off, so that the action itself never checks out a fork's code:
```yaml
- uses: anthropics/claude-code-action@v1
with:
allow_fork_prs: "false"
```
The action then skips a pull request whose branch is in another repository, or in a fork that was deleted. It does so for every event, including comments on the conversation, whoever opened the pull request and whether it is open or closed. It also skips `workflow_run` events for a fork's run. It logs a warning, checks nothing out and doesn't start Claude, and the step succeeds without Claude's outputs, so a required check that only runs the action passes. Any value other than `true` turns forks off, including an empty one.
For a comment on the conversation, the action asks GitHub where the branch is, after it has checked the commenter's write access. The token needs `pull-requests: read` for that. If the answer can't be read, the run fails before anything is checked out.
This covers what the action does with the event that started it. It doesn't cover:
- **Other steps.** A step that checks out a fork's code, such as `actions/checkout` with the head of a `pull_request_target` event, does so before the action can refuse. A step after the action still runs.
- **Events that don't name a pull request**, such as `issues`, `workflow_dispatch`, `repository_dispatch` and `schedule`. A `workflow_run` event is judged by the run right before it, so one that follows a comment's run, or another `workflow_run`, counts as yours.
- **Commands in a session.** On an issue, for example, Claude could fetch a fork's commits with commands that you allow or that auto mode's review passes, such as `git` or `gh`.
- **Commits already in your repository**, such as a `pr-N` branch from an earlier run, or a fork's commits that someone pushed to a branch of yours.
- **Text.** Issues and comments from outside contributors still reach Claude.
### Running code from pull requests
A command that runs repository code (`npm install`, `bun test`, `make`) runs whatever the checkout contains, in a job that holds your workflow's tokens. Tag mode checks out the pull request's head, so on a pull request from an outside contributor that code is theirs, even when a maintainer triggered the run.
By default, when the pull request's author does not have write access, the action denies commands that commonly run repository code (package managers, test runners, interpreters and shells), even if `claude_args` allows them. Claude can still read and edit files, commit and push. Bots count as authors without write access, even if they are listed in `allowed_bots`, and if the check fails the author is treated the same way. Issues and pull requests from authors with write access are not affected. Runs on `workflow_run`, `workflow_dispatch`, `repository_dispatch` and `schedule` events are not covered, so follow the checkout guidance above if those check out pull request code. On a closed or merged pull request, tag mode works from the base branch rather than the pull request's code, so nothing is denied.
- **This is a list of common commands, not a sandbox.** A rule such as `Bash(./scripts/test.sh)` stays allowed, and the action logs a warning with the number of `Bash` rules in `claude_args` that the list does not cover. Rules in the `settings` input or settings files are not checked. Other ways of writing a command, such as `git --no-pager -c`, may not match the list. Add your own rules with `--disallowedTools` if they can run pull request code.
- **Trust follows the pull request's author.** If someone with write access opens a pull request from another person's branch, or pushes someone else's commits, that code is treated as theirs. Review it before asking Claude to run it.
- **Commands outside the list follow your permission mode.** With `--permission-mode auto`, a command that no rule covers runs only if Claude Code's safety review passes it, which we recommend where it is available. In the modes that ask for approval, such a command is denied, because a workflow has no one to approve it. Either way, a rule in `--allowedTools` skips that decision, so don't rely on the review for commands you allow. If `claude_args` turns off permission checks, for example with `bypassPermissions`, commands outside the list run without approval, and the action logs a warning.
- **Text can still steer Claude.** On an issue or a trusted author's pull request these commands stay allowed, so text from an outside contributor could try to get Claude to edit a file and then run it. Keep the rules you grant in `claude_args` as narrow as you can.
- **Other broad rules weaken the protections.** `Write` lets Claude write outside the workspace, and `Bash(git:*)` can still run programs through forms this list does not deny, such as `git rebase --exec`. Avoid them in workflows that handle outside contributions.
**Pushes on pull requests from forks:** in tag mode on an open pull request from a fork, Claude pushes to `pr-N` in your repository. That branch holds the fork's commits as well as Claude's, and GitHub's limits on runs from forks don't apply to it. A workflow or other CI that builds it runs the contributor's code and can have your secrets.
- Keep `pr-*` out of CI that holds secrets, for example with `branches-ignore: ["pr-*"]` on `push` triggers, and check CI outside GitHub too.
- Don't deploy or release from `pr-N`. Review the fork's commits before you merge the branch or open a pull request from it, and delete it when you are done.
- To stop Claude's pushes in a workflow, on all pull requests and issues, add `--disallowedTools "Bash(git push:*)"` to `claude_args`. With `use_commit_signing`, also deny `mcp__github_file_ops`.
**Getting test results:** on pull requests from forks, run the tests in a separate `pull_request` workflow that needs no secrets, and let Claude read the outcome. The pull request's code then runs outside the job that holds Claude's tokens, which is not the case with `restrict_untrusted_prs: false`.
1. In the test workflow, set `permissions: contents: read`. GitHub also limits the token to read access and withholds secrets on pull requests from forks of a public repository. That limit is the real protection, because a pull request can edit the workflow file.
2. Set your repository to require approval for all external contributors. Review the changes, including workflow files, and then approve the runs. GitHub doesn't ask for approval on pull requests from members of your organization.
3. In Claude's workflow, add `actions: read` to the job's `permissions`. In tag mode, Claude can then see the pull request's workflow runs and their results.
4. When the runs finish, mention `@claude` in a comment on the pull request's conversation. Mentions in reviews don't work on pull requests from forks, because GitHub gives those runs no credentials.
Treat what Claude reports as a pointer to the run. Results are for the latest commit on the pull request, so they don't cover edits Claude makes. A run with the conclusion `action_required` is waiting for approval and has neither passed nor failed. A pull request can also change what a workflow does and what it is called.
Job logs, and job and step names, are text the pull request controls. To turn off the tools that read them, add `--disallowedTools "mcp__github_ci__download_job_log,mcp__github_ci__get_workflow_run_details"` to `claude_args`. Claude may still reach logs through other commands that you allow or that auto mode's review passes, such as `gh`.
Bots often open pull requests from branches in your own repository. Runs there usually need no approval and can get secrets, so these steps don't protect you in that case.
**Upgrading:** workflows that use a floating tag or branch such as `@v1` or `@main` get this behavior without changing anything. If a workflow relied on commands like `npm test` running on pull requests from outside contributors or bots, those commands are now denied, Claude's config and tool config come from your default branch or `base_branch` even if the pull request targets another branch, as described [above](#which-files-come-from-the-base-branch-on-pull-requests), Git LFS files from the pull request stay pointer files if they aren't downloaded yet, and a `cd` no longer carries over to the next Bash command. The run still succeeds, so check the log for the warning. Set `restrict_untrusted_prs: false` to turn this off.
### What a session can leave behind
Steps that run after Claude, including post steps such as `actions/checkout`'s cleanup, trust `GITHUB_ENV`, `GITHUB_PATH` and git's config files. When the session ends, the action restores those files, including config pulled in through `include.path`, to how they were before it started, and logs a warning naming any it changed. A session therefore can't pass environment variables to later steps through `GITHUB_ENV`, and `git config` changes it made, such as the upstream that `git push -u` records, don't last. Use `GITHUB_OUTPUT` or `--json-schema` to pass results on. What a session writes to `GITHUB_OUTPUT` and `GITHUB_STEP_SUMMARY` is kept, unless it replaced the file with a link or something else that isn't a file.
While Claude runs, and for the action's own git commands afterwards, git ignores `core.fsmonitor`. Git hooks are off for the action's own commits after the session, and from before the checkout on pull requests that `restrict_untrusted_prs` applies to. During the session these are set in the repository's git config and by the wrapper scripts. Global git config also sets `safe.bareRepository=explicit` for the session, so git won't use a repository it finds in a folder that has no `.git`. A workflow that has Claude work in a bare repository needs to pass `--git-dir`. If a run is cancelled before it restores this, the setting stays in the global config of a self-hosted runner that keeps its home directory. Git LFS uploads through a hook, so the action uploads objects itself for its own commits, but a push Claude makes on such a pull request won't include them.
The action's own git commands use `safe.bareRepository=explicit` too. If the action starts in a bare repository, it doesn't set this before the session, unless `restrict_untrusted_prs` applies. On pull requests that `restrict_untrusted_prs` applies to, the action also checks out the pull request with `GIT_LFS_SKIP_SMUDGE=1`, so files that Git LFS hasn't downloaded yet stay as pointer files. The variable is set for that checkout only.
Claude's wrapper scripts for `git add`, `git rm`, `git commit` and `git push` always work on the repository the action checked out, and take only a short list of flags, so `git add -f` is refused. They ignore settings from the environment: `GIT_SSH_COMMAND`, `GIT_ASKPASS`, `GIT_CONFIG_*`, any other `GIT_*` variable except the author and committer ones, `XDG_CONFIG_HOME`, and a `HOME` changed during the session. A workflow that relies on one of those, such as pushing over SSH with `GIT_SSH_COMMAND`, can allow its own command in `claude_args`, such as `Bash(git push:*)`, which gives up these protections for that workflow.
This has limits:
- It is not a sandbox. Code running on the runner, for example through an allowed `npm test`, can stop the action before it restores anything, or change the files later steps run, including the code of this and other actions. An allow rule for `Write` can reach those files too, so treat it like allowing code to run.
- A run that is cancelled or times out during the session does not restore anything, and later `always()` and post steps still run.
- Submodule config, a repository checked out to a subdirectory, and an `include.path` target that does not exist yet are not restored.
- The git settings need git 2.31 or newer, and `safe.bareRepository` needs 2.38. On older versions they have no effect.
### `claude-code-action` vs `claude-code-base-action`
`claude-code-base-action` is a lower-level building block that installs and runs Claude Code with the inputs you provide. It does not perform actor permission checks or restore project configuration from the base ref. If you need those behaviors, use this action (`claude-code-action`). See the [base-action README](../base-action/README.md#trust-model) for details.
`claude-code-base-action` is a lower-level building block that installs and runs Claude Code with the inputs you provide. It does not perform actor permission checks, refuse pull requests from forks, restrict commands on pull requests from outside contributors, or restore project configuration from the base ref. If you need those behaviors, use this action (`claude-code-action`). See the [base-action README](../base-action/README.md#trust-model) for details.
## Pull Request Creation
In its default configuration, **Claude does not create pull requests automatically** when responding to `@claude` mentions. Instead:
- Claude commits code changes to a new branch
- Claude commits code changes to a new branch, or to the pull request's branch, through wrapper scripts in the action. A deny rule for `git add`, `git rm`, `git commit` or `git push` in `claude_args` also covers those scripts, including a narrower rule such as `git add -A:*`. A deny rule in a settings file doesn't, so pass it in `claude_args`
- Claude's push command can't update any other branch. For a pull request from a fork, its branch is `pr-N` in your repository, which also holds the [fork's commits](#running-code-from-pull-requests), and the fork's branch is not updated
- Branch protection on your default branch is still worth having: a workflow can allow other push commands, and a pull request opened from the default branch has that branch as its own
- Claude provides a **link to the GitHub PR creation page** in its response
- **The user must click the link and create the PR themselves**, ensuring human oversight before any code is proposed for merging
@@ -69,6 +165,10 @@ This design ensures that users retain full control over what pull requests are c
On public repos, you can also use `include_comments_by_actor` to allowlist which users' comments are passed to Claude, reducing exposure to untrusted input. Use `exclude_comments_by_actor` to filter out noisy bot comments (e.g., `dependabot[bot]`, `renovate[bot]`). If an actor matches both lists, exclusion takes priority. See [Usage](./usage.md) for details.
### Auto mode is a review, not a boundary
By default the action uses [auto mode](./usage.md#permission-mode), where a safety review decides tool calls that your allow rules don't cover. It catches many unsafe actions, but a call can still run if the review passes it. For a strict list of what Claude may run, use `--permission-mode acceptEdits` or `default`.
## GitHub App Permissions
The [Claude Code GitHub app](https://github.com/apps/claude) requests the following permissions:
@@ -104,6 +204,8 @@ This uses GitHub's API to create commits, which automatically signs them as veri
This is the simplest option and requires no additional setup. However, because it uses the GitHub API instead of git CLI, it cannot perform complex git operations like rebasing, cherry-picking, or interactive history manipulation.
With `prompt`, a run on an issue commits to `issue-N`. A run on a pull request commits to the pull request's branch, or to `pr-N` if that branch isn't known to be in this repository. `issue-N` and `pr-N` start from the base branch. To use another branch, set `CLAUDE_BRANCH` in the step's `env`.
### Option 2: SSH Signing Key (ssh_signing_key)
This uses an SSH key to sign commits via git CLI. Use this option when you need both signed commits AND standard git operations (rebasing, cherry-picking, etc.):
+58 -10
View File
@@ -33,9 +33,9 @@ jobs:
# --model claude-4-0-sonnet-20250805
# Optional: add custom plugin marketplaces
# plugin_marketplaces: "https://github.com/user/marketplace1.git\nhttps://github.com/user/marketplace2.git"
# plugin_marketplaces: "https://github.com/anthropics/claude-plugins-official.git\nhttps://github.com/user/marketplace.git"
# Optional: install Claude Code plugins
# plugins: "code-review@claude-code-plugins\nfeature-dev@claude-code-plugins"
# plugins: "code-review@claude-plugins-official\nfeature-dev@claude-plugins-official"
# Optional: add custom trigger phrase (default: @claude)
# trigger_phrase: "/claude"
@@ -65,14 +65,14 @@ jobs:
| `track_progress` | Force tag mode with tracking comments. Only works with specific PR/issue events. Preserves GitHub context | No | `false` |
| `include_fix_links` | Include 'Fix this' links in PR code review feedback that open Claude Code with context to fix the identified issue | No | `true` |
| `claude_args` | Additional [arguments to pass directly to Claude CLI](https://docs.claude.com/en/docs/claude-code/cli-reference#cli-flags) (e.g., `--max-turns 10 --model claude-4-0-sonnet-20250805`) | No | "" |
| `base_branch` | The base branch to use for creating new branches (e.g., 'main', 'develop') | No | - |
| `base_branch` | The base branch to use for creating new branches (e.g., 'main', 'develop'). See also [Security](./security.md#which-files-come-from-the-base-branch-on-pull-requests) | No | - |
| `use_sticky_comment` | Use just one comment to deliver PR comments (only applies for pull_request event workflows) | No | `false` |
| `classify_inline_comments` | Buffer inline comments without `confirmed: true` and classify them (real review vs test/probe) via Haiku before posting after the session ends. Prevents subagent test comments. Set `'false'` to post all inline comments immediately | No | `true` |
| `github_token` | GitHub token for Claude to operate with. **Only include this if you're connecting a custom GitHub app of your own!** | No | - |
| `use_bedrock` | Use Amazon Bedrock with OIDC authentication instead of direct Anthropic API | No | `false` |
| `use_vertex` | Use Google Vertex AI with OIDC authentication instead of direct Anthropic API | No | `false` |
| `assignee_trigger` | The assignee username that triggers the action (e.g. @claude). Only used for issue assignment | No | - |
| `label_trigger` | The label name that triggers the action when applied to an issue (e.g. "claude") | No | - |
| `label_trigger` | The label name that triggers the action when applied to an issue (e.g. "claude"). With `prompt` and `track_progress`, an empty value accepts any label | No | - |
| `trigger_phrase` | The trigger phrase to look for in comments, issue/PR bodies, and issue titles | No | `@claude` |
| `branch_prefix` | The prefix to use for Claude branches (defaults to 'claude/', use 'claude-' for dash format) | No | `claude/` |
| `settings` | Claude Code settings as JSON string or path to settings JSON file | No | "" |
@@ -84,6 +84,8 @@ jobs:
| `include_comments_by_actor` | Comma-separated list of actor usernames to INCLUDE in comments. Supports the `*[bot]` wildcard to match all bot accounts. Empty (default) includes all actors | No | "" |
| `exclude_comments_by_actor` | Comma-separated list of actor usernames to EXCLUDE from comments. Supports the `*[bot]` wildcard to match all bot accounts. If an actor matches both lists, exclusion takes priority | No | "" |
| `allowed_bots` | Comma-separated list of allowed bot usernames, or '\*' to allow all bots. Empty string (default) allows no bots. **⚠️ On public repos with `'*'`, external Apps may be able to invoke this action.** See [Security](./security.md) | No | "" |
| `allow_fork_prs` | Whether to run on pull requests from forks, and on `workflow_run` events for a fork's run. With any value other than `true`, the action skips them. See [Security](./security.md#refusing-pull-requests-from-forks). | No | `true` |
| `restrict_untrusted_prs` | On pull requests whose author does not have write access, deny commands that commonly run code from the checkout, even if `claude_args` allows them. See [Security](./security.md#running-code-from-pull-requests). | No | `true` |
| `allowed_non_write_users` | **⚠️ RISKY**: Comma-separated list of usernames to allow without write permissions, or '\*' for all users. Only works with `github_token` input. See [Security](./security.md) | No | "" |
| `path_to_claude_code_executable` | Optional path to a custom Claude Code executable. Skips automatic installation. Useful for Nix, custom containers, or specialized environments | No | "" |
| `path_to_bun_executable` | Optional path to a custom Bun executable. Skips automatic Bun installation. Useful for Nix, custom containers, or specialized environments | No | "" |
@@ -99,7 +101,7 @@ These inputs are deprecated and will be removed in a future version:
| `mode` | **DEPRECATED**: Mode is now automatically detected based on workflow context | Remove this input; the action auto-detects the correct mode |
| `direct_prompt` | **DEPRECATED**: Use `prompt` instead | Replace with `prompt` |
| `override_prompt` | **DEPRECATED**: Use `prompt` with template variables or `claude_args` with `--system-prompt` | Use `prompt` for templates or `claude_args` for system prompts |
| `custom_instructions` | **DEPRECATED**: Use `claude_args` with `--system-prompt` or include in `prompt` | Move instructions to `prompt` or use `claude_args` |
| `custom_instructions` | **DEPRECATED**: Use `claude_args` with `--append-system-prompt` or include in `prompt` | Move instructions to `prompt` or use `claude_args` |
| `max_turns` | **DEPRECATED**: Use `claude_args` with `--max-turns` instead | Use `claude_args: "--max-turns 5"` |
| `model` | **DEPRECATED**: Use `claude_args` with `--model` instead | Use `claude_args: "--model claude-4-0-sonnet-20250805"` |
| `fallback_model` | **DEPRECATED**: Use `claude_args` with fallback configuration | Configure fallback in `claude_args` or `settings` |
@@ -112,6 +114,46 @@ These inputs are deprecated and will be removed in a future version:
> **Note**: This action is currently in beta. Features and APIs may change as we continue to improve the integration.
## Permission Mode
The action runs Claude in [auto mode](https://code.claude.com/docs/en/permission-modes) by default. Reads and file edits in the workspace run without asking, and a safety review decides any other tool call that your allow rules don't cover. Narrow allow rules like `Bash(npm test)` skip the review, and broad ones like `Bash(*)` or package-manager run commands are ignored in this mode.
### Opting out
Set another mode in `claude_args`. These are the earlier defaults:
```yaml
- uses: anthropics/claude-code-action@v1
with:
# For @claude mentions: allow file edits in the workspace, and deny what no rule covers
claude_args: --permission-mode acceptEdits
# For workflows with a prompt: deny what no rule covers
# claude_args: --permission-mode default
```
Opt out if your workflow relies on `--allowedTools` as a strict list, because auto mode can run other tool calls when the review passes.
### When auto mode doesn't apply
- **Unsupported model or organization setting:** auto mode needs a [supported model](https://code.claude.com/docs/en/permission-modes#eliminate-prompts-with-auto-mode), and an administrator can turn it off. The session then starts in Manual. For `@claude` mentions, the action logs a warning and switches to `acceptEdits`. With a `prompt`, it stays in Manual and denies what no rule covers.
- **`allowed_non_write_users` or `allowed_bots` is set:** sessions keep the earlier defaults, because other people's text reaches these runs without a write-access check, so the review can't pass a call those defaults deny. Add `--permission-mode auto` to opt in.
- **`ANTHROPIC_BASE_URL` is set:** sessions keep the earlier defaults, because the safety check depends on what a gateway passes through. Add `--permission-mode auto` once your gateway supports it.
- **`path_to_claude_code_executable` is set:** sessions keep the earlier defaults, because a Claude Code version that predates auto mode exits on the flag. Add `--permission-mode auto` if your version supports it.
- **The `settings` input sets `env.ANTHROPIC_BASE_URL`, sets a `permissions.defaultMode` other than `auto`, or is a file path:** sessions keep the earlier defaults, including for a file that sets neither, because a pull request can change the file. Workflows with a prompt then use the mode in your settings. For `@claude` mentions, `acceptEdits` still overrides it, so pass yours in `claude_args`.
- **`--dangerously-skip-permissions`:** this still takes priority over any mode.
### Provider notes
- Where the server's check doesn't reach a session, Claude Code makes its own [classifier requests, which count toward token usage](https://code.claude.com/docs/en/auto-mode-classifier-billing).
- On Amazon Bedrock, your account may need access to the classifier's model, or actions that need the review are blocked. See the [error reference](https://code.claude.com/docs/en/errors.md#auto-mode-cannot-determine-the-safety-of-an-action).
### Upgrading
- Tool calls that used to be denied can now run if the review passes.
- Allow rules that auto mode ignores no longer approve those commands on their own.
- The action checks only JSON in the `settings` input, not settings files such as `.claude/settings.json` in your repository. If one of those points `ANTHROPIC_BASE_URL` to a gateway or sets `permissions.defaultMode`, auto mode still applies, so set the mode you want in `claude_args`.
## Upgrading from v0.x?
For a comprehensive guide on migrating from v0.x to v1.0, including step-by-step instructions and examples, see our **[Migration Guide](./migration-guide.md)**.
@@ -139,7 +181,7 @@ For a comprehensive guide on migrating from v0.x to v1.0, including step-by-step
anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
claude_args: |
--max-turns 10
--system-prompt "Focus on security"
--append-system-prompt "Focus on security"
```
#### Automation Workflows
@@ -231,21 +273,25 @@ All structured output fields are available in the `structured_output` output as
**In GitHub Actions expressions:**
```yaml
if: fromJSON(steps.analyze.outputs.structured_output).is_flaky == true
run: |
CONFIDENCE=${{ fromJSON(steps.analyze.outputs.structured_output).confidence }}
if: |
fromJSON(steps.analyze.outputs.structured_output).is_flaky == true &&
fromJSON(steps.analyze.outputs.structured_output).confidence >= 0.7
run: ./retry-tests.sh
```
**In bash with jq:**
```yaml
- name: Process results
env:
OUTPUT: ${{ steps.analyze.outputs.structured_output }}
run: |
OUTPUT='${{ steps.analyze.outputs.structured_output }}'
IS_FLAKY=$(echo "$OUTPUT" | jq -r '.is_flaky')
SUMMARY=$(echo "$OUTPUT" | jq -r '.summary')
```
Always pass these values to a script through `env:`, as above. GitHub puts a `${{ }}` expression's value into the script before the shell reads it, so a quote or other shell syntax in Claude's answer can break the script or run as a command.
**Note**: Due to GitHub Actions limitations, composite actions cannot expose dynamic outputs. All fields are bundled in the single `structured_output` JSON string.
### Complete Example
@@ -305,3 +351,5 @@ Upload a screenshot of a bug and ask Claude to fix it:
```
Claude can see and analyze images, making it easy to fix visual bugs or UI issues.
If a description is edited after the event that triggered the run, Claude gets the text from that event without its images. A comment or review edited after that event is left out, or loses its images. Ticking a checkbox counts as an edit, and re-running the job does not help. If the images are still in the description, mentioning Claude in a new comment includes them.
+39
View File
@@ -0,0 +1,39 @@
# Require human approvals on PRs that contain agent-authored commits.
#
# Both triggers run the workflow file from the BASE/DEFAULT branch, so a PR
# cannot edit this check to approve itself. (`pull_request_review` is not
# used because it runs from the merge ref, not the default branch; native
# Approve reviews are picked up on the next synchronize or `/approve`
# comment.)
#
# After adding this workflow, mark `agent-approval-check` as a required
# status check on your protected branches.
name: agent-approval-check
on:
pull_request_target:
types: [opened, synchronize, reopened, ready_for_review]
issue_comment:
types: [created]
permissions:
contents: read
pull-requests: write
statuses: write
jobs:
check:
# issue_comment also fires on plain issues; skip those early.
if: github.event_name != 'issue_comment' || github.event.issue.pull_request
runs-on: ubuntu-latest
steps:
- uses: anthropics/claude-code-action/agent-approval-check@main
with:
required_approvals: 2
agent_emails: [email protected]
agent_logins: claude[bot],claude-code[bot]
# Uncomment to tune:
# excluded_approvers: dependabot[bot]
# exempt_path_prefixes: docs/
# protected_bases: main,release
+4 -3
View File
@@ -55,8 +55,8 @@ jobs:
GH_TOKEN: ${{ github.token }}
WORKFLOW_NAME: ${{ github.event.workflow_run.name }}
HEAD_BRANCH: ${{ github.event.workflow_run.head_branch }}
OUTPUT: ${{ steps.detect.outputs.structured_output }}
run: |
OUTPUT='${{ steps.detect.outputs.structured_output }}'
CONFIDENCE=$(echo "$OUTPUT" | jq -r '.confidence')
SUMMARY=$(echo "$OUTPUT" | jq -r '.summary')
@@ -72,8 +72,9 @@ jobs:
if: |
fromJSON(steps.detect.outputs.structured_output).is_flaky == true &&
fromJSON(steps.detect.outputs.structured_output).confidence < 0.7
env:
OUTPUT: ${{ steps.detect.outputs.structured_output }}
run: |
OUTPUT='${{ steps.detect.outputs.structured_output }}'
CONFIDENCE=$(echo "$OUTPUT" | jq -r '.confidence')
echo "⚠️ Possible flaky test but confidence too low ($CONFIDENCE)"
@@ -85,8 +86,8 @@ jobs:
env:
GH_TOKEN: ${{ github.token }}
HEAD_BRANCH: ${{ github.event.workflow_run.head_branch }}
OUTPUT: ${{ steps.detect.outputs.structured_output }}
run: |
OUTPUT='${{ steps.detect.outputs.structured_output }}'
IS_FLAKY=$(echo "$OUTPUT" | jq -r '.is_flaky')
CONFIDENCE=$(echo "$OUTPUT" | jq -r '.confidence')
SUMMARY=$(echo "$OUTPUT" | jq -r '.summary')
+3 -3
View File
@@ -12,17 +12,17 @@
"dependencies": {
"@actions/core": "^1.10.1",
"@actions/github": "^6.0.1",
"@anthropic-ai/claude-agent-sdk": "^0.3.161",
"@anthropic-ai/claude-agent-sdk": "^0.3.330",
"@modelcontextprotocol/sdk": "^1.11.0",
"@octokit/graphql": "^8.2.2",
"@octokit/rest": "^21.1.1",
"@octokit/webhooks-types": "^7.6.1",
"node-fetch": "^3.3.2",
"shell-quote": "^1.8.3",
"shell-quote": "^1.8.4",
"zod": "^3.24.4"
},
"devDependencies": {
"@types/bun": "1.2.11",
"@types/bun": "1.4.3",
"@types/node": "^20.0.0",
"@types/node-fetch": "^2.6.12",
"@types/shell-quote": "^1.7.5",
+27
View File
@@ -0,0 +1,27 @@
#!/usr/bin/env bash
set -euo pipefail
# Wrapper around `git add` that only stages files. Without it, an allow rule for
# `git add` also permits flags that open an editor or read a file of paths.
#
# Usage:
# git-add.sh [-A] [-u] [-v] [--chmod=+x|--chmod=-x] [--] <paths>
source "$(dirname "${BASH_SOURCE[0]}")/git-env.sh"
paths_only=false
for arg in "$@"; do
if [[ "$paths_only" == true ]]; then
continue
fi
case "$arg" in
-A | --all | -u | --update | -v | --verbose | --chmod=+x | --chmod=-x) ;;
--) paths_only=true ;;
-*)
echo "Error: only -A, -u, -v, --chmod and file paths are allowed (got: $arg)" >&2
exit 1
;;
esac
done
exec git "${git_args[@]}" add "$@"
+96
View File
@@ -0,0 +1,96 @@
#!/usr/bin/env bash
set -euo pipefail
# Wrapper around `git commit` that only allows a literal message. Without it, an
# allow rule for `git commit` also permits flags that make git read a file into
# the commit, such as -F .git/config, which a later push would then publish.
#
# Usage:
# git-commit.sh -m "<message>" [-m "<message>"] [-a] [--allow-empty] [-- <paths>]
source "$(dirname "${BASH_SOURCE[0]}")/git-env.sh"
has_message=false
expect_message=false
paths_only=false
has_path=false
all=false
args=()
for arg in "$@"; do
# The arguments without -a, for when the files it adds are staged separately below.
kept="$arg"
if [[ "$expect_message" == true ]]; then
expect_message=false
has_message=true
elif [[ "$paths_only" == true ]]; then
has_path=true
else
case "$arg" in
-m | --message)
expect_message=true
;;
-am)
expect_message=true
all=true
kept="-m"
;;
--message=*)
has_message=true
;;
-a | --all)
all=true
continue
;;
--allow-empty | -q | --quiet | -s | --signoff) ;;
--)
paths_only=true
;;
-*)
echo "Error: only -m with a literal message and basic commit flags are allowed (got: $arg)" >&2
exit 1
;;
*)
has_path=true
;;
esac
fi
args+=("$kept")
done
if [[ "$expect_message" == true ]]; then
echo "Error: -m needs a message" >&2
exit 1
fi
if [[ "$has_message" != true ]]; then
echo "Error: a message is required, with -m" >&2
exit 1
fi
# The action took these files from a branch other than the pull request's base, so a
# commit must not carry that branch's version of them. It stays in the working tree.
restored=()
if [[ -e "$script_dir/.restored-paths" ]]; then
while IFS= read -r pathspec; do
restored+=("$pathspec")
done <"$script_dir/.restored-paths"
fi
if [[ ${#restored[@]} -gt 0 ]]; then
if [[ "$has_path" == true ]]; then
for pathspec in "${restored[@]}"; do
args+=(":(exclude,${pathspec#:(}")
done
elif [[ "$all" == true ]]; then
git "${git_args[@]}" add -u :/
fi
if ! git "${git_args[@]}" diff --cached --quiet -- "${restored[@]}"; then
git "${git_args[@]}" reset -q -- "${restored[@]}"
echo "Note: these files come from another branch in this run, so their changes are not committed: ${restored[*]}" >&2
fi
if [[ "$has_path" != true || "$all" != true ]]; then
set -- "${args[@]}"
fi
fi
exec git "${git_args[@]}" commit "$@"
+25
View File
@@ -0,0 +1,25 @@
# Sourced by the git wrappers. Clears settings in the environment that change what
# git runs, and sets the arguments every wrapper passes to git. Files beside this
# script, written by the action, record the session's repository and hook setting.
script_dir="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
while IFS= read -r name; do
case "$name" in
GIT_AUTHOR_* | GIT_COMMITTER_*) ;;
GIT_*) unset "$name" ;;
esac
done < <(compgen -e)
unset XDG_CONFIG_HOME EDITOR VISUAL PAGER SSH_ASKPASS
git_args=(-c core.fsmonitor=false -c safe.bareRepository=explicit)
if [[ -e "$script_dir/.no-hooks" ]]; then
git_args+=(-c core.hooksPath=/dev/null)
fi
if [[ -e "$script_dir/.git-repo" ]]; then
{ IFS= read -r git_dir; IFS= read -r work_tree; IFS= read -r home || true; } <"$script_dir/.git-repo"
git_args+=("--git-dir=$git_dir" "--work-tree=$work_tree")
if [[ -n "$home" ]]; then
export HOME="$home"
fi
fi
+22 -5
View File
@@ -1,7 +1,10 @@
#!/usr/bin/env bash
set -euo pipefail
# Wrapper around `git push` that only allows `origin <ref>` with no flags.
# Wrapper around `git push` that only allows `origin <ref>` with no flags, and only
# updates the branch this session was set up on. That stops a prompt-injected session
# from updating main or anyone else's branch. Tag mode records the branch in a file
# beside this script. Elsewhere, set CLAUDE_PUSH_BRANCH.
# Defends against --receive-pack / --exec RCE and arbitrary-remote exfiltration
# (H1 #3556799). `git push:*` in allowedTools permits `git push --receive-pack='sh -c ...' ext::sh`
# which runs arbitrary shell on the Actions runner. This wrapper closes that.
@@ -27,10 +30,24 @@ if [[ "$1" != "origin" ]]; then
exit 1
fi
REF="$2"
if [[ "$REF" != "HEAD" ]] && ! git check-ref-format --branch "$REF" >/dev/null 2>&1; then
echo "Error: invalid ref: $REF" >&2
BRANCH="${CLAUDE_PUSH_BRANCH:-}"
source "$(dirname "${BASH_SOURCE[0]}")/git-env.sh"
if [[ -e "$script_dir/.push-branch" ]]; then
BRANCH="$(<"$script_dir/.push-branch")"
fi
if [[ -z "$BRANCH" ]]; then
echo "Error: CLAUDE_PUSH_BRANCH is not set, so there is no branch to push to" >&2
exit 1
fi
exec git push origin "$REF"
case "$2" in
HEAD) SOURCE="HEAD" ;;
"$BRANCH") SOURCE="refs/heads/$BRANCH" ;;
*)
echo "Error: this session can only push to $BRANCH (got: $2)" >&2
exit 1
;;
esac
# The full refspec fixes the target, whatever HEAD points at, and can't be forced.
exec git "${git_args[@]}" push origin "$SOURCE:refs/heads/$BRANCH"
+26
View File
@@ -0,0 +1,26 @@
#!/usr/bin/env bash
set -euo pipefail
# Wrapper around `git rm` that only takes basic flags and file paths.
#
# Usage:
# git-rm.sh [-r] [-f] [--cached] [-q] [--] <paths>
source "$(dirname "${BASH_SOURCE[0]}")/git-env.sh"
paths_only=false
for arg in "$@"; do
if [[ "$paths_only" == true ]]; then
continue
fi
case "$arg" in
-r | -f | -rf | -fr | --force | --cached | -q | --quiet) ;;
--) paths_only=true ;;
-*)
echo "Error: only -r, -f, --cached, -q and file paths are allowed (got: $arg)" >&2
exit 1
;;
esac
done
exec git "${git_args[@]}" rm "$@"
+90 -39
View File
@@ -1,8 +1,9 @@
#!/usr/bin/env bun
import * as core from "@actions/core";
import { writeFile, mkdir } from "fs/promises";
import { writeFile, mkdir, rm } from "fs/promises";
import type { FetchDataResult } from "../github/data/fetcher";
import { isSameLogin } from "../github/validation/actor";
import {
formatContext,
formatBody,
@@ -17,20 +18,23 @@ import {
isPullRequestReviewEvent,
isPullRequestReviewCommentEvent,
} from "../github/context";
import type { ParsedGitHubContext } from "../github/context";
import type { GitHubContext, ParsedGitHubContext } from "../github/context";
import type { CommonFields, PreparedContext, EventData } from "./types";
import { GITHUB_SERVER_URL } from "../github/api/config";
import { extractUserRequest } from "../utils/extract-user-request";
export type { CommonFields, PreparedContext } from "./types";
const GIT_ADD_WRAPPER = `${process.env.GITHUB_ACTION_PATH}/scripts/git-add.sh`;
const GIT_RM_WRAPPER = `${process.env.GITHUB_ACTION_PATH}/scripts/git-rm.sh`;
const GIT_PUSH_WRAPPER = `${process.env.GITHUB_ACTION_PATH}/scripts/git-push.sh`;
const GIT_COMMIT_WRAPPER = `${process.env.GITHUB_ACTION_PATH}/scripts/git-commit.sh`;
/** Filename for the user request file, read by the SDK runner */
const USER_REQUEST_FILENAME = "claude-user-request.txt";
// Tag mode defaults - these tools are needed for tag mode to function.
// Edit/MultiEdit/Write are intentionally omitted: acceptEdits permission mode
// auto-allows file edits inside $GITHUB_WORKSPACE and denies writes outside it.
// Edit/MultiEdit/Write are intentionally omitted: the permission mode allows file
// edits inside $GITHUB_WORKSPACE and doesn't allow every write outside it.
const BASE_ALLOWED_TOOLS = ["Glob", "Grep", "LS", "Read"];
export function buildAllowedToolsString(
@@ -53,10 +57,10 @@ export function buildAllowedToolsString(
} else {
// When not using commit signing, add specific Bash git commands
baseTools.push(
"Bash(git add:*)",
"Bash(git commit:*)",
`Bash(${GIT_ADD_WRAPPER}:*)`,
`Bash(${GIT_COMMIT_WRAPPER}:*)`,
`Bash(${GIT_PUSH_WRAPPER}:*)`,
"Bash(git rm:*)",
`Bash(${GIT_RM_WRAPPER}:*)`,
);
}
@@ -122,6 +126,7 @@ export function prepareContext(
// Extract trigger username and comment data based on event type
let triggerUsername: string | undefined;
let triggerUserId: number | undefined;
let commentId: string | undefined;
let commentBody: string | undefined;
@@ -129,15 +134,19 @@ export function prepareContext(
commentId = context.payload.comment.id.toString();
commentBody = context.payload.comment.body;
triggerUsername = context.payload.comment.user.login;
triggerUserId = context.payload.comment.user.id;
} else if (isPullRequestReviewEvent(context)) {
commentBody = context.payload.review.body ?? "";
triggerUsername = context.payload.review.user.login;
triggerUserId = context.payload.review.user.id;
} else if (isPullRequestReviewCommentEvent(context)) {
commentId = context.payload.comment.id.toString();
commentBody = context.payload.comment.body;
triggerUsername = context.payload.comment.user.login;
triggerUserId = context.payload.comment.user.id;
} else if (isIssuesEvent(context)) {
triggerUsername = context.payload.issue.user.login;
triggerUserId = context.payload.issue.user.id;
}
// Create infrastructure fields object
@@ -146,6 +155,7 @@ export function prepareContext(
claudeCommentId,
triggerPhrase,
...(triggerUsername && { triggerUsername }),
...(triggerUserId && { triggerUserId }),
...(prompt && { prompt }),
...(claudeBranch && { claudeBranch }),
};
@@ -276,7 +286,7 @@ export function prepareContext(
...(assigneeTrigger && { assigneeTrigger }),
};
} else if (eventAction === "labeled") {
if (!labelTrigger) {
if (!labelTrigger && !prompt) {
throw new Error("LABEL_TRIGGER is required for issue labeled event");
}
eventData = {
@@ -286,7 +296,7 @@ export function prepareContext(
issueNumber,
baseBranch,
claudeBranch,
labelTrigger,
...(labelTrigger && { labelTrigger }),
};
} else if (eventAction === "opened") {
eventData = {
@@ -364,7 +374,9 @@ export function getEventTypeAndContext(envVars: PreparedContext): {
} else if (eventData.eventAction === "labeled") {
return {
eventType: "ISSUE_LABELED",
triggerContext: `issue labeled with '${eventData.labelTrigger}'`,
triggerContext: eventData.labelTrigger
? `issue labeled with '${eventData.labelTrigger}'`
: "issue labeled",
};
}
return {
@@ -394,9 +406,16 @@ function getCommitInstructions(
context: PreparedContext,
useCommitSigning: boolean,
): string {
const triggerName = githubData.triggerDisplayName ?? context.triggerUsername;
const triggerEmail =
context.triggerUserId && context.triggerUsername
? `${context.triggerUserId}+${context.triggerUsername}@users.noreply.github.com`
: context.triggerUsername
? `${context.triggerUsername}@users.noreply.github.com`
: undefined;
const coAuthorLine =
(githubData.triggerDisplayName ?? context.triggerUsername) !== "Unknown"
? `Co-authored-by: ${githubData.triggerDisplayName ?? context.triggerUsername} <${context.triggerUsername}@users.noreply.github.com>`
triggerName && triggerName !== "Unknown" && triggerEmail
? `Co-authored-by: ${triggerName} <${triggerEmail}>`
: "";
if (useCommitSigning) {
@@ -419,12 +438,12 @@ function getCommitInstructions(
if (eventData.isPR && !eventData.claudeBranch) {
return `
- Use git commands via the Bash tool to commit and push your changes:
- Stage files: Bash(git add <files>)
- Commit with a descriptive message: Bash(git commit -m "<message>")
- Stage files: Bash(${GIT_ADD_WRAPPER} <files>)
- Commit with a descriptive message: Bash(${GIT_COMMIT_WRAPPER} -m "<message>")
${
coAuthorLine
? `- When committing and the trigger user is not "Unknown", include a Co-authored-by trailer:
Bash(git commit -m "<message>\\n\\n${coAuthorLine}")`
Bash(${GIT_COMMIT_WRAPPER} -m "<message>\\n\\n${coAuthorLine}")`
: ""
}
- Push to the remote: Bash(${GIT_PUSH_WRAPPER} origin HEAD)`;
@@ -433,12 +452,12 @@ function getCommitInstructions(
return `
- You are already on the correct branch (${eventData.claudeBranch || "the PR branch"}). Do not create a new branch.
- Use git commands via the Bash tool to commit and push your changes:
- Stage files: Bash(git add <files>)
- Commit with a descriptive message: Bash(git commit -m "<message>")
- Stage files: Bash(${GIT_ADD_WRAPPER} <files>)
- Commit with a descriptive message: Bash(${GIT_COMMIT_WRAPPER} -m "<message>")
${
coAuthorLine
? `- When committing and the trigger user is not "Unknown", include a Co-authored-by trailer:
Bash(git commit -m "<message>\\n\\n${coAuthorLine}")`
Bash(${GIT_COMMIT_WRAPPER} -m "<message>\\n\\n${coAuthorLine}")`
: ""
}
- Push to the remote: Bash(${GIT_PUSH_WRAPPER} origin ${branchName})`;
@@ -811,12 +830,12 @@ ${
? `- Use mcp__github_file_ops__commit_files for making commits (works for both new and existing files, single or multiple). Use mcp__github_file_ops__delete_files for deleting files (supports deleting single or multiple files atomically), or mcp__github__delete_file for deleting a single file. Edit files locally, and the tool will read the content from the same path on disk.
Tool usage examples:
- mcp__github_file_ops__commit_files: {"files": ["path/to/file1.js", "path/to/file2.py"], "message": "feat: add new feature"}
- mcp__github_file_ops__delete_files: {"files": ["path/to/old.js"], "message": "chore: remove deprecated file"}`
- mcp__github_file_ops__delete_files: {"paths": ["path/to/old.js"], "message": "chore: remove deprecated file"}`
: `- Use git commands via the Bash tool for version control (remember that you have access to these git commands):
- Stage files: Bash(git add <files>)
- Commit changes: Bash(git commit -m "<message>")
- Push to remote: Bash(${GIT_PUSH_WRAPPER} origin <branch>)
- Delete files: Bash(git rm <files>) followed by commit and push
- Stage files: Bash(${GIT_ADD_WRAPPER} <files>)
- Commit changes: Bash(${GIT_COMMIT_WRAPPER} -m "<message>")
- Push to remote: Bash(${GIT_PUSH_WRAPPER} origin HEAD)
- Delete files: Bash(${GIT_RM_WRAPPER} <files>) followed by commit and push
- Check status: Bash(git status)
- View diff: Bash(git diff)${eventData.isPR && eventData.baseBranch ? `\n - IMPORTANT: For PR diffs, use: Bash(git diff origin/${eventData.baseBranch}...HEAD)` : ""}`
}
@@ -843,7 +862,7 @@ What You CANNOT Do:
- Submit formal GitHub PR reviews
- Approve pull requests (for security reasons)
- Post multiple comments (you only update your initial comment)
- Execute commands outside the repository context${useCommitSigning ? "\n- Run arbitrary Bash commands (unless explicitly allowed via allowed_tools configuration)" : ""}
- Execute commands outside the repository context${useCommitSigning ? "\n- Run arbitrary Bash commands (unless explicitly allowed via claude_args with --allowedTools)" : ""}
- Perform branch operations (cannot merge branches, rebase, or perform other git operations beyond creating and pushing commits)
- Modify files in the .github/workflows directory (GitHub App permissions do not allow workflow modifications)
@@ -864,6 +883,22 @@ f. If you are unable to complete certain steps, such as running a linter or test
return promptContent;
}
function getCommentOrReviewAuthor(
context: GitHubContext | undefined,
): string | undefined {
if (!context) return undefined;
if (
isIssueCommentEvent(context) ||
isPullRequestReviewCommentEvent(context)
) {
return context.payload.comment.user?.login;
}
if (isPullRequestReviewEvent(context)) {
return context.payload.review.user?.login;
}
return undefined;
}
/**
* Extracts the user's request from the prepared context and GitHub data.
*
@@ -880,12 +915,17 @@ f. If you are unable to complete certain steps, such as running a linter or test
* // Issue body with "@claude fix this" -> returns "fix this"
* // Issue assigned without @claude in body -> returns null
*/
function extractUserRequestFromContext(
export function extractUserRequestFromContext(
context: PreparedContext,
githubData: FetchDataResult,
): string | null {
const { eventData, triggerPhrase } = context;
// The request gets the highest trust in the prompt and can run slash commands,
// so it only comes from the person who passed the permission check. When
// someone else triggers the run, for example by labeling another person's
// issue or editing their comment, that text stays ordinary context.
// For comment events, extract from comment body
if (
"commentBody" in eventData &&
@@ -894,11 +934,23 @@ function extractUserRequestFromContext(
eventData.eventName === "pull_request_review_comment" ||
eventData.eventName === "pull_request_review")
) {
return extractUserRequest(eventData.commentBody, triggerPhrase);
return isSameLogin(
getCommentOrReviewAuthor(context.githubContext),
context.githubContext?.actor,
)
? extractUserRequest(eventData.commentBody, triggerPhrase)
: null;
}
// For issue/PR events triggered by body content, extract from the body
if (githubData.contextData?.body) {
// For issue/PR events triggered by body content, extract from the body.
// GraphQL gives a bot's login without the "[bot]" suffix that actors have.
const author = githubData.contextData?.author;
const authorLogin =
author?.__typename === "Bot" ? `${author.login}[bot]` : author?.login;
if (
githubData.contextData?.body &&
isSameLogin(authorLogin, context.githubContext?.actor)
) {
const request = extractUserRequest(
githubData.contextData.body,
triggerPhrase,
@@ -930,9 +982,14 @@ export async function createPrompt(
claudeBranch,
);
await mkdir(`${process.env.RUNNER_TEMP || "/tmp"}/claude-prompts`, {
recursive: true,
});
// Clear any stale prompt files from a prior invocation. RUNNER_TEMP is documented
// to be emptied between jobs, but on non-ephemeral self-hosted runners this is
// not reliably honored — a stale claude-user-request.txt left behind by a prior
// mention-mode invocation would not be overwritten by a subsequent agent-mode
// invocation, and would leak into the model's context.
const promptDir = `${process.env.RUNNER_TEMP || "/tmp"}/claude-prompts`;
await rm(promptDir, { recursive: true, force: true });
await mkdir(promptDir, { recursive: true });
// Generate the prompt directly
const promptContent = generatePrompt(
@@ -948,10 +1005,7 @@ export async function createPrompt(
console.log("=======================");
// Write the prompt file
await writeFile(
`${process.env.RUNNER_TEMP || "/tmp"}/claude-prompts/claude-prompt.txt`,
promptContent,
);
await writeFile(`${promptDir}/claude-prompt.txt`, promptContent);
// Extract and write the user request separately for SDK multi-block messaging
// This allows the CLI to process slash commands (e.g., "@claude /review-pr")
@@ -960,10 +1014,7 @@ export async function createPrompt(
githubData,
);
if (userRequest) {
await writeFile(
`${process.env.RUNNER_TEMP || "/tmp"}/claude-prompts/${USER_REQUEST_FILENAME}`,
userRequest,
);
await writeFile(`${promptDir}/${USER_REQUEST_FILENAME}`, userRequest);
console.log("===== USER REQUEST =====");
console.log(userRequest);
console.log("========================");
+2 -1
View File
@@ -5,6 +5,7 @@ export type CommonFields = {
claudeCommentId: string;
triggerPhrase: string;
triggerUsername?: string;
triggerUserId?: number;
prompt?: string;
claudeBranch?: string;
};
@@ -75,7 +76,7 @@ type IssueLabeledEvent = {
issueNumber: string;
baseBranch: string;
claudeBranch: string;
labelTrigger: string;
labelTrigger?: string;
};
type PullRequestBaseEvent = {
+2
View File
@@ -6,6 +6,8 @@ export function collectActionInputsPresence(): string {
base_branch: "",
branch_prefix: "claude/",
allowed_bots: "",
allow_fork_prs: "true",
restrict_untrusted_prs: "true",
mode: "tag",
model: "",
anthropic_model: "",
Executable → Regular
+20 -11
View File
@@ -2,6 +2,7 @@
import { readFileSync, existsSync } from "fs";
import { exit } from "process";
import { redactSecrets } from "../github/utils/sanitizer";
export type ToolUse = {
type: string;
@@ -163,8 +164,15 @@ export function formatResultContent(content: any): string {
typeof parsedContent[0] === "object" &&
parsedContent[0]?.type === "text"
) {
// Extract the text field from the first item
contentStr = parsedContent[0]?.text || "";
// Keep every text block, not just the first: a tool result may split its
// output across several, and dropping the rest silently loses findings,
// file paths and follow-up instructions from the rendered summary. Blocks
// of other types (for example images) are skipped. Tool output is
// arbitrary, so `text` is not guaranteed to be a string.
contentStr = parsedContent
.filter((block: any) => block?.type === "text")
.map((block: any) => String(block?.text || ""))
.join("\n");
} else {
contentStr = String(content).trim();
}
@@ -172,6 +180,10 @@ export function formatResultContent(content: any): string {
contentStr = String(content).trim();
}
// Redact before truncating so a credential cannot be split at the cut and
// slip past the final redaction pass.
contentStr = redactSecrets(contentStr);
// Truncate very long results
if (contentStr.length > 3000) {
contentStr = contentStr.substring(0, 2997) + "...";
@@ -268,7 +280,8 @@ export function groupTurnsNaturally(data: Turn[]): GroupedContent[] {
type: "system_init",
tools_count: tools.length,
});
} else {
} else if (subtype !== "thinking_tokens") {
// Skip thinking_tokens - internal progress events not meant for summary
groupedContent.push({
type: "system_other",
data: turn,
@@ -419,7 +432,9 @@ export function formatTurnsFromData(data: Turn[]): string {
// Generate markdown
const markdown = formatGroupedContent(groupedContent);
return markdown;
// Runtime output may contain credentials that are not registered as
// workflow secrets, so redact known formats before this gets published.
return redactSecrets(markdown);
}
function main(): void {
@@ -446,14 +461,8 @@ function main(): void {
const fileContent = readFileSync(jsonFile, "utf-8");
const data: Turn[] = JSON.parse(fileContent);
// Group turns naturally
const groupedContent = groupTurnsNaturally(data);
// Generate markdown
const markdown = formatGroupedContent(groupedContent);
// Print to stdout (so it can be captured by shell)
console.log(markdown);
console.log(formatTurnsFromData(data));
} catch (error) {
console.error(`Error processing file: ${error}`);
exit(1);
@@ -11,6 +11,7 @@
*/
import { readFileSync } from "fs";
import { createOctokit } from "../github/api/client";
import { redactSecrets } from "../github/utils/sanitizer";
const BUFFER_PATH = "/tmp/inline-comments-buffer.jsonl";
@@ -120,7 +121,7 @@ async function postComment(
owner,
repo,
pull_number,
body: c.body,
body: redactSecrets(c.body),
path: c.path,
side: c.side || "RIGHT",
commit_id: c.commit_id || headSha,
-99
View File
@@ -1,99 +0,0 @@
#!/usr/bin/env bun
/**
* Prepare the Claude action by checking trigger conditions, verifying human actor,
* and creating the initial tracking comment
*/
import * as core from "@actions/core";
import { setupGitHubToken } from "../github/token";
import { checkWritePermissions } from "../github/validation/permissions";
import { createOctokit } from "../github/api/client";
import { parseGitHubContext, isEntityContext } from "../github/context";
import { detectMode } from "../modes/detector";
import { prepareTagMode } from "../modes/tag";
import { prepareAgentMode } from "../modes/agent";
import { checkContainsTrigger } from "../github/validation/trigger";
import { collectActionInputsPresence } from "./collect-inputs";
async function run() {
try {
collectActionInputsPresence();
// Parse GitHub context first to enable mode detection
const context = parseGitHubContext();
// Auto-detect mode based on context
const modeName = detectMode(context);
console.log(
`Auto-detected mode: ${modeName} for event: ${context.eventName}`,
);
// Setup GitHub token
const githubToken = await setupGitHubToken();
const octokit = createOctokit(githubToken);
// Step 3: Check write permissions (only for entity contexts)
if (isEntityContext(context)) {
// Check if github_token was provided as input (not from app)
const githubTokenProvided = !!process.env.OVERRIDE_GITHUB_TOKEN;
const hasWritePermissions = await checkWritePermissions(
octokit.rest,
context,
context.inputs.allowedNonWriteUsers,
githubTokenProvided,
);
if (!hasWritePermissions) {
throw new Error(
"Actor does not have write permissions to the repository",
);
}
}
// Check trigger conditions
const containsTrigger =
modeName === "tag"
? isEntityContext(context) && checkContainsTrigger(context)
: !!context.inputs?.prompt;
// Debug logging
console.log(`Mode: ${modeName}`);
console.log(`Context prompt: ${context.inputs?.prompt || "NO PROMPT"}`);
console.log(`Trigger result: ${containsTrigger}`);
// Set output for action.yml to check
core.setOutput("contains_trigger", containsTrigger.toString());
if (!containsTrigger) {
console.log("No trigger found, skipping remaining steps");
// Still set github_token output even when skipping
core.setOutput("github_token", githubToken);
return;
}
// Run prepare
console.log(
`Preparing with mode: ${modeName} for event: ${context.eventName}`,
);
if (modeName === "tag") {
await prepareTagMode({ context, octokit, githubToken });
} else {
await prepareAgentMode({ context, octokit, githubToken });
}
// MCP config is handled by individual modes (tag/agent) and included in their claude_args output
// Expose the GitHub token (Claude App token) as an output
core.setOutput("github_token", githubToken);
} catch (error) {
const errorMessage = error instanceof Error ? error.message : String(error);
core.setFailed(`Prepare step failed with error: ${errorMessage}`);
// Also output the clean error message for the action to capture
core.setOutput("prepare_error", errorMessage);
process.exit(1);
}
}
if (import.meta.main) {
run();
}
+194 -24
View File
@@ -12,7 +12,15 @@ import { spawn } from "child_process";
import { appendFile } from "fs/promises";
import { existsSync, readFileSync } from "fs";
import { setupGitHubToken, WorkflowValidationSkipError } from "../github/token";
import { checkWritePermissions } from "../github/validation/permissions";
import {
checkBotActorPullRequestAuthor,
checkWritePermissions,
} from "../github/validation/permissions";
import {
isCommentOnRefusedForkPr,
isRefusedForkEvent,
} from "../github/validation/fork-pr";
import { getUntrustedPrDisallowedTools } from "../github/validation/untrusted-pr";
import { createOctokit } from "../github/api/client";
import type { Octokits } from "../github/api/client";
import {
@@ -21,18 +29,44 @@ import {
isPullRequestEvent,
isPullRequestReviewEvent,
isPullRequestReviewCommentEvent,
isWorkflowRunEvent,
} from "../github/context";
import type { GitHubContext } from "../github/context";
import { detectMode } from "../modes/detector";
import { prepareTagMode } from "../modes/tag";
import { prepareAgentMode } from "../modes/agent";
import { checkContainsTrigger } from "../github/validation/trigger";
import { restoreConfigFromBase } from "../github/operations/restore-config";
import {
checkContainsTrigger,
isModeration,
} from "../github/validation/trigger";
import {
clearSessionState,
disableGitPrograms,
pinRestoredPaths,
prepareGitForSession,
snapshotSessionFiles,
} from "../github/operations/session-guard";
import { chooseConfigBranch } from "../github/operations/config-branch";
import {
fetchBranch,
keepHeadHistory,
restoreConfigFromBase,
} from "../github/operations/restore-config";
import {
restoreToolConfigFromBase,
toolConfigDenyRules,
} from "../github/operations/tool-config";
import {
namedConfigPaths,
restoreConfigPathsFromBase,
} from "../github/operations/config-paths";
import { validateBranchName } from "../github/operations/branch";
import type { BranchInfo } from "../github/operations/branch";
import { collectActionInputsPresence } from "./collect-inputs";
import { updateCommentLink } from "./update-comment-link";
import { formatTurnsFromData } from "./format-turns";
import type { Turn } from "./format-turns";
import { redactSecrets } from "../github/utils/sanitizer";
// Base-action imports (used directly instead of subprocess)
import { setupWorkloadIdentity } from "../../base-action/src/workload-identity";
import type { WorkloadIdentityHandle } from "../../base-action/src/workload-identity";
@@ -44,6 +78,13 @@ import { runClaude } from "../../base-action/src/run-claude";
import type { ClaudeRunResult } from "../../base-action/src/run-claude-sdk";
import { setExecutionFileOutputIfPresent } from "../../base-action/src/execution-file";
// Exported for unit testing. `set -o pipefail` makes curl's non-zero exit
// propagate through the pipe so the install retry logic actually triggers
// on 429/403 instead of silently succeeding (see #1136).
export function buildInstallCommand(version: string): string {
return `set -o pipefail; curl -fsSL https://claude.ai/install.sh | bash -s -- ${version}`;
}
/**
* Install Claude Code CLI, handling retry logic and custom executable paths.
* Returns the absolute path to the claude executable.
@@ -68,7 +109,7 @@ async function installClaudeCode(): Promise<string> {
return customExecutable;
}
const claudeCodeVersion = "2.1.161";
const claudeCodeVersion = "2.1.330";
console.log(`Installing Claude Code v${claudeCodeVersion}...`);
for (let attempt = 1; attempt <= 3; attempt++) {
@@ -77,10 +118,7 @@ async function installClaudeCode(): Promise<string> {
await new Promise<void>((resolve, reject) => {
const child = spawn(
"bash",
[
"-c",
`curl -fsSL https://claude.ai/install.sh | bash -s -- ${claudeCodeVersion}`,
],
["-c", buildInstallCommand(claudeCodeVersion)],
{ stdio: "inherit" },
);
child.on("close", (code) => {
@@ -132,7 +170,7 @@ async function writeStepSummary(executionFile: string): Promise<void> {
fallback +=
"Failed to format output (please report). Here's the raw JSON:\n\n";
fallback += "```json\n";
fallback += readFileSync(executionFile, "utf-8");
fallback += redactSecrets(readFileSync(executionFile, "utf-8"));
fallback += "\n```\n";
await appendFile(summaryFile, fallback);
} catch {
@@ -141,7 +179,13 @@ async function writeStepSummary(executionFile: string): Promise<void> {
}
}
const FORK_REFUSED =
"Skipped, because allow_fork_prs is not true and this pull request or workflow run is from a fork.";
async function run() {
// Bun has read this by now, and it's only for the action's own scripts
delete process.env.BUN_RUNTIME_TRANSPILER_CACHE_PATH;
let githubToken: string | undefined;
let commentId: number | undefined;
let claudeBranch: string | undefined;
@@ -153,12 +197,22 @@ async function run() {
let context: GitHubContext | undefined;
let octokit: Octokits | undefined;
let workloadIdentity: WorkloadIdentityHandle | undefined;
// Paths reverted to the PR base branch, which cleanup must not commit back
// onto the PR author's branch. Empty unless restoreConfigFromBase ran.
let restoredConfigPaths: string[] = [];
// Track whether we've completed prepare phase, so we can attribute errors correctly
let prepareCompleted = false;
try {
// Phase 1: Prepare
const actionInputsPresent = collectActionInputsPresence();
context = parseGitHubContext();
// Before a token, which GitHub doesn't give to some of these events
if (isRefusedForkEvent(context)) {
core.warning(FORK_REFUSED);
return;
}
const modeName = detectMode(context);
console.log(
`Auto-detected mode: ${modeName} for event: ${context.eventName}`,
@@ -181,8 +235,10 @@ async function run() {
process.env.GITHUB_TOKEN = githubToken;
process.env.GH_TOKEN = githubToken;
// Check write permissions (only for entity contexts)
if (isEntityContext(context)) {
// Check write permissions for entity contexts, and for workflow_run
// events, whose upstream run may have been started by an actor without
// write access (e.g. the author of a fork pull request)
if (isEntityContext(context) || isWorkflowRunEvent(context)) {
const hasWritePermissions = await checkWritePermissions(
octokit.rest,
context,
@@ -200,7 +256,8 @@ async function run() {
const containsTrigger =
modeName === "tag"
? isEntityContext(context) && checkContainsTrigger(context)
: !!context.inputs?.prompt;
: !!context.inputs?.prompt &&
!(isEntityContext(context) && isModeration(context));
console.log(`Mode: ${modeName}`);
console.log(`Context prompt: ${context.inputs?.prompt || "NO PROMPT"}`);
console.log(`Trigger result: ${containsTrigger}`);
@@ -211,13 +268,46 @@ async function run() {
return;
}
// Before the action checks anything out
if (await isCommentOnRefusedForkPr(octokit.rest, context)) {
core.warning(FORK_REFUSED);
return;
}
const botActorAllowed = await checkBotActorPullRequestAuthor(
octokit.rest,
context,
context.inputs.allowedNonWriteUsers,
!!process.env.OVERRIDE_GITHUB_TOKEN,
);
if (!botActorAllowed) {
throw new Error(
"A bot acted on this pull request, but its author does not have write access",
);
}
// Run prepare
// Commands that run checkout code are denied on PRs from authors without write
// access. Hooks go off here, before the action checks out such a PR's head, and
// keep running for trusted authors, whose repositories may rely on them.
const untrustedPrDisallowedTools = isEntityContext(context)
? await getUntrustedPrDisallowedTools(octokit.rest, context)
: [];
const hooksOff = untrustedPrDisallowedTools.length > 0;
disableGitPrograms({ hooks: hooksOff });
clearSessionState();
console.log(
`Preparing with mode: ${modeName} for event: ${context.eventName}`,
);
const prepareResult =
modeName === "tag"
? await prepareTagMode({ context, octokit, githubToken })
? await prepareTagMode({
context,
octokit,
githubToken,
skipLfsDownloads: hooksOff,
})
: await prepareAgentMode({ context, octokit, githubToken });
commentId = prepareResult.commentId;
@@ -240,8 +330,16 @@ async function run() {
validateEnvironmentVariables();
// Tag mode sets claudeBranch only when it branched from the base instead of
// checking out the PR's head, as for a closed or merged PR.
const checkedOutPrHead = modeName !== "tag" || !claudeBranch;
const isUntrustedCheckout =
checkedOutPrHead && untrustedPrDisallowedTools.length > 0;
let restoredToolConfig: string[] = [];
let configPathDenyRules: string[] = [];
// On PRs, .claude/ and .mcp.json in the checkout are attacker-controlled.
// Restore them from the base branch before the CLI reads them.
// Restore them from a trusted branch before the CLI reads them.
//
// We read pull_request.base.ref from the payload directly because agent
// mode's branchInfo.baseBranch defaults to the repo's default branch rather
@@ -260,7 +358,43 @@ async function run() {
validateBranchName(restoreBase);
}
if (restoreBase) {
restoreConfigFromBase(restoreBase);
const { isFork }: BranchInfo = prepareResult.branchInfo;
const configBranch = chooseConfigBranch(
context,
restoreBase,
untrustedPrDisallowedTools.length > 0,
isFork,
);
const isOtherBranch = configBranch !== restoreBase;
if (isOtherBranch) {
core.warning(
`Claude's config comes from ${configBranch}, not from the base branch ${restoreBase}, because the pull request's author is not confirmed to have write access. See docs/security.md in anthropics/claude-code-action.`,
);
}
const restoreHeadHistory = isOtherBranch
? keepHeadHistory()
: undefined;
restoredConfigPaths = restoreConfigFromBase(configBranch);
// Before plugins are installed, which can already start package managers here.
if (isUntrustedCheckout) {
restoredToolConfig = restoreToolConfigFromBase(configBranch);
const configPaths = restoreConfigPathsFromBase(
configBranch,
namedConfigPaths(prepareResult.claudeArgs),
);
configPathDenyRules = configPaths.denyRules;
restoredConfigPaths.push(
...restoredToolConfig,
...configPaths.replaced,
);
}
if (isOtherBranch) {
// A branch from the base has none of the pull request's version of them.
if (checkedOutPrHead) pinRestoredPaths(restoredConfigPaths);
// Claude still compares the pull request with its base.
restoreHeadHistory?.();
fetchBranch(restoreBase);
}
}
}
@@ -280,13 +414,46 @@ async function run() {
promptFile,
});
const claudeResult: ClaudeRunResult = await runClaude(promptConfig.path, {
claudeArgs: prepareResult.claudeArgs,
appendSystemPrompt: process.env.APPEND_SYSTEM_PROMPT,
model: process.env.ANTHROPIC_MODEL,
pathToClaudeCodeExecutable: claudeExecutable,
showFullOutput: process.env.INPUT_SHOW_FULL_OUTPUT,
});
// Passed separately because text added to claudeArgs can be lost: a comment at
// the end of the user's claude_args would swallow it.
const disallowedTools = isUntrustedCheckout
? [
...untrustedPrDisallowedTools,
...toolConfigDenyRules(restoredToolConfig),
...configPathDenyRules,
]
: [];
// A hook runs in the session's current directory, and only the repository root's tool
// config comes from the base branch. This keeps Bash from moving the session.
if (isUntrustedCheckout) {
process.env.CLAUDE_BASH_MAINTAIN_PROJECT_WORKING_DIR = "1";
}
const restoreSessionFiles = snapshotSessionFiles(() =>
prepareGitForSession({ hooks: hooksOff }),
);
let claudeResult: ClaudeRunResult;
try {
claudeResult = await runClaude(promptConfig.path, {
claudeArgs: prepareResult.claudeArgs,
disallowedTools: disallowedTools.join(","),
appendSystemPrompt: process.env.APPEND_SYSTEM_PROMPT,
model: process.env.ANTHROPIC_MODEL,
pathToClaudeCodeExecutable: claudeExecutable,
showFullOutput: process.env.INPUT_SHOW_FULL_OUTPUT,
// Tag mode has no allow rules for file edits, so it needs a mode that allows them.
...(modeName === "tag" && { fallbackPermissionMode: "acceptEdits" }),
});
} finally {
restoreSessionFiles();
// The action's own git commands from here on hold the job's credentials,
// and the session could have edited a tracked hook.
disableGitPrograms({ hooks: true });
// Revoking the app token depends on this output, and a session can write it.
core.setOutput("skipped_due_to_workflow_validation_mismatch", "false");
}
claudeSuccess = claudeResult.conclusion === "success";
executionFile = claudeResult.executionFile;
@@ -310,11 +477,13 @@ async function run() {
prepareSuccess = false;
prepareError = errorMessage;
}
core.setFailed(`Action failed with error: ${errorMessage}`);
core.setFailed(`Action failed with error: ${redactSecrets(errorMessage)}`);
} finally {
// Phase 4: Cleanup (always runs)
clearSessionState();
// Stop refreshing the workload identity token file
// Stop refreshing the workload identity token file and delete the token
// material so it doesn't outlive this step
workloadIdentity?.stop();
// Update tracking comment
@@ -339,6 +508,7 @@ async function run() {
prepareSuccess,
prepareError,
useCommitSigning: context.inputs.useCommitSigning,
restoredConfigPaths,
});
} catch (error) {
console.error("Error updating comment with job link:", error);
+10 -1
View File
@@ -16,6 +16,7 @@ import type { ParsedGitHubContext } from "../github/context";
import { GITHUB_SERVER_URL } from "../github/api/config";
import { checkAndCommitOrDeleteBranch } from "../github/operations/branch-cleanup";
import { updateClaudeComment } from "../github/operations/comments/update-claude-comment";
import { encodeBranchNameForUrl } from "../github/operations/comments/common";
export type UpdateCommentLinkParams = {
commentId: number;
@@ -30,6 +31,12 @@ export type UpdateCommentLinkParams = {
prepareSuccess: boolean;
prepareError?: string;
useCommitSigning: boolean;
/**
* Paths restored from the PR base branch by restoreConfigFromBase. The
* auto-commit in checkAndCommitOrDeleteBranch must leave these alone, or it
* commits the revert onto the PR author's branch.
*/
restoredConfigPaths?: string[];
};
export async function updateCommentLink(
@@ -43,6 +50,7 @@ export async function updateCommentLink(
context,
octokit,
useCommitSigning,
restoredConfigPaths = [],
} = params;
const { owner, repo } = context.repository;
@@ -116,6 +124,7 @@ export async function updateCommentLink(
claudeBranch,
baseBranch,
useCommitSigning,
restoredConfigPaths,
);
// Check if we need to add PR URL when we have a new branch
@@ -151,7 +160,7 @@ export async function updateCommentLink(
const prBody = encodeURIComponent(
`This PR addresses ${entityType.toLowerCase()} #${context.entityNumber}\n\nGenerated with [Claude Code](https://claude.ai/code)`,
);
const prUrl = `${serverUrl}/${owner}/${repo}/compare/${baseBranch}...${claudeBranch}?quick_pull=1&title=${prTitle}&body=${prBody}`;
const prUrl = `${serverUrl}/${owner}/${repo}/compare/${encodeBranchNameForUrl(baseBranch)}...${encodeBranchNameForUrl(claudeBranch)}?quick_pull=1&title=${prTitle}&body=${prBody}`;
prLink = `\n[Create a PR](${prUrl})`;
}
} catch (error) {
+2 -2
View File
@@ -1,6 +1,6 @@
import { Octokit } from "@octokit/rest";
import { graphql } from "@octokit/graphql";
import { GITHUB_API_URL } from "./config";
import { GITHUB_API_URL, GITHUB_GRAPHQL_URL } from "./config";
export type Octokits = {
rest: Octokit;
@@ -14,7 +14,7 @@ export function createOctokit(token: string): Octokits {
baseUrl: GITHUB_API_URL,
}),
graphql: graphql.defaults({
baseUrl: GITHUB_API_URL,
baseUrl: GITHUB_GRAPHQL_URL,
headers: {
authorization: `token ${token}`,
},
+13
View File
@@ -2,3 +2,16 @@ export const GITHUB_API_URL =
process.env.GITHUB_API_URL || "https://api.github.com";
export const GITHUB_SERVER_URL =
process.env.GITHUB_SERVER_URL || "https://github.com";
// GraphQL base URL for @octokit/graphql. GitHub Actions exposes the full GraphQL
// endpoint in GITHUB_GRAPHQL_URL (e.g. "https://HOST/api/graphql"), while
// @octokit/graphql appends "/graphql" to whatever baseUrl it is given, so a
// single trailing "/graphql" is stripped here to avoid "/graphql/graphql".
// When GITHUB_GRAPHQL_URL is unset we fall back to GITHUB_API_URL, preserving the
// existing behavior where @octokit/graphql rewrites a REST ".../api/v3" base to
// ".../api/graphql". The trailing-slash trim keeps that rewrite working.
export const GITHUB_GRAPHQL_URL = (
process.env.GITHUB_GRAPHQL_URL || GITHUB_API_URL
)
.replace(/\/+$/, "")
.replace(/\/graphql$/, "");
+9 -2
View File
@@ -7,6 +7,7 @@ export const PR_QUERY = `
title
body
author {
__typename
login
}
baseRefName
@@ -25,7 +26,7 @@ export const PR_QUERY = `
additions
deletions
state
labels(first: 1) {
labels(first: 100) {
nodes {
name
}
@@ -57,6 +58,7 @@ export const PR_QUERY = `
databaseId
body
author {
__typename
login
}
createdAt
@@ -70,6 +72,7 @@ export const PR_QUERY = `
id
databaseId
author {
__typename
login
}
body
@@ -84,7 +87,9 @@ export const PR_QUERY = `
body
path
line
diffHunk
author {
__typename
login
}
createdAt
@@ -107,13 +112,14 @@ export const ISSUE_QUERY = `
title
body
author {
__typename
login
}
createdAt
updatedAt
lastEditedAt
state
labels(first: 1) {
labels(first: 100) {
nodes {
name
}
@@ -124,6 +130,7 @@ export const ISSUE_QUERY = `
databaseId
body
author {
__typename
login
}
createdAt
+6
View File
@@ -282,6 +282,12 @@ export function isPullRequestReviewCommentEvent(
return context.eventName === "pull_request_review_comment";
}
export function isWorkflowRunEvent(
context: GitHubContext,
): context is AutomationContext & { payload: WorkflowRunEvent } {
return context.eventName === "workflow_run";
}
export function isIssuesAssignedEvent(
context: GitHubContext,
): context is ParsedGitHubContext & { payload: IssuesAssignedEvent } {
+185 -33
View File
@@ -1,5 +1,7 @@
import { execFileSync } from "child_process";
import type { IssuesEvent } from "@octokit/webhooks-types";
import type { Octokits } from "../api/client";
import { isSameLogin } from "../validation/actor";
import { ISSUE_QUERY, PR_QUERY, USER_QUERY } from "../api/queries/github";
import {
isIssueCommentEvent,
@@ -22,6 +24,7 @@ import type { CommentWithImages } from "../utils/image-downloader";
import { downloadCommentImages } from "../utils/image-downloader";
import {
parseActorFilter,
resolveActorName,
shouldIncludeCommentByActor,
} from "../utils/actor-filter";
@@ -29,6 +32,12 @@ import {
* Extracts the trigger timestamp from the GitHub webhook payload.
* This timestamp represents when the triggering comment/review/event was created.
*
* For `issues` and `pull_request` events there is no dedicated trigger
* object in the payload, so the issue/PR's own timestamps from the webhook
* snapshot are used: `created_at` for opened events, otherwise `updated_at`
* (falling back to `created_at`). For issues labeled/assigned events,
* prefer resolveTriggerTimestamp() which looks up the exact event time.
*
* @param context - Parsed GitHub context from webhook
* @returns ISO timestamp string or undefined if not available
*/
@@ -41,11 +50,151 @@ export function extractTriggerTimestamp(
return context.payload.review.submitted_at || undefined;
} else if (isPullRequestReviewCommentEvent(context)) {
return context.payload.comment.created_at || undefined;
} else if (isIssuesEvent(context)) {
const issue = context.payload.issue;
if (context.eventAction === "opened") {
return issue?.created_at || issue?.updated_at || undefined;
}
// updated_at reflects the last comment or edit on the issue, so the
// newest pre-existing comment can share this timestamp and be excluded
// along with anything newer.
return issue?.updated_at || issue?.created_at || undefined;
} else if (isPullRequestEvent(context)) {
const pullRequest = context.payload.pull_request;
if (context.eventAction === "opened") {
return pullRequest?.created_at || pullRequest?.updated_at || undefined;
}
return pullRequest?.updated_at || pullRequest?.created_at || undefined;
}
return undefined;
}
/**
* Resolves the trigger timestamp for the event, consulting the GitHub API
* where the webhook payload does not carry an exact time for the triggering
* action.
*
* For issues labeled/assigned events the label/assignment carries no
* timestamp of its own in the payload, so the matching entry in the issue's
* event history is looked up and its `created_at` is used. If the lookup
* fails, this falls back to extractTriggerTimestamp().
*
* @param context - Parsed GitHub context from webhook
* @param octokits - GitHub API clients
* @returns ISO timestamp string or undefined if not available
*/
export async function resolveTriggerTimestamp(
context: ParsedGitHubContext,
octokits: Octokits,
): Promise<string | undefined> {
if (
isIssuesEvent(context) &&
(context.eventAction === "labeled" || context.eventAction === "assigned")
) {
const eventTime = await findIssueEventTime(context, octokits);
if (eventTime) {
return eventTime;
}
console.warn(
`Could not resolve the ${context.eventAction} event time for issue #${context.entityNumber}; falling back to the webhook payload timestamps`,
);
}
return extractTriggerTimestamp(context);
}
/**
* Looks up the most recent labeled/assigned event on the issue that matches
* the label or assignee in the webhook payload, returning its created_at.
*/
async function findIssueEventTime(
context: ParsedGitHubContext & { payload: IssuesEvent },
octokits: Octokits,
): Promise<string | undefined> {
const payload = context.payload;
let matches: (event: {
event: string;
label?: { name?: string | null };
assignee?: { login?: string } | null;
actor?: { login?: string } | null;
assigner?: { login?: string } | null;
}) => boolean;
// Only the event from the person who passed the permission check counts. Someone
// else removing and re-adding the label must not move the cutoff for comments.
// Assignments name the person who made them separately, as the assigner.
const byActor = (event: {
actor?: { login?: string } | null;
assigner?: { login?: string } | null;
}) => isSameLogin((event.assigner ?? event.actor)?.login, context.actor);
if (payload.action === "labeled") {
const labelName = payload.label?.name;
if (!labelName) return undefined;
matches = (event) =>
event.event === "labeled" &&
event.label?.name === labelName &&
byActor(event);
} else if (payload.action === "assigned") {
const assigneeLogin = payload.assignee?.login;
if (!assigneeLogin) return undefined;
matches = (event) =>
event.event === "assigned" &&
event.assignee?.login === assigneeLogin &&
byActor(event);
} else {
return undefined;
}
try {
const events = await octokits.rest.paginate(
octokits.rest.issues.listEvents,
{
owner: context.repository.owner,
repo: context.repository.repo,
issue_number: context.entityNumber,
per_page: 100,
},
);
let latest: (typeof events)[number] | undefined;
for (const event of events.filter(matches)) {
if (
!latest ||
new Date(event.created_at).getTime() >
new Date(latest.created_at).getTime()
) {
latest = event;
}
}
// Labeling/assignment does not bump the issue's updated_at, so the event
// that fired this webhook cannot predate the payload snapshot's
// updated_at. An older match means the current event is not visible in
// the events API yet; ignore it rather than adopt a stale boundary.
const snapshotUpdatedAt = payload.issue?.updated_at;
if (
latest &&
snapshotUpdatedAt &&
new Date(latest.created_at).getTime() <
new Date(snapshotUpdatedAt).getTime()
) {
console.warn(
`Latest matching ${payload.action} event on issue #${context.entityNumber} predates the issue's updated_at; treating it as stale`,
);
return undefined;
}
return latest?.created_at || undefined;
} catch (error) {
console.warn(
`Failed to fetch events for issue #${context.entityNumber}:`,
error,
);
return undefined;
}
}
/**
* Extracts the original title from the GitHub webhook payload.
* This is the title as it existed when the trigger event occurred.
@@ -204,11 +353,9 @@ export function isBodySafeToUse(
* @param excludeActors - Comma-separated actors to exclude
* @returns Filtered array of comments
*/
export function filterCommentsByActor<T extends { author: { login: string } }>(
comments: T[],
includeActors: string = "",
excludeActors: string = "",
): T[] {
export function filterCommentsByActor<
T extends { author: { login: string; __typename?: string } | null },
>(comments: T[], includeActors: string = "", excludeActors: string = ""): T[] {
const includeParsed = parseActorFilter(includeActors);
const excludeParsed = parseActorFilter(excludeActors);
@@ -219,7 +366,10 @@ export function filterCommentsByActor<T extends { author: { login: string } }>(
return comments.filter((comment) =>
shouldIncludeCommentByActor(
comment.author.login,
// Normalizes App actors to their "[bot]"-suffixed name, which is the form
// filter patterns are written in. Also maps deleted ("ghost") accounts,
// whose author is null, to "ghost" so filtering never dereferences null.
resolveActorName(comment.author),
includeParsed,
excludeParsed,
),
@@ -290,7 +440,12 @@ export async function fetchGitHubData({
if (prResult.repository.pullRequest) {
const pullRequest = prResult.repository.pullRequest;
contextData = pullRequest;
changedFiles = pullRequest.files.nodes || [];
if (pullRequest.files === null) {
console.warn(
`GitHub did not return the file list for PR #${prNumber} (diff likely too large); proceeding without file-level context`,
);
}
changedFiles = pullRequest.files?.nodes ?? [];
comments = filterCommentsByActor(
filterCommentsToTriggerTime(
pullRequest.comments?.nodes || [],
@@ -378,34 +533,26 @@ export async function fetchGitHubData({
body: c.body,
}));
// Filter review bodies to trigger time
const filteredReviewBodies = reviewData?.nodes
? filterReviewsToTriggerTime(reviewData.nodes, triggerTime).filter(
(r) => r.body,
)
: [];
const reviewBodies: CommentWithImages[] = filteredReviewBodies.map((r) => ({
type: "review_body" as const,
id: r.databaseId,
pullNumber: prNumber,
body: r.body,
}));
// Filter review comments to trigger time and by actor
// Filter reviews and inline review comments to trigger time and by actor
// before building anything from them. The trigger-time filter is the TOCTOU
// protection applied to issue/PR comments and the body above: it drops
// anything submitted, created, or edited at/after the trigger so an attacker
// cannot inject content into the prompt after an authorized trigger. Without
// it, review bodies and inline review comments would reach the prompt
// verbatim regardless of when they landed.
if (reviewData && reviewData.nodes) {
// Filter reviews by actor
// Drop reviews submitted or edited after the trigger, then filter by actor.
reviewData.nodes = filterCommentsByActor(
reviewData.nodes,
filterReviewsToTriggerTime(reviewData.nodes, triggerTime),
includeCommentsByActor,
excludeCommentsByActor,
);
// Also filter inline review comments within each review
// Apply the same trigger-time + actor filtering to inline review comments.
reviewData.nodes.forEach((review) => {
if (review.comments?.nodes) {
review.comments.nodes = filterCommentsByActor(
review.comments.nodes,
filterCommentsToTriggerTime(review.comments.nodes, triggerTime),
includeCommentsByActor,
excludeCommentsByActor,
);
@@ -413,14 +560,19 @@ export async function fetchGitHubData({
});
}
const allReviewComments =
reviewData?.nodes?.flatMap((r) => r.comments?.nodes ?? []) ?? [];
const filteredReviewComments = filterCommentsToTriggerTime(
allReviewComments,
triggerTime,
);
// Build the image-processing lists from the already-filtered review nodes,
// so reviews/comments excluded from the prompt are not processed for images.
const reviewBodies: CommentWithImages[] = (reviewData?.nodes ?? [])
.filter((r) => r.body)
.map((r) => ({
type: "review_body" as const,
id: r.databaseId,
pullNumber: prNumber,
body: r.body,
}));
const reviewComments: CommentWithImages[] = filteredReviewComments
const reviewComments: CommentWithImages[] = (reviewData?.nodes ?? [])
.flatMap((r) => r.comments?.nodes ?? [])
.filter((c) => c.body && !c.isMinimized)
.map((c) => ({
type: "review_comment" as const,
+23 -7
View File
@@ -8,6 +8,11 @@ import type {
import type { GitHubFileWithSHA } from "./fetcher";
import { sanitizeContent } from "../utils/sanitizer";
function formatLabels(labelNodes: Array<{ name: string }>): string {
if (labelNodes.length === 0) return "none";
return labelNodes.map((l) => l.name).join(", ");
}
export function formatContext(
contextData: GitHubPullRequest | GitHubIssue,
isPR: boolean,
@@ -16,19 +21,21 @@ export function formatContext(
const prData = contextData as GitHubPullRequest;
const sanitizedTitle = sanitizeContent(prData.title);
return `PR Title: ${sanitizedTitle}
PR Author: ${prData.author.login}
PR Author: ${prData.author?.login ?? "ghost"}
PR Branch: ${prData.headRefName} -> ${prData.baseRefName}
PR State: ${prData.state}
PR Labels: ${formatLabels(prData.labels.nodes)}
PR Additions: ${prData.additions}
PR Deletions: ${prData.deletions}
Total Commits: ${prData.commits.totalCount}
Changed Files: ${prData.files.nodes.length} files`;
Changed Files: ${prData.files ? `${prData.files.nodes.length} files` : "unknown (file list unavailable)"}`;
} else {
const issueData = contextData as GitHubIssue;
const sanitizedTitle = sanitizeContent(issueData.title);
return `Issue Title: ${sanitizedTitle}
Issue Author: ${issueData.author.login}
Issue State: ${issueData.state}`;
Issue Author: ${issueData.author?.login ?? "ghost"}
Issue State: ${issueData.state}
Issue Labels: ${formatLabels(issueData.labels.nodes)}`;
}
}
@@ -64,7 +71,7 @@ export function formatComments(
body = sanitizeContent(body);
return `[${comment.author.login} at ${comment.createdAt}]: ${body}`;
return `[${comment.author?.login ?? "ghost"} at ${comment.createdAt}]: ${body}`;
})
.join("\n\n");
}
@@ -78,7 +85,7 @@ export function formatReviewComments(
}
const formattedReviews = reviewData.nodes.map((review) => {
let reviewOutput = `[Review by ${review.author.login} at ${review.submittedAt}]: ${review.state}`;
let reviewOutput = `[Review by ${review.author?.login ?? "ghost"} at ${review.submittedAt}]: ${review.state}`;
if (review.body && review.body.trim()) {
let body = review.body;
@@ -111,7 +118,16 @@ export function formatReviewComments(
body = sanitizeContent(body);
return ` [Comment on ${comment.path}:${comment.line || "?"}]: ${body}`;
let formatted = ` [Comment on ${comment.path}:${comment.line || "?"}]: ${body}`;
// The diff hunk is the code the comment was left on. Without it the
// comment arrives without the context it was written against.
if (comment.diffHunk) {
const diffHunk = sanitizeContent(comment.diffHunk);
formatted += `\n Diff context:\n\`\`\`diff\n${diffHunk}\n\`\`\``;
}
return formatted;
})
.join("\n");
if (comments) {
+47 -7
View File
@@ -1,7 +1,18 @@
import type { Octokits } from "../api/client";
import { GITHUB_SERVER_URL } from "../api/config";
import { encodeBranchNameForUrl } from "./comments/common";
import { $ } from "bun";
async function pushLfsObjects(branch: string): Promise<void> {
let tracked: string;
try {
tracked = await $`git lfs ls-files --name-only`.quiet().text();
} catch {
return; // Git LFS isn't installed here
}
if (tracked.trim()) await $`git lfs push origin ${branch}`;
}
export async function checkAndCommitOrDeleteBranch(
octokit: Octokits,
owner: string,
@@ -9,10 +20,32 @@ export async function checkAndCommitOrDeleteBranch(
claudeBranch: string | undefined,
baseBranch: string,
useCommitSigning: boolean,
restoredConfigPaths: string[] = [],
): Promise<{ shouldDeleteBranch: boolean; branchLink: string }> {
let branchLink = "";
let shouldDeleteBranch = false;
// On pull requests, restoreConfigFromBase replaces .claude/, CLAUDE.md and
// friends with the base branch's versions and leaves them unstaged so the
// revert does not reach a commit. Auto-committing with a bare `git add -A`
// would stage them anyway and push a silent revert of the PR author's own
// config onto their branch.
//
// The exclusion is driven by what was actually restored rather than applied
// unconditionally: this path also runs for issues, where no restore happens
// and Claude may legitimately have been asked to edit CLAUDE.md or
// .claude/settings.json. Excluding those there would silently drop the work.
const pathspecArgs =
restoredConfigPaths.length > 0
? ["--", ".", ...restoredConfigPaths.map((p) => `:(exclude,literal)${p}`)]
: [];
if (pathspecArgs.length > 0) {
console.log(
`Excluding base-restored config from auto-commit: ${restoredConfigPaths.join(", ")}`,
);
}
if (claudeBranch) {
// First check if the branch exists remotely
let branchExistsRemotely = false;
@@ -57,21 +90,28 @@ export async function checkAndCommitOrDeleteBranch(
// Check for uncommitted changes using git status
try {
const gitStatus = await $`git status --porcelain`.quiet();
// Scoped the same way as the staging below: if the restored config
// is the only dirty entry there is no real work, and the branch
// should be treated as empty rather than receiving a pure revert.
const gitStatus =
await $`git status --porcelain ${pathspecArgs}`.quiet();
const hasUncommittedChanges =
gitStatus.stdout.toString().trim().length > 0;
if (hasUncommittedChanges) {
console.log("Found uncommitted changes, committing them...");
// Add all changes
await $`git add -A`;
// Add all changes, minus anything restored from the base branch
await $`git add -A ${pathspecArgs}`;
// Commit with a descriptive message
const runId = process.env.GITHUB_RUN_ID || "unknown";
const commitMessage = `Auto-commit: Save uncommitted changes from Claude\n\nRun ID: ${runId}`;
await $`git commit -m ${commitMessage}`;
// Hooks are off by now, and Git LFS uploads from a hook.
await pushLfsObjects(claudeBranch);
// Push the changes
await $`git push origin ${claudeBranch}`;
@@ -80,7 +120,7 @@ export async function checkAndCommitOrDeleteBranch(
);
// Set branch link since we now have commits
const branchUrl = `${GITHUB_SERVER_URL}/${owner}/${repo}/tree/${claudeBranch}`;
const branchUrl = `${GITHUB_SERVER_URL}/${owner}/${repo}/tree/${encodeBranchNameForUrl(claudeBranch)}`;
branchLink = `\n[View branch](${branchUrl})`;
} else {
console.log(
@@ -91,7 +131,7 @@ export async function checkAndCommitOrDeleteBranch(
} catch (gitError) {
console.error("Error checking/committing changes:", gitError);
// If we can't check git status, assume the branch might have changes
const branchUrl = `${GITHUB_SERVER_URL}/${owner}/${repo}/tree/${claudeBranch}`;
const branchUrl = `${GITHUB_SERVER_URL}/${owner}/${repo}/tree/${encodeBranchNameForUrl(claudeBranch)}`;
branchLink = `\n[View branch](${branchUrl})`;
}
} else {
@@ -102,13 +142,13 @@ export async function checkAndCommitOrDeleteBranch(
}
} else {
// Only add branch link if there are commits
const branchUrl = `${GITHUB_SERVER_URL}/${owner}/${repo}/tree/${claudeBranch}`;
const branchUrl = `${GITHUB_SERVER_URL}/${owner}/${repo}/tree/${encodeBranchNameForUrl(claudeBranch)}`;
branchLink = `\n[View branch](${branchUrl})`;
}
} catch (error) {
console.error("Error comparing commits on Claude branch:", error);
// If we can't compare but the branch exists remotely, include the branch link
const branchUrl = `${GITHUB_SERVER_URL}/${owner}/${repo}/tree/${claudeBranch}`;
const branchUrl = `${GITHUB_SERVER_URL}/${owner}/${repo}/tree/${encodeBranchNameForUrl(claudeBranch)}`;
branchLink = `\n[View branch](${branchUrl})`;
}
}
+133 -29
View File
@@ -9,10 +9,14 @@
import { $ } from "bun";
import { execFileSync } from "child_process";
import type { ParsedGitHubContext } from "../context";
import { isPullRequestEvent } from "../context";
import { retryWithBackoff } from "../../utils/retry";
import type { GitHubPullRequest } from "../types";
import type { Octokits } from "../api/client";
import type { FetchDataResult } from "../data/fetcher";
import { generateBranchName } from "../../utils/branch-template";
import { fetchDepthArgs } from "./fetch-depth";
import { forksAllowed } from "../validation/fork-pr";
/**
* Extracts the first label from GitHub data, or returns undefined if no labels exist
@@ -27,14 +31,15 @@ function extractFirstLabel(githubData: FetchDataResult): string | undefined {
* This prevents command injection by ensuring only safe characters are used.
*
* Valid branch names:
* - Start with alphanumeric character (not dash, to prevent option injection)
* - Contain only alphanumeric, forward slash, hyphen, underscore, period, or hash (#)
* - Start with alphanumeric character, underscore, or @ (not dash, to prevent option injection)
* - Contain only alphanumeric, forward slash, hyphen, underscore, period, hash (#), plus (+), comma (,), at sign (@), or parentheses
* - Do not start or end with a period
* - Do not end with a slash
* - Do not contain '..' (path traversal)
* - Do not contain '//' (consecutive slashes)
* - Do not end with '.lock'
* - Do not contain '@{'
* - Are not the single character '@' (HEAD shorthand in git revision syntax)
* - Do not contain control characters or special git characters (~^:?*[\])
*/
export function validateBranchName(branchName: string): void {
@@ -58,18 +63,26 @@ export function validateBranchName(branchName: string): void {
);
}
// Strict whitelist pattern: alphanumeric start, then alphanumeric/slash/hyphen/underscore/period/hash/plus/comma.
// Strict whitelist pattern: alphanumeric or @ start, then alphanumeric/slash/hyphen/underscore/period/hash/plus/comma/at-sign/parentheses.
// # is valid per git-check-ref-format and commonly used in branch names like "fix/#123-description".
// + is valid per git-check-ref-format and generated by Claude Code's EnterWorktree tool when
// converting worktree names containing "/" (e.g. "feat/foo" becomes "worktree-feat+foo").
// , is valid per git-check-ref-format and commonly appears in branch names derived from titles
// or external identifiers (e.g. place names like "feature/paris,france").
// @ is valid per git-check-ref-format anywhere in a ref name, including the first character
// (e.g. ticket conventions like "TICKET-123@add-feature" or prefixes like "@hotfix/...");
// the bare name "@" (HEAD shorthand) and the "@{" sequence (reflog syntax) are rejected below.
// _ is valid per git-check-ref-format anywhere in a ref name, including the first character;
// leading underscores are a common convention for release/internal branches (e.g.
// "_release/v1.2.3"), which previously failed validation as a PR's base branch.
// Parentheses are valid per git-check-ref-format and commonly appear in branch names that
// use Conventional Commit-style scopes (e.g. "feat(parser)-handle-empty-input").
// All git calls use execFileSync (not shell interpolation), so none of these characters carry injection risk.
const validPattern = /^[a-zA-Z0-9][a-zA-Z0-9/_.#+,-]*$/;
const validPattern = /^[a-zA-Z0-9@_][a-zA-Z0-9/_.#+,@()-]*$/;
if (!validPattern.test(branchName)) {
throw new Error(
`Invalid branch name: "${branchName}". Branch names must start with an alphanumeric character and contain only alphanumeric characters, forward slashes, hyphens, underscores, periods, hashes (#), plus signs (+), or commas (,).`,
`Invalid branch name: "${branchName}". Branch names must start with an alphanumeric character, underscore, or '@' and contain only alphanumeric characters, forward slashes, hyphens, underscores, periods, hashes (#), plus signs (+), commas (,), at signs (@), or parentheses.`,
);
}
@@ -112,6 +125,15 @@ export function validateBranchName(branchName: string): void {
`Invalid branch name: "${branchName}". Branch names cannot contain '@{'`,
);
}
// Per git-check-ref-format, a refname cannot be the single character "@"; "@" also
// resolves to HEAD in git revision syntax, so a bare "@" must never reach git as a
// branch argument where it could be interpreted as a revision instead.
if (branchName === "@") {
throw new Error(
`Invalid branch name: "@". Branch names cannot be the single character '@'.`,
);
}
}
/**
@@ -123,20 +145,24 @@ export function validateBranchName(branchName: string): void {
*
* @param args - Git command arguments (e.g., ["checkout", "branch-name"])
*/
function execGit(args: string[]): void {
execFileSync("git", args, { stdio: "inherit", env: process.env });
function execGit(args: string[], env: NodeJS.ProcessEnv = process.env): void {
execFileSync("git", args, { stdio: "inherit", env });
}
export type BranchInfo = {
baseBranch: string;
claudeBranch?: string;
currentBranch: string;
/** Set when the head of an open pull request was checked out. */
isFork?: boolean;
};
export async function setupBranch(
octokits: Octokits,
githubData: FetchDataResult,
context: ParsedGitHubContext,
triggerTime: string | undefined,
skipLfsDownloads: boolean,
): Promise<BranchInfo> {
const { owner, repo } = context.repository;
const entityNumber = context.entityNumber;
@@ -154,40 +180,90 @@ export async function setupBranch(
);
// Fall through to create a new branch like we do for issues
} else {
// run.ts has skipped these already.
if (prData.isCrossRepository && !forksAllowed()) {
throw new Error(`PR #${entityNumber} is from a fork`);
}
// Handle open PR: Checkout the PR branch
console.log("This is an open PR, checking out PR branch...");
const branchName = prData.headRefName;
const headSha = prData.headRefOid;
// Determine optimal fetch depth based on PR commit count, with a minimum of 20
// Determine optimal fetch depth based on PR commit count, with a minimum
// of 20. Only applied to a checkout that is already shallow — see
// fetchDepthArgs.
const commitCount = prData.commits.totalCount;
const fetchDepth = Math.max(commitCount, 20);
const depthArgs = fetchDepthArgs(fetchDepth);
console.log(
`PR #${entityNumber}: ${commitCount} commits, using fetch depth ${fetchDepth}`,
`PR #${entityNumber}: ${commitCount} commits, ${
depthArgs.length > 0
? `using fetch depth ${fetchDepth}`
: "fetching without a depth limit (checkout has full history)"
}`,
);
// Validate branch names before use to prevent command injection
validateBranchName(branchName);
if (!/^[0-9a-f]{40}$/.test(headSha)) {
throw new Error(
`Unexpected head SHA for PR #${entityNumber}: ${headSha}`,
);
}
// For cross-repository (fork) PRs, fetch via the pull ref since the
// branch only exists on the fork's remote, not on origin.
if (prData.isCrossRepository) {
console.log(
`PR #${entityNumber} is from a fork, fetching via refs/pull/${entityNumber}/head...`,
if (isPullRequestEvent(context)) {
const eventSha = context.payload.pull_request.head.sha;
if (eventSha !== headSha) {
throw new Error(
`PR #${entityNumber} was updated after the trigger (${eventSha} -> ${headSha}). Push again or mention Claude to run on the latest version.`,
);
}
} else if (triggerTime) {
if (!prData.headRepository) {
throw new Error(
`PR #${entityNumber}: head repository is not available, cannot pin the PR version`,
);
}
const pushedAt = await headPushedAt(
octokits,
{
owner: prData.headRepository.owner.login,
repo: prData.headRepository.name,
},
branchName,
headSha,
);
if (new Date(pushedAt).getTime() >= new Date(triggerTime).getTime()) {
throw new Error(
`PR #${entityNumber} was updated after the trigger (${pushedAt} >= ${triggerTime}). Mention Claude again to run on the latest version.`,
);
}
}
const localBranch = prData.isCrossRepository
? `pr-${entityNumber}`
: branchName;
console.log(`PR #${entityNumber}: pinning ${localBranch} to ${headSha}`);
execGit(["fetch", "origin", ...depthArgs, headSha]);
execGit(
["checkout", "-B", localBranch, headSha, "--"],
// Files that Git LFS would download stay as pointer files.
skipLfsDownloads
? { ...process.env, GIT_LFS_SKIP_SMUDGE: "1" }
: process.env,
);
const checkedOut = execFileSync("git", ["rev-parse", "HEAD"], {
encoding: "utf8",
}).trim();
if (checkedOut !== headSha) {
throw new Error(
`PR #${entityNumber}: checked out ${checkedOut}, expected ${headSha}`,
);
execGit([
"fetch",
"origin",
`--depth=${fetchDepth}`,
`pull/${entityNumber}/head:${branchName}`,
]);
} else {
// Execute git commands to checkout PR branch (dynamic depth based on PR size)
// Using execFileSync instead of shell template literals for security
execGit(["fetch", "origin", `--depth=${fetchDepth}`, branchName]);
}
execGit(["checkout", branchName, "--"]);
console.log(`Successfully checked out PR branch for PR #${entityNumber}`);
@@ -197,7 +273,8 @@ export async function setupBranch(
return {
baseBranch,
currentBranch: branchName,
currentBranch: localBranch,
isFork: prData.isCrossRepository,
};
}
}
@@ -272,6 +349,11 @@ export async function setupBranch(
// Branch doesn't exist (non-zero exit code), continue with generated name
}
// Validate before either path uses the name. The signing path hands it to
// the file ops server rather than to git, so without this an invalid
// template only surfaces as a 422 on the first commit.
validateBranchName(newBranch);
// For commit signing, defer branch creation to the file ops server
if (context.inputs.useCommitSigning) {
console.log(
@@ -281,7 +363,7 @@ export async function setupBranch(
// Ensure we're on the source branch
console.log(`Fetching and checking out source branch: ${sourceBranch}`);
validateBranchName(sourceBranch);
execGit(["fetch", "origin", sourceBranch, "--depth=1"]);
execGit(["fetch", "origin", sourceBranch, ...fetchDepthArgs(1)]);
execGit(["checkout", sourceBranch, "--"]);
return {
@@ -299,8 +381,7 @@ export async function setupBranch(
// Fetch and checkout the source branch first to ensure we branch from the correct base
console.log(`Fetching and checking out source branch: ${sourceBranch}`);
validateBranchName(sourceBranch);
validateBranchName(newBranch);
execGit(["fetch", "origin", sourceBranch, "--depth=1"]);
execGit(["fetch", "origin", sourceBranch, ...fetchDepthArgs(1)]);
execGit(["checkout", sourceBranch, "--"]);
// Create and checkout the new branch from the source branch
@@ -320,3 +401,26 @@ export async function setupBranch(
process.exit(1);
}
}
async function headPushedAt(
octokits: Octokits,
headRepo: { owner: string; repo: string },
branchName: string,
headSha: string,
): Promise<string> {
const { data } = await retryWithBackoff(() =>
octokits.rest.request("GET /repos/{owner}/{repo}/activity", {
owner: headRepo.owner,
repo: headRepo.repo,
ref: `refs/heads/${branchName}`,
per_page: 1,
}),
);
const latest = (data as Array<{ after: string; timestamp: string }>)[0];
if (!latest || latest.after !== headSha) {
throw new Error(
`Could not determine when ${headSha} was pushed to ${headRepo.owner}/${headRepo.repo}:${branchName}`,
);
}
return latest.timestamp;
}
+7 -3
View File
@@ -1,4 +1,6 @@
import { GITHUB_SERVER_URL } from "../api/config";
import { redactSecrets } from "../utils/sanitizer";
import { encodeBranchNameForUrl } from "./comments/common";
export type ExecutionDetails = {
total_cost_usd?: number;
@@ -160,7 +162,7 @@ export function updateCommentBody(input: CommentUpdateInput): string {
// Extract owner/repo from jobUrl
const repoMatch = jobUrl.match(/github\.com\/([^\/]+)\/([^\/]+)\//);
if (repoMatch) {
branchUrl = `${GITHUB_SERVER_URL}/${repoMatch[1]}/${repoMatch[2]}/tree/${finalBranchName}`;
branchUrl = `${GITHUB_SERVER_URL}/${repoMatch[1]}/${repoMatch[2]}/tree/${encodeBranchNameForUrl(finalBranchName)}`;
}
}
@@ -181,9 +183,11 @@ export function updateCommentBody(input: CommentUpdateInput): string {
// Build the new body with blank line between header and separator
let newBody = `${header}${links}`;
// Add error details if available
// Add error details if available. The message may embed runtime credentials
// (e.g. a token in a git remote URL) that are not registered as workflow
// secrets, so redact known formats before posting.
if (actionFailed && errorDetails) {
newBody += `\n\n\`\`\`\n${errorDetails}\n\`\`\``;
newBody += `\n\n\`\`\`\n${redactSecrets(errorDetails)}\n\`\`\``;
}
newBody += `\n\n---\n`;
+6 -1
View File
@@ -12,12 +12,17 @@ export function createJobRunLink(
return `[View job run](${jobRunUrl})`;
}
/** Encode Git-ref path segments without turning `/` into `%2F`. */
export function encodeBranchNameForUrl(branchName: string): string {
return branchName.split("/").map(encodeURIComponent).join("/");
}
export function createBranchLink(
owner: string,
repo: string,
branchName: string,
): string {
const branchUrl = `${GITHUB_SERVER_URL}/${owner}/${repo}/tree/${branchName}`;
const branchUrl = `${GITHUB_SERVER_URL}/${owner}/${repo}/tree/${encodeBranchNameForUrl(branchName)}`;
return `\n[View branch](${branchUrl})`;
}
+55
View File
@@ -0,0 +1,55 @@
import type { ParsedGitHubContext } from "../context";
import { validateBranchName } from "./branch";
// Tag mode pushes a fork's commits to these.
const FORK_BRANCH = /^pr-\d+$/;
/**
* Returns the branch to restore Claude's config from on a pull request: the base, or
* for an untrusted author the base_branch input or the default branch. A pull request
* that targets the default branch keeps its base, as does a bot's in this repository,
* unless the base holds a fork's commits.
*
* @param isFork - Whether the checked-out head is in a fork. Read only for
* issue_comment events. Undefined counts as a fork.
*/
export function chooseConfigBranch(
context: ParsedGitHubContext,
baseBranch: string,
isAuthorUntrusted: boolean,
isFork?: boolean,
): string {
const { payload, repository, inputs } = context;
const { user } =
"pull_request" in payload ? payload.pull_request : payload.issue;
// A deleted fork has no head repository.
const isSameRepository =
"pull_request" in payload
? payload.pull_request.head.repo?.full_name?.toLowerCase() ===
repository.full_name.toLowerCase()
: isFork === false;
const { default_branch: defaultBranch } = repository;
// The pull request's author names its head branch, so that is never trusted.
const inputBranch =
"pull_request" in payload &&
payload.pull_request.head.ref === inputs.baseBranch
? undefined
: inputs.baseBranch;
if (
!isAuthorUntrusted ||
(isSameRepository &&
user?.type === "Bot" &&
!FORK_BRANCH.test(baseBranch)) ||
(defaultBranch && baseBranch === defaultBranch)
) {
return baseBranch;
}
const branch = inputBranch || defaultBranch;
if (!branch) {
throw new Error(
"Could not find the repository's default branch in the event. Set the base_branch input to the branch to take Claude's config from.",
);
}
validateBranchName(branch);
return branch;
}
+381
View File
@@ -0,0 +1,381 @@
import * as core from "@actions/core";
import { mkdirSync, symlinkSync, writeFileSync } from "fs";
import { isAbsolute, posix, relative, resolve, sep } from "path";
import {
ACCUMULATE_DELIMITER,
parseClaudeArgsToExtraArgs,
} from "../../../base-action/src/parse-sdk-options";
import { SENSITIVE_PATHS } from "./restore-config";
import {
fold,
fsPath,
git,
hasRealParents,
isFile,
isLink,
isSubmodule,
onDisk,
readBlobs,
remove,
snapshotForReview,
} from "./tool-config";
import type { Entry } from "./tool-config";
// The arguments that name config in files.
const FILE_FLAGS = new Set([
"agents",
"append-subagent-system-prompt-file",
"append-system-prompt-file",
"mcp-config",
"plugin-dir",
"plugin-dir-no-mcp",
"settings",
"system-prompt-file",
]);
// Only the project's config in this directory.
const PROJECT_FLAG = "project-config-root";
const MARKETPLACE_FILE = ".claude-plugin/marketplace.json";
// Git's entries by path, relative to the working directory. As in tool-config.ts,
// Latin-1 characters stand for bytes.
type Entries = Map<string, Entry>;
const toBytes = (path: string): string => Buffer.from(path).toString("latin1");
const toText = (path: string): string => fsPath(path).toString();
const isSameOrInside = (path: string, parent: string): boolean =>
path === parent || path.startsWith(`${parent}/`);
const isSensitive = (path: string): boolean =>
SENSITIVE_PATHS.some((sensitive) => isSameOrInside(path, sensitive));
// Each is "<mode> <other fields>\t<path>".
function parseEntries(listing: Buffer, oidField: number): Entries {
const entries: Entries = new Map();
for (const line of listing.toString("latin1").split("\0")) {
const [info, path] = line.split(/\t(.*)/s);
if (!info || !path) continue;
const fields = info.split(" ");
entries.set(path, { mode: fields[0]!, oid: fields[oidField]! });
}
return entries;
}
/**
* A path as it is relative to the checkout, or undefined if it's outside. Links are not
* followed, because the pull request can change them. The workspace can have another
* name than the working directory, which is free of links.
*/
function inCheckout(path: string): string | undefined {
const roots = [process.cwd(), process.env.GITHUB_WORKSPACE || process.cwd()];
for (const root of roots) {
const inRoot = relative(root, resolve(path));
if (!isAbsolute(inRoot) && inRoot.split(sep)[0] !== "..") {
return toBytes(inRoot.split(sep).join("/"));
}
}
return undefined;
}
function isJson(value: string): boolean {
if (!value.startsWith("{")) return false;
try {
JSON.parse(value);
return true;
} catch {
return false;
}
}
const WHOLE_REPOSITORY =
"The workflow takes a plugin from the whole repository, which can't come from a trusted branch on a pull request from an author without write access. Put the plugin in a subdirectory, or set restrict_untrusted_prs to false.";
const UNCLEAR_PATH =
'A path in the workflow\'s inputs has ".." after another name, so it can\'t come from a trusted branch on a pull request from an author without write access. Write it without "..", or set restrict_untrusted_prs to false.';
const OTHER_SUBMODULE =
"The pull request and the trusted branch have different commits of a submodule that the workflow's inputs name, so that branch's version can't be used. Bring the pull request up to date, put the files in the repository, or set restrict_untrusted_prs to false.";
class PathError extends Error {}
/**
* Whether a path has ".." after a name in the checkout, which the pull request could
* make a link.
*/
function goesBack(path: string): boolean {
const parts = path.split(/[\\/]/);
return parts.some(
(part, index) =>
part === ".." && !!inCheckout(parts.slice(0, index).join("/") || "/"),
);
}
type NamedPaths = { paths: string[]; marketplaces: string[] };
/** The paths in this checkout that the workflow's inputs name as config. */
export function namedConfigPaths(claudeArgs: string): NamedPaths {
const args = Object.entries(parseClaudeArgsToExtraArgs(claudeArgs)).flatMap(
([key, value]) => {
const [flag, inlineValue] = key.split(/=(.*)/s);
return [...(value?.split(ACCUMULATE_DELIMITER) ?? []), inlineValue ?? ""]
.map((path) => path.trim())
.flatMap((path) => (path ? [{ flag: flag!, path }] : []));
},
);
const fromArgs = args.flatMap(({ flag, path }) =>
flag === PROJECT_FLAG
? SENSITIVE_PATHS.map((name) => `${path}/${name}`)
: FILE_FLAGS.has(flag)
? [path]
: [],
);
const toPaths = (values: string[]): string[] =>
values
.map((value) => value.trim())
.filter((value) => value && !isJson(value))
.flatMap((value) => {
if (goesBack(value)) throw new PathError(UNCLEAR_PATH);
const path = inCheckout(value);
return path === undefined ? [] : [path];
});
return {
paths: toPaths([process.env.INPUT_SETTINGS ?? "", ...fromArgs]),
marketplaces: toPaths(
(process.env.INPUT_PLUGIN_MARKETPLACES ?? "")
.split("\n")
.filter((entry) => /^\s*((\.\.?)?\/|[a-z]:[\\/])/i.test(entry)),
),
};
}
type RestoredPaths = { replaced: string[]; denyRules: string[] };
/**
* On a pull request from an author without write access, takes the files that the
* workflow's inputs name from the base branch. Links are restored but not followed. Only
* files that the pull request changes, or whose .gitattributes it changes, are replaced,
* so what an earlier step created or changed stays. restoreConfigFromBase must run
* first, because it fetches the base branch. Like it, this changes the working tree and
* not the index.
*
* @returns The paths that were replaced, all named by the base branch or the workflow,
* and rules that stop a session from editing these and the rest of that config.
* @throws If the base branch's version can't be put in place.
*/
export function restoreConfigPathsFromBase(
baseBranch: string,
named: NamedPaths,
): RestoredPaths {
if (named.paths.length + named.marketplaces.length === 0) {
return { replaced: [], denyRules: [] };
}
let restored: RestoredPaths;
try {
restored = restoreConfigPaths(`refs/remotes/origin/${baseBranch}`, named);
} catch (error) {
if (error instanceof PathError) throw error;
const { code, status } = error as { code?: string; status?: number };
core.warning(
`Could not replace config files (${code ?? `exit code ${status ?? "unknown"}`})`,
);
throw new Error(
"Could not use the trusted branch's version of the files that the workflow's inputs name",
);
}
if (restored.replaced.length > 0) {
core.warning(
`The pull request and ${baseBranch} differ in files that the workflow's inputs name (${restored.replaced.join(", ")}), so that branch's version is used and Claude can't edit it. See docs/security.md.`,
);
}
return restored;
}
function restoreConfigPaths(ref: string, named: NamedPaths): RestoredPaths {
const base = parseEntries(git(["ls-tree", "-r", "-z", ref]), 2);
const pr = parseEntries(git(["ls-files", "--stage", "-z"]), 1);
// The rest of Claude's config has been restored already.
const targets = [
...named.paths,
...named.marketplaces.flatMap((path) => marketplacePaths(path, base)),
]
.map((path) => posix.join(path || ".").replace(/\/$/, ""))
.filter((path) => !isSameOrInside(path, "..") && !isSensitive(path));
if (targets.includes(".")) throw new PathError(WHOLE_REPOSITORY);
const replaced = replaceFiles([...new Set(targets)], base, pr).map(toText);
return {
replaced,
denyRules: [...new Set([...targets.map(toText), ...replaced])]
.map((path) => path.replace(/[\\*?[\]]/g, "\\$&"))
.flatMap((path) => [`Edit(./${path})`, `Edit(./${path}/**)`]),
};
}
/**
* What a local marketplace is made of. One at the root of the repository has only its
* own plugins in other directories.
*/
function marketplacePaths(path: string, base: Entries): string[] {
const isManifest = path.endsWith(".json");
// The CLI takes the directory above the manifest's as the marketplace's root.
if (isManifest && !path.includes("/")) return [path];
const root = isManifest
? posix.join(posix.dirname(path), "..")
: path === ""
? "."
: path;
if (root !== ".") return [root];
const manifest = isManifest ? path : MARKETPLACE_FILE;
const own = isManifest ? path : posix.dirname(MARKETPLACE_FILE);
const entry = base.get(manifest);
// Without one that the base branch has, there is nothing to take plugins from.
if (!entry || !isFile(entry)) return [own];
const { plugins, metadata } = JSON.parse(
readBlobs([entry])
.get(entry.oid)!
.toString()
.replace(/^\uFEFF/, ""),
) as {
plugins?: ({ source?: unknown } | null)[];
metadata?: { pluginRoot?: unknown };
};
const pluginRoot =
typeof metadata?.pluginRoot === "string" ? metadata.pluginRoot : "";
const sources = (plugins ?? []).flatMap((plugin) =>
typeof plugin?.source === "string" ? [plugin.source] : [],
);
return [
own,
...sources.flatMap((source) =>
// A bare name can be in the marketplace's plugin root.
[source, ...(pluginRoot ? [posix.join(pluginRoot, source)] : [])].map(
(dir) => toBytes(posix.join(".", dir)),
),
),
];
}
/**
* Whether a path that differs from a target at most in case, or in how a character is
* encoded, is in it or above it on this file system.
*/
function namesSameFile(name: string, target: string): boolean {
const names = name.split("/");
const targets = target.split("/");
const spelled = [
...targets.slice(0, names.length),
...names.slice(targets.length),
].join("/");
if (name === spelled) return true;
const [stats, other] = [onDisk(name), onDisk(spelled)];
return !!stats && !!other && stats.ino === other.ino;
}
/** The directories whose .gitattributes the pull request changes, ignoring case. */
function changedAttributes(base: Entries, pr: Entries): string[] {
const names = new Set([...base.keys(), ...pr.keys()]);
return [...names].flatMap((name) =>
fold(posix.basename(name)) === ".gitattributes" &&
base.get(name)?.oid !== pr.get(name)?.oid
? [fold(posix.dirname(name))]
: [],
);
}
const isInside = (name: string, foldedDir: string): boolean =>
foldedDir === "." || fold(name).startsWith(`${foldedDir}/`);
/** Puts the base branch's version of these paths in place. Returns those that differed. */
function replaceFiles(targets: string[], base: Entries, pr: Entries): string[] {
const foldedPr = [...pr.keys()].map((name) => [name, fold(name)] as const);
const namesByTarget = new Map(
targets.map((target) => {
const folded = fold(target);
const names = [
...[...base.keys()].filter(
(name) =>
isSameOrInside(name, target) || isSameOrInside(target, name),
),
// A pull request can also have a file or link where the base branch has a
// directory.
...foldedPr.flatMap(([name, other]) =>
(isSameOrInside(other, folded) || isSameOrInside(folded, other)) &&
namesSameFile(name, target)
? [name]
: [],
),
];
return [target, [...new Set(names)]] as const;
}),
);
const attributes = changedAttributes(base, pr);
const differs = (name: string): boolean => {
const [ours, theirs] = [base.get(name), pr.get(name)];
return (
ours?.oid !== theirs?.oid ||
ours?.mode !== theirs?.mode ||
// Git can write even an unchanged file differently then.
(isFile(theirs!) && attributes.some((dir) => isInside(name, dir)))
);
};
const differing = [...namesByTarget].flatMap(([target, names]) => {
// Names that differ only in case can be the same file.
const folded = new Set(names.filter(differs).map(fold));
const changed = names.filter((name) => folded.has(fold(name)));
return changed.length > 0 ? [[target, changed] as const] : [];
});
const changed = [...new Set(differing.flatMap(([, names]) => names))];
// Claude Code goes on without a plugin that is missing.
if (changed.some((name) => base.has(name) && isSubmodule(base.get(name)!))) {
throw new PathError(OTHER_SUBMODULE);
}
// All of these go before anything is written. A directory where the pull request has
// a file or link came from an earlier step.
const removed = new Set(
changed.filter((name) => {
const entry = pr.get(name);
const stats = entry && onDisk(name);
return (
stats &&
(!stats.isDirectory() || isSubmodule(entry)) &&
hasRealParents(name)
);
}),
);
for (const name of removed) {
const entry = pr.get(name)!;
if (entry.oid !== base.get(name)?.oid) snapshotForReview(name, entry);
remove(name);
}
// One that the pull request has but is not on disk was left out or deleted by the
// workflow. What is on disk now came from an earlier step.
const toWrite = changed.flatMap((name) => {
const entry = base.get(name);
const wanted = pr.has(name) ? removed.has(name) : true;
return entry && !isSubmodule(entry) && wanted && !onDisk(name)
? [{ name, entry }]
: [];
});
const blobs = readBlobs(toWrite.map(({ entry }) => entry));
for (const { name, entry } of toWrite) {
if (!hasRealParents(name)) continue;
const content = blobs.get(entry.oid);
if (!content) throw new Error("missing object");
mkdirSync(fsPath(posix.dirname(name)), { recursive: true });
if (isLink(entry)) symlinkSync(content, fsPath(name));
else {
const mode = entry.mode === "100755" ? 0o755 : 0o644;
writeFileSync(fsPath(name), content, { mode, flag: "wx" });
}
}
// Where the pull request has a file above a target, that is gone too.
return [
...new Set(
differing.flatMap(([target, names]) => [
target,
...names.filter((name) => target.startsWith(`${name}/`)),
]),
),
];
}
+40
View File
@@ -0,0 +1,40 @@
import { execFileSync } from "child_process";
/**
* Builds the `--depth` argument for a `git fetch`, unless the checkout still
* has its full history.
*
* `--depth` does not only cap what gets downloaded. Against a complete checkout
* (`actions/checkout` with `fetch-depth: 0`) it also truncates the history that
* is already there and marks the repository shallow, which drops the merge base
* with the base branch: `git log origin/<base>..HEAD` then quietly lists
* commits that are already merged, and `git diff origin/<base>...HEAD` fails
* with "no merge base". Those are the commands the prompt tells Claude to run
* to scope its work to the PR.
*
* A shallow checkout (the `fetch-depth: 1` default) has no history left to
* lose, so the limit still applies there and large repositories keep the fetch
* savings it was added for.
*/
export function fetchDepthArgs(depth: number): string[] {
return isShallowRepository() ? [`--depth=${depth}`] : [];
}
function isShallowRepository(): boolean {
try {
const output = execFileSync(
"git",
["rev-parse", "--is-shallow-repository"],
{
encoding: "utf8",
stdio: ["ignore", "pipe", "pipe"],
},
);
return output.trim() === "true";
} catch {
// No repository yet, or a git old enough not to know the flag. Treat the
// checkout as complete: fetching more than necessary is recoverable,
// truncating history is not.
return false;
}
}
+81 -12
View File
@@ -42,18 +42,74 @@ export async function configureGitAuth(
await $`git config user.email "${botId}+${botName}@${noreplyDomain}"`;
console.log(`✓ Set git user as ${botName}`);
await replaceCheckoutCredentials(githubToken, context);
console.log("Git authentication configured successfully");
}
/**
* Replace the credential that actions/checkout persisted in the working tree.
*
* actions/checkout stores its token as an `http.<server>/.extraheader` entry
* in .git/config for the duration of the job. Claude and the tools it invokes
* run inside this working tree, so remove that entry and back git with the
* action's own token instead (a credential helper when non-write users or bots
* are allowed, otherwise the origin URL). This applies to every mode, including API
* commit signing where no other git configuration is needed.
*
* actions/checkout < v6 stored the header directly in the repo-local config,
* where `git config --unset-all` removes it. Since v6.0.0 (backported to
* v5.0.1 and v4.3.1) the header is written to a separate file under
* RUNNER_TEMP that the repo config pulls in via `include.path`; `--unset-all`
* on the local config cannot touch an include-provided value, so the removal
* was a silent no-op and the checkout credential (typically the workflow
* GITHUB_TOKEN) stayed usable by git for the rest of the job. Clear the
* header from the local config AND from every included file so it can no
* longer authenticate while Claude runs.
*/
export async function replaceCheckoutCredentials(
githubToken: string,
context: GitHubContext,
) {
const serverUrl = new URL(GITHUB_SERVER_URL);
// Remove the authorization header that actions/checkout sets
console.log("Removing existing git authentication headers...");
const extraheaderKey = `http.${GITHUB_SERVER_URL}/.extraheader`;
let removedHeader = false;
try {
await $`git config --unset-all http.${GITHUB_SERVER_URL}/.extraheader`;
console.log("✓ Removed existing authentication headers");
} catch (e) {
console.log("No existing authentication headers to remove");
await $`git config --unset-all ${extraheaderKey}`;
removedHeader = true;
} catch {
// No extraheader in the local config (expected on the v6+ include layout).
}
try {
const includePaths =
await $`git config --local --get-all include.path`.text();
for (const includePath of includePaths.split("\n")) {
const path = includePath.trim();
if (!path) continue;
try {
await $`git config --file ${path} --unset-all ${extraheaderKey}`;
removedHeader = true;
} catch {
// This include does not define the header; leave it untouched.
}
}
} catch {
// No include.path entries in the local config.
}
console.log(
removedHeader
? "✓ Removed existing authentication headers"
: "No existing authentication headers to remove",
);
if (process.env.ALLOWED_NON_WRITE_USERS) {
if (process.env.ALLOWED_NON_WRITE_USERS || context.inputs.allowedBots) {
// When processing content from non-write users, use a credential helper
// instead of embedding the token in the remote URL. The helper script reads
// instead of embedding the token in the remote URL. Allowed bots count too:
// their access isn't checked, and they relay text that other people wrote.
// The helper script reads
// from GH_TOKEN at auth time, so .git/config stays token-free. Written as a
// file to avoid shell-escaping the helper body; placed under
// GITHUB_ACTION_PATH so it sits alongside the action source.
@@ -63,15 +119,30 @@ export async function configureGitAuth(
process.env.GITHUB_ACTION_PATH || homedir(),
".git-credential-gh-token",
);
// With no token it answers nothing, so that git goes on to a later step's own
// credentials. An empty password would count as an answer.
await writeFile(
helperPath,
'#!/bin/sh\necho username=x-access-token\necho password="$GH_TOKEN"\n',
'#!/bin/sh\n[ -n "$GH_TOKEN" ] || exit 0\necho username=x-access-token\necho password="$GH_TOKEN"\n',
{ mode: 0o700 },
);
const cleanUrl = `https://${serverUrl.host}/${context.repository.owner}/${context.repository.repo}.git`;
await $`git remote set-url origin ${cleanUrl}`;
await $`git config credential.helper ${helperPath}`;
// The username holds no secret. It keeps the URL out of rewrite rules on the
// server's prefix, as the token URL below is, and they would bypass the helper.
const helperUrl = `https://x-access-token@${serverUrl.host}/${context.repository.owner}/${context.repository.repo}.git`;
await $`git remote set-url origin ${helperUrl}`;
// git runs the helper through sh, which would strip Windows backslashes and split
// a path with a space.
const shellSafePath = helperPath
.replaceAll("\\", "/")
.replaceAll("'", "'\\''");
// Only for this server: git asks a helper about every host it contacts, and the
// helper answers without looking at which one.
const helperKey = `credential.https://${serverUrl.host}.helper`;
await $`git config ${helperKey} ${`!'${shellSafePath}'`}`;
console.log("✓ Configured credential helper");
console.log(
"Later steps that use the origin remote need their own credentials: GH_TOKEN, GIT_ASKPASS, a credential saved for the username x-access-token, or their own remote URL",
);
} else {
// Update the remote URL to include the token for authentication
console.log("Updating remote URL with authentication...");
@@ -79,8 +150,6 @@ export async function configureGitAuth(
await $`git remote set-url origin ${remoteUrl}`;
console.log("✓ Updated remote URL with authentication token");
}
console.log("Git authentication configured successfully");
}
/**
+282 -23
View File
@@ -1,13 +1,20 @@
import * as core from "@actions/core";
import { execFileSync } from "child_process";
import {
appendFileSync,
cpSync,
existsSync,
lstatSync,
mkdirSync,
readFileSync,
readlinkSync,
realpathSync,
rmSync,
statSync,
writeFileSync,
} from "fs";
import { dirname } from "path";
import { dirname, join, posix, relative, sep } from "path";
import { fetchDepthArgs } from "./fetch-depth";
// Paths that are both PR-controllable and read from cwd at CLI startup.
//
@@ -17,7 +24,7 @@ import { dirname } from "path";
// .gitconfig — git reads ~/.gitconfig and .git/config, never cwd/.gitconfig.
// .bashrc etc. — shells source these from $HOME; checkout cannot reach $HOME.
// .vscode/.idea— IDE config; nothing in the CLI's startup path reads them.
const SENSITIVE_PATHS = [
export const SENSITIVE_PATHS = [
".claude",
".mcp.json",
".claude.json",
@@ -25,12 +32,183 @@ const SENSITIVE_PATHS = [
".ripgreprc",
"CLAUDE.md",
"CLAUDE.local.md",
"AGENTS.md",
".husky",
];
const CLAUDE_PR_EXCLUDE_PATTERN = "/.claude-pr/";
function ensureClaudePrExcludedFromGit(): void {
function isSameOrInside(child: string, parent: string): boolean {
return child === parent || child.startsWith(`${parent}${sep}`);
}
// Repository paths (relative to cwd, `/`-separated) that a link may resolve
// to: `files` are tracked files whose working-tree content is unchanged from
// HEAD, `dirs` are directories that contain at least one tracked file.
type TrackedPaths = { files: Set<string>; dirs: Set<string> };
// Built from the superproject only: `git ls-files` reports a submodule as a
// single entry, so paths inside a checked-out submodule are in neither set and
// links into one are recorded as placeholders.
function listTrackedPaths(): TrackedPaths {
const gitPathList = (args: string[]) =>
execFileSync("git", args, { encoding: "utf8", maxBuffer: Infinity })
.split("\0")
.filter(Boolean);
const modified = new Set(
gitPathList([
"diff",
"--name-only",
"-z",
"--relative",
"--ignore-submodules",
"HEAD",
"--",
]),
);
const tracked: TrackedPaths = { files: new Set(), dirs: new Set() };
for (const file of gitPathList(["ls-files", "-z"])) {
if (!modified.has(file)) {
tracked.files.add(file);
}
for (
let dir = posix.dirname(file);
dir !== "." && !tracked.dirs.has(dir);
dir = posix.dirname(dir)
) {
tracked.dirs.add(dir);
}
}
return tracked;
}
// The snapshot is scoped to tracked repository content and never contains
// links. An entry is copied with its content only when all of these hold:
// 1. its real target (through any links) lies inside the working tree;
// 2. no component of the target's path inside the tree is `.git`, and the
// target is not inside the snapshot directory itself;
// 3. the target does not contain a directory already on the entry's own
// path (which would recurse);
// 4. if the entry is reached through a link (it is one, or a directory above
// it inside the sensitive path is), a file target must be tracked in the
// checkout with its content unchanged from HEAD, and a directory target
// must contain at least one tracked file (see listTrackedPaths). Directory
// targets that pass are descended into and their children are checked
// individually.
// Files and directories at their literal, non-linked location are unaffected
// by rule 4 and are copied as-is. Every other entry — targets outside the
// tree, dangling or looping links, git metadata, submodule contents, untracked
// or locally modified files reached through a link — is recorded as a
// placeholder file (see recordPlaceholder), so nothing in the snapshot
// resolves anywhere else.
function shouldSnapshotContent(
entryPath: string,
workTreeRealPath: string,
tracked: TrackedPaths,
): boolean {
try {
const targetRealPath = realpathSync(entryPath);
if (!isSameOrInside(targetRealPath, workTreeRealPath)) {
return false;
}
const targetParts = relative(workTreeRealPath, targetRealPath).split(sep);
if (targetParts.includes(".git") || targetParts[0] === ".claude-pr") {
return false;
}
for (let dir = dirname(entryPath); ; dir = dirname(dir)) {
if (isSameOrInside(realpathSync(dir), targetRealPath)) {
return false;
}
if (dir === dirname(dir)) {
break;
}
}
const literalPath = join(
workTreeRealPath,
relative(process.cwd(), entryPath),
);
if (targetRealPath === literalPath) {
return true;
}
const targetRepoPath = targetParts.join("/");
return statSync(targetRealPath).isDirectory()
? tracked.dirs.has(targetRepoPath)
: tracked.files.has(targetRepoPath);
} catch {
return false;
}
}
// Writes a short regular file at `dest` describing the entry that was left
// out, so the snapshot records that something was there without linking to it.
function recordPlaceholder(src: string, dest: string): void {
console.warn(
`Snapshot: ${src} not included in snapshot; recording a placeholder`,
);
let description = "is not included in this snapshot";
try {
description = `was a symbolic link to ${JSON.stringify(readlinkSync(src))}; the link target is not included in this snapshot`;
} catch {
// Not a link (or no longer present).
}
mkdirSync(dirname(dest), { recursive: true });
writeFileSync(dest, `Snapshot placeholder: ${src} ${description}.\n`);
}
/**
* Copies a sensitive path into the review snapshot. Entries that pass the
* check above are copied dereferenced (reviewers see the effective content);
* every other entry is recorded as a placeholder file, never as a link.
* Applies per entry, including links nested inside a real directory.
*/
function snapshotSensitivePath(
src: string,
dest: string,
workTreeRealPath: string,
tracked: TrackedPaths,
): void {
const excluded: Array<{ src: string; dest: string }> = [];
const keepOrExclude =
(keep: (entry: string) => boolean) =>
(entrySrc: string, entryDest: string) => {
if (keep(entrySrc)) {
return true;
}
excluded.push({ src: entrySrc, dest: entryDest });
return false;
};
try {
cpSync(src, dest, {
recursive: true,
dereference: true,
filter: keepOrExclude((entry) =>
shouldSnapshotContent(entry, workTreeRealPath, tracked),
),
});
} catch (error) {
// Dangling links are normally caught by the filter above. If a target
// disappears between that check and the copy, the dereferencing copy
// throws ENOENT; start over without following links, recording every link
// as a placeholder, instead of failing the restore.
if (
!(error instanceof Error && "code" in error && error.code === "ENOENT")
) {
throw error;
}
rmSync(dest, { recursive: true, force: true });
excluded.length = 0;
cpSync(src, dest, {
recursive: true,
filter: keepOrExclude((entry) => !lstatSync(entry).isSymbolicLink()),
});
}
for (const entry of excluded) {
recordPlaceholder(entry.src, entry.dest);
}
}
export function ensureClaudePrExcludedFromGit(): void {
const excludePath = execFileSync(
"git",
["rev-parse", "--git-path", "info/exclude"],
@@ -52,8 +230,70 @@ function ensureClaudePrExcludedFromGit(): void {
appendFileSync(excludePath, `${prefix}${CLAUDE_PR_EXCLUDE_PATTERN}\n`);
}
function headHistoryLength(): number {
const head = execFileSync("git", ["rev-parse", "--verify", "HEAD^{commit}"], {
encoding: "utf8",
}).trim();
return Number(
execFileSync("git", ["rev-list", "--count", head, "--"], {
encoding: "utf8",
}),
);
}
/**
* Restores security-sensitive config paths from the PR base branch.
* Fetching a branch into a shallow checkout cuts HEAD's history at that branch's tip if
* it is in there. Returns a function that fetches HEAD's history again if it was cut
* since. No path from HEAD is longer than its number of commits.
*/
export function keepHeadHistory(): () => void {
const before = headHistoryLength();
return () => {
if (headHistoryLength() >= before) return;
try {
const head = execFileSync("git", ["rev-parse", "HEAD"], {
encoding: "utf8",
}).trim();
fetchRef(head, before);
} catch {
// For example, a commit that an earlier step made isn't on the remote.
core.warning(
"Could not fetch the checked-out history again, so comparing with the base branch can fail.",
);
}
};
}
function fetchRef(ref: string, depth: number): void {
// --no-recurse-submodules: a .gitmodules file must not start a submodule fetch,
// whatever fetch.recurseSubmodules says.
execFileSync(
"git",
[
"fetch",
"origin",
ref,
...fetchDepthArgs(depth),
// Following tags would cut a shallow history at each tag it fetches.
"--no-tags",
"--no-recurse-submodules",
],
{
stdio: "inherit",
env: process.env,
},
);
}
export function fetchBranch(branch: string): void {
// In full, so that a tag of the same name isn't fetched instead and the
// remote-tracking branch is updated whatever the checkout's refspec covers.
fetchRef(`+refs/heads/${branch}:refs/remotes/origin/${branch}`, 1);
}
/**
* Restores security-sensitive config paths from the PR base branch, or from the
* branch that chooseConfigBranch picks in its place.
*
* The CLI's non-interactive mode trusts cwd: it reads `.mcp.json`,
* `.claude/settings.json`, and `.claude/settings.local.json` from the working
@@ -71,10 +311,27 @@ function ensureClaudePrExcludedFromGit(): void {
* commits with `git add -A`, the revert will be included in that commit. This
* is a narrow UX tradeoff for closing the RCE surface.
*
* @param baseBranch - PR base branch name. Must be pre-validated (branch.ts
* calls validateBranchName on it before returning).
* Only the paths listed in SENSITIVE_PATHS come from the base branch; the rest
* of the working tree stays at the PR head. A base-branch hook or setting that
* calls out through files a PR can change — package-manager scripts
* (`bun run`, `npm run`, `yarn`, `pnpm run`), Makefile or task-runner targets,
* repo-relative script paths, or tools that load executable project config —
* therefore runs whatever the PR head provides. Keep restored hooks
* self-contained: invoke the tool binary directly, pin its version, and pass
* config on the command line rather than reading it from the checkout. This
* extends to the runtime itself: `bunx <tool>` runs the tool under `node` when
* `node` is on PATH, but on a Bun-only runner Bun executes the script and reads
* `bunfig.toml` (e.g. `preload`) from the checkout, so `bunfig.toml` and
* `.npmrc` there are PR-controlled runtime config too. On PRs from authors without
* write access, restoreToolConfigFromBase replaces those.
*
* @param baseBranch - The branch to restore from, see chooseConfigBranch. Must be
* pre-validated with validateBranchName.
* @returns The paths whose working-tree state now comes from the base branch
* rather than the PR. Callers that stage files must exclude these, or they
* will commit the revert back onto the PR author's branch.
*/
export function restoreConfigFromBase(baseBranch: string): void {
export function restoreConfigFromBase(baseBranch: string): string[] {
console.log(
`Restoring ${SENSITIVE_PATHS.join(", ")} from origin/${baseBranch} (PR head is untrusted)`,
);
@@ -82,11 +339,15 @@ export function restoreConfigFromBase(baseBranch: string): void {
// Snapshot every PR-authored sensitive path into .claude-pr/ before deletion
// so review agents can inspect what the PR changes without those files ever
// being executed. Captured before the security delete so it reflects the
// PR-authored version.
// PR-authored version. Links are followed only to tracked, unmodified content
// inside the working tree; anything else is recorded as a placeholder file,
// so the snapshot itself never contains links.
rmSync(".claude-pr", { recursive: true, force: true });
const workTreeRealPath = realpathSync(process.cwd());
const tracked = listTrackedPaths();
for (const p of SENSITIVE_PATHS) {
if (existsSync(p)) {
cpSync(p, `.claude-pr/${p}`, { recursive: true, dereference: true });
if (lstatSync(p, { throwIfNoEntry: false })) {
snapshotSensitivePath(p, `.claude-pr/${p}`, workTreeRealPath, tracked);
}
}
if (existsSync(".claude-pr")) {
@@ -109,22 +370,15 @@ export function restoreConfigFromBase(baseBranch: string): void {
rmSync(p, { recursive: true, force: true });
}
// --no-recurse-submodules: explicitly suppress submodule fetching regardless of
// fetch.recurseSubmodules config. Defense-in-depth alongside the delete above.
execFileSync(
"git",
["fetch", "origin", baseBranch, "--depth=1", "--no-recurse-submodules"],
{
stdio: "inherit",
env: process.env,
},
);
fetchBranch(baseBranch);
for (const p of SENSITIVE_PATHS) {
try {
execFileSync("git", ["checkout", `origin/${baseBranch}`, "--", p], {
stdio: "pipe",
});
execFileSync(
"git",
["checkout", `refs/remotes/origin/${baseBranch}`, "--", p],
{ stdio: "pipe" },
);
} catch {
// Path doesn't exist on base — it stays deleted.
}
@@ -139,4 +393,9 @@ export function restoreConfigFromBase(baseBranch: string): void {
} catch {
// Nothing was staged, or paths don't exist on HEAD — either is fine.
}
// Every sensitive path is reported, not just the ones that changed: the
// restore also deletes paths the PR added that are absent on base, and those
// deletions are stageable too.
return [...SENSITIVE_PATHS];
}
+413
View File
@@ -0,0 +1,413 @@
import * as core from "@actions/core";
import { execFileSync } from "child_process";
import {
closeSync,
constants,
fstatSync,
lstatSync,
mkdirSync,
openSync,
readFileSync,
readlinkSync,
realpathSync,
rmSync,
writeFileSync,
} from "fs";
import { homedir } from "os";
import { dirname, isAbsolute, join, resolve } from "path";
// A session may write to one that keeps its writes, but not put a link in its place.
type TrackedFile = { label: string; path: string; keepsWrites?: boolean };
// "opaque" is anything that can't be saved and compared: a directory, a FIFO, a
// device, a file too large to hold, or one that can't be read.
type FileSnapshot = TrackedFile & {
state: "absent" | "file" | "opaque";
wasSymlink: boolean;
content?: Buffer;
mode?: number;
names?: number;
// For a symlink: what it points to, and the regular file that ends up at.
link?: { target: string; resolved: string };
};
const MAX_SNAPSHOT_BYTES = 64 * 1024 * 1024;
const SESSION_STATE_FILES = [
".push-branch",
".no-hooks",
".git-repo",
".restored-paths",
];
const NO_IMPLICIT_BARE_REPOSITORY: [string, string] = [
"safe.bareRepository",
"explicit",
];
// How many GIT_CONFIG_* entries were there before disableGitPrograms added any.
let entriesBeforeAction: number | undefined;
const codeOf = (error: unknown): string =>
(error as { code?: string }).code ?? "unknown error";
function gitProgramSettings({ hooks }: { hooks: boolean }): [string, string][] {
return hooks
? [
["core.fsmonitor", "false"],
["core.hooksPath", "/dev/null"],
]
: [["core.fsmonitor", "false"]];
}
function resolveGitPath(file: string): string | null {
try {
const path = execFileSync("git", ["rev-parse", "--git-path", file], {
encoding: "utf8",
stdio: ["ignore", "pipe", "ignore"],
env: { ...process.env },
}).trim();
return isAbsolute(path) ? path : join(process.cwd(), path);
} catch {
return null;
}
}
/**
* Every config file git reads here, which includes those pulled in through
* include.path. actions/checkout keeps its config in one of those.
*/
function gitConfigFilesInUse(): TrackedFile[] {
try {
const listing = execFileSync(
"git",
["config", "--list", "--includes", "--show-origin", "--name-only", "-z"],
{
encoding: "utf8",
stdio: ["ignore", "pipe", "ignore"],
env: { ...process.env },
},
);
return listing
.split("\0")
.filter((token, index) => index % 2 === 0 && token.startsWith("file:"))
.map((origin) => ({
label: "a git config file",
path: origin.slice("file:".length),
}));
} catch {
return [];
}
}
/**
* Files that later steps trust and that a session can change. The runner applies
* GITHUB_ENV and GITHUB_PATH to every later step, post steps included, and those
* steps run git, which reads these config files and can run programs they name.
*/
function filesLaterStepsTrust(): TrackedFile[] {
const home = process.env.HOME || homedir();
const candidates: [string, string | null | undefined][] = [
["GITHUB_ENV", process.env.GITHUB_ENV],
["GITHUB_PATH", process.env.GITHUB_PATH],
[".git/config", resolveGitPath("config")],
[".git/config.worktree", resolveGitPath("config.worktree")],
["~/.gitconfig", join(home, ".gitconfig")],
["GIT_CONFIG_GLOBAL", process.env.GIT_CONFIG_GLOBAL],
[
"git config in XDG_CONFIG_HOME",
join(process.env.XDG_CONFIG_HOME || join(home, ".config"), "git/config"),
],
];
const named = candidates.flatMap(([label, path]) =>
path ? [{ label, path }] : [],
);
// The action adds to these after the session, so a link there would redirect that to
// a file later steps trust.
const written = ["GITHUB_OUTPUT", "GITHUB_STEP_SUMMARY"].flatMap((label) => {
const path = process.env[label];
return path ? [{ label, path, keepsWrites: true }] : [];
});
const seen = new Set<string>();
return [...named, ...gitConfigFilesInUse(), ...written].filter(({ path }) => {
const absolute = resolve(path);
if (seen.has(absolute)) return false;
seen.add(absolute);
return true;
});
}
function readSnapshot(file: TrackedFile): FileSnapshot {
let wasSymlink: boolean;
try {
wasSymlink = lstatSync(file.path).isSymbolicLink();
} catch {
return { ...file, state: "absent", wasSymlink: false };
}
let fd: number | undefined;
try {
// Non-blocking, because opening a FIFO would otherwise wait for a writer.
fd = openSync(file.path, constants.O_RDONLY | (constants.O_NONBLOCK ?? 0));
const stats = fstatSync(fd);
if (!stats.isFile() || stats.size > MAX_SNAPSHOT_BYTES) {
return { ...file, state: "opaque", wasSymlink };
}
return {
...file,
state: "file",
wasSymlink,
content: readFileSync(fd),
mode: stats.mode & 0o777,
names: stats.nlink,
...(wasSymlink && {
link: {
target: readlinkSync(file.path),
resolved: realpathSync(file.path),
},
}),
};
} catch {
return { ...file, state: "opaque", wasSymlink };
} finally {
if (fd !== undefined) closeSync(fd);
}
}
function isUnchanged(before: FileSnapshot, now: FileSnapshot): boolean {
if (before.state !== now.state || before.wasSymlink !== now.wasSymlink) {
return false;
}
return before.state !== "file" || before.content!.equals(now.content!);
}
/**
* Where to restore a file. A symlink that was there before the session and still
* leads to the same file is left alone, and that file is restored, so a runner's
* own link keeps working. One a session added or re-pointed gets replaced.
*/
function restorePath(before: FileSnapshot): string {
if (!before.link) return before.path;
try {
return readlinkSync(before.path) === before.link.target &&
realpathSync(before.path) === before.link.resolved
? before.link.resolved
: before.path;
} catch {
return before.path;
}
}
/**
* Whether what is written to a file now ends up in that file and no other. A new name
* for a file that later steps trust is as good as a symlink to it.
*/
function isItsOwnFile(before: FileSnapshot): boolean {
try {
const stats = lstatSync(restorePath(before));
return stats.isFile() && stats.nlink <= (before.names ?? 1);
} catch {
return false;
}
}
/**
* Records the files later steps trust, and returns a function that puts them
* back. Call the returned function as soon as the session ends and before the
* action runs git again. Replacing a file rather than writing to it means a
* session cannot redirect the write by swapping in a symlink.
*
* This stops a session that can only write through its tools. It is not a
* sandbox: code running as the runner's user can stop the action before it
* restores anything.
*/
export function snapshotSessionFiles(beforeSession?: () => void): () => void {
const files = filesLaterStepsTrust();
const snapshots = files.map(readSnapshot);
// What the action changes for the session goes back too, without a warning.
beforeSession?.();
const atStart = files.map(readSnapshot);
return () => {
const restored: string[] = [];
for (const [index, before] of snapshots.entries()) {
try {
const now = readSnapshot(before);
if (before.keepsWrites && isItsOwnFile(before)) continue;
// The file it leads to can have the same content so far.
const leadsElsewhere = before.keepsWrites && now.state === "file";
if (!leadsElsewhere && isUnchanged(before, now)) continue;
const changedBySession =
leadsElsewhere || !isUnchanged(atStart[index]!, now);
if (before.state === "opaque") {
core.warning(
`Could not restore ${before.label} after the session: it was not a regular file before`,
);
continue;
}
const path = restorePath(before);
rmSync(path, { force: true, recursive: true });
if (before.state === "file") {
writeFileSync(path, before.content!, { mode: before.mode });
}
if (changedBySession) restored.push(before.label);
} catch (error) {
// The label is fixed text, and the error code carries no file content.
core.warning(
`Could not restore ${before.label} after the session (${codeOf(error)})`,
);
}
}
if (restored.length > 0) {
core.warning(
`The session changed files that later steps trust, so they were restored: ${restored.join(", ")}`,
);
}
};
}
/**
* Makes git ignore settings that run programs, for this process and everything
* it starts. GIT_CONFIG_* outranks every config file, and existing entries are
* kept. It is not written to GITHUB_ENV, so later steps are unaffected.
*/
export function disableGitPrograms({ hooks }: { hooks: boolean }): void {
let count = Number.parseInt(process.env.GIT_CONFIG_COUNT ?? "", 10) || 0;
entriesBeforeAction ??= count;
const current = new Map<string, string | undefined>();
for (let i = 0; i < count; i++) {
current.set(
process.env[`GIT_CONFIG_KEY_${i}`] ?? "",
process.env[`GIT_CONFIG_VALUE_${i}`],
);
}
const settings = gitProgramSettings({ hooks });
// A workflow may run the action in a bare repository, where this setting would stop git.
const setting = NO_IMPLICIT_BARE_REPOSITORY.join("=");
const stopsGit =
gitOutput("rev-parse", "--git-dir") !== null &&
gitOutput("-c", setting, "rev-parse", "--git-dir") === null;
if (hooks || !stopsGit) settings.push(NO_IMPLICIT_BARE_REPOSITORY);
for (const [key, value] of settings) {
if (current.get(key) === value) continue;
process.env[`GIT_CONFIG_KEY_${count}`] = key;
process.env[`GIT_CONFIG_VALUE_${count}`] = value;
count++;
}
process.env.GIT_CONFIG_COUNT = String(count);
}
const sessionFile = (name: string): string | undefined =>
process.env.GITHUB_ACTION_PATH
? join(process.env.GITHUB_ACTION_PATH, "scripts", name)
: undefined;
function writeSessionFile(name: string, content: string): void {
const path = sessionFile(name);
if (!path) return;
mkdirSync(dirname(path), { recursive: true });
rmSync(path, { force: true, recursive: true });
writeFileSync(path, content, { mode: 0o444 });
}
/** Records the only branch the push wrapper may update. */
export function pinPushBranch(branch: string): void {
try {
writeSessionFile(".push-branch", branch);
} catch (error) {
// With no file and no variable, the wrapper refuses to push.
core.warning(`Could not record the session's branch (${codeOf(error)})`);
}
}
/**
* Records the paths that the commit wrapper leaves out, because the action took them
* from a branch other than the pull request's base.
*/
export function pinRestoredPaths(paths: string[]): void {
const pathspecs = paths.map((path) =>
// Its commands are replaced in subdirectories too.
path === "package.json"
? ":(top,glob)**/package.json"
: `:(top,literal)${path}`,
);
try {
writeSessionFile(".restored-paths", `${pathspecs.join("\n")}\n`);
} catch (error) {
core.warning(
`Could not record the restored files, so a commit can include them (${codeOf(error)})`,
);
}
}
/** Removes what the wrappers read, so another step in this job can't inherit it. */
export function clearSessionState(): void {
for (const name of SESSION_STATE_FILES) {
const path = sessionFile(name);
if (!path) continue;
try {
rmSync(path, { force: true, recursive: true });
} catch (error) {
core.warning(`Could not remove ${name} (${codeOf(error)})`);
}
}
}
const gitOutput = (...args: string[]): string | null => {
try {
return execFileSync("git", args, {
encoding: "utf8",
stdio: ["ignore", "pipe", "ignore"],
env: { ...process.env },
}).trim();
} catch {
return null;
}
};
/**
* Moves the settings from disableGitPrograms out of the environment for the session.
* Repository config and the wrappers keep them in force instead, and
* snapshotSessionFiles puts the config back afterwards. Elsewhere the concern is a
* repository git finds in a plain folder, with its own config. Git ignores
* safe.bareRepository in repository config, so the session gets it from global config.
*/
export function prepareGitForSession({ hooks }: { hooks: boolean }): void {
const count = Number.parseInt(process.env.GIT_CONFIG_COUNT ?? "", 10) || 0;
const before = entriesBeforeAction ?? count;
for (let i = before; i < count; i++) {
delete process.env[`GIT_CONFIG_KEY_${i}`];
delete process.env[`GIT_CONFIG_VALUE_${i}`];
}
if (before > 0) process.env.GIT_CONFIG_COUNT = String(before);
else delete process.env.GIT_CONFIG_COUNT;
entriesBeforeAction = undefined;
const setConfig = (scope: string, entries: [string, string][]): void => {
for (const [key, value] of entries) {
if (gitOutput("config", scope, "--replace-all", key, value) === null) {
core.warning(`Could not set ${key} for the session`);
}
}
};
// Before the global setting, which stops git from finding a bare repository here.
const gitDir = gitOutput("rev-parse", "--absolute-git-dir");
const workTree = gitDir && gitOutput("rev-parse", "--show-toplevel");
if (gitDir) setConfig("--local", gitProgramSettings({ hooks }));
setConfig("--global", [NO_IMPLICIT_BARE_REPOSITORY]);
try {
if (hooks) writeSessionFile(".no-hooks", "");
if (gitDir && workTree) {
const home = process.env.HOME || homedir();
writeSessionFile(".git-repo", `${gitDir}\n${workTree}\n${home}\n`);
}
} catch (error) {
core.warning(
`Could not record git settings for the wrappers (${codeOf(error)})`,
);
}
}
+422
View File
@@ -0,0 +1,422 @@
import * as core from "@actions/core";
import { execFileSync } from "child_process";
import {
lstatSync,
mkdirSync,
readdirSync,
readFileSync,
rmdirSync,
rmSync,
symlinkSync,
writeFileSync,
} from "fs";
import { dirname } from "path";
import {
ensureClaudePrExcludedFromGit,
SENSITIVE_PATHS,
} from "./restore-config";
// What npm and Bun read from the repository root to decide which program to run and
// where to download it from. Config from the base branch can start them, for example
// in a hook or an MCP server. node_modules counts because its .bin comes first on PATH,
// and package.json because npx looks for commands there first.
const TOOL_CONFIG = [".npmrc", "bunfig.toml", "node_modules"];
const PACKAGE_JSON = "package.json";
// The fields that provide commands. A bin that is a string is named after the package,
// and npx looks in workspaces too.
const COMMAND_FIELDS = ["name", "bin", "directories", "workspaces"];
const NO_COMMANDS = "none";
const ENCODING = "working-tree-encoding";
const MAX_OUTPUT_BYTES = 1024 ** 3;
const MAX_PACKAGE_JSON_BYTES = 4 * 1024 * 1024;
export type Entry = { mode: string; oid: string };
type Replacement = { path: string; entry: Entry; content?: Buffer };
// File names can be any bytes. Latin-1 keeps each one intact.
export const fsPath = (path: string): Buffer => Buffer.from(path, "latin1");
export const isFile = (entry: Entry): boolean => entry.mode.startsWith("100");
export const isLink = (entry: Entry): boolean => entry.mode === "120000";
export const isSubmodule = (entry: Entry): boolean => entry.mode === "160000";
// No file name or content, which the pull request controls, reaches the log.
function warn(what: string, error: unknown): void {
const { code, status } = error as { code?: string; status?: number };
core.warning(`${what} (${code ?? `exit code ${status ?? "unknown"}`})`);
}
export function git(args: string[], input: string | Buffer = ""): Buffer {
return execFileSync("git", args, {
stdio: ["pipe", "pipe", "ignore"],
input,
env: { ...process.env },
maxBuffer: MAX_OUTPUT_BYTES,
});
}
// A parent that is a file or a looping link means there is nothing there.
export function onDisk(path: string) {
try {
return lstatSync(fsPath(path));
} catch {
return undefined;
}
}
// File systems can ignore case, and some the way a character is encoded.
export const fold = (path: string): string =>
fsPath(path).toString().normalize("NFKC").toLowerCase();
/**
* Whether a path is in a directory that comes from the base branch already. The index
* has the pull request's version of it, so nothing there may be rebuilt from that.
*/
function isRestored(path: string, topLevel: Set<string>): boolean {
const [dir] = path.split("/") as [string];
return SENSITIVE_PATHS.some(
(restored) =>
dir === restored ||
// Another spelling is a directory of its own only if the file system lists it.
(fold(dir) === fold(restored) && !topLevel.has(dir)),
);
}
/** Which tool config a path belongs to. */
function rootName(path: string): string | undefined {
const folded = fold(path);
// A directory with one of these names is in the file's place, whatever it holds.
const name = TOOL_CONFIG.find(
(name) => folded === name || folded.startsWith(`${name}/`),
);
if (name) return name;
return folded === PACKAGE_JSON || folded.endsWith(`/${PACKAGE_JSON}`)
? PACKAGE_JSON
: undefined;
}
/** Whether no parent of a path is a link, which could lead out of the checkout. */
export function hasRealParents(path: string): boolean {
for (let dir = dirname(path); dir !== "."; dir = dirname(dir)) {
if (onDisk(dir)?.isSymbolicLink()) return false;
}
return true;
}
export function remove(path: string): void {
rmSync(fsPath(path), { recursive: true, force: true });
// An empty directory would stand in the way of a file of that name.
try {
for (let dir = dirname(path); dir !== "."; dir = dirname(dir)) {
rmdirSync(fsPath(dir));
}
} catch {
// It isn't empty.
}
}
// Each entry is "<mode> <other fields>\t<path>", relative to the working directory.
function parseEntries(listing: Buffer, oidField: number): Map<string, Entry> {
const entries = new Map<string, Entry>();
const topLevel = new Set(
readdirSync(".", { encoding: "buffer" }).map((name) =>
name.toString("latin1"),
),
);
for (const line of listing.toString("latin1").split("\0")) {
const [info, path] = line.split(/\t(.*)/s);
if (!info || !path || !rootName(path) || isRestored(path, topLevel)) {
continue;
}
const fields = info.split(" ");
entries.set(path, { mode: fields[0]!, oid: fields[oidField]! });
}
return entries;
}
const pullRequestEntries = (): Map<string, Entry> =>
parseEntries(git(["ls-files", "--stage", "-z"]), 1);
// A checkout with a narrow refspec has no remote-tracking branch, only FETCH_HEAD.
function baseEntries(baseBranch: string): Map<string, Entry> {
let ref = `refs/remotes/origin/${baseBranch}`;
try {
git(["rev-parse", "--verify", "--quiet", `${ref}^{commit}`]);
} catch {
ref = "FETCH_HEAD";
}
return parseEntries(git(["ls-tree", "-r", "-z", ref]), 2);
}
/** The stored content of each entry, which no .gitattributes has been applied to. */
export function readBlobs(entries: Entry[]): Map<string, Buffer> {
const blobs = new Map<string, Buffer>();
const oids = [...new Set(entries.map(({ oid }) => oid))];
if (oids.length === 0) return blobs;
const output = git(["cat-file", "--batch"], `${oids.join("\n")}\n`);
// "<oid> <type> <size>\n<content>\n", or "<oid> missing\n"
for (let at = 0; at < output.length; ) {
const end = output.indexOf("\n", at);
const [oid, , size] = output.toString("latin1", at, end).split(" ");
at = end + 1;
if (size === undefined) continue;
blobs.set(oid!, output.subarray(at, at + Number(size)));
at += Number(size) + 1;
}
return blobs;
}
type Json = Record<string, unknown>;
/** Parses a package.json, or returns undefined if it can't be understood. */
function parsePackageJson(size: number, read: () => Buffer): Json | undefined {
if (size > MAX_PACKAGE_JSON_BYTES) return undefined;
// trimStart also drops a byte order mark.
const json = JSON.parse(read().toString().trimStart());
return json && typeof json === "object" ? json : undefined;
}
/** The package.json that git stores. */
function readPackageJson(entry: Entry): Json | undefined {
if (!isFile(entry)) return undefined;
try {
const size = Number(git(["cat-file", "-s", entry.oid]).toString());
return parsePackageJson(size, () => git(["cat-file", "blob", entry.oid]));
} catch {
return undefined;
}
}
/** The package.json on disk, which .gitattributes can make differ from what git stores. */
function readPackageJsonOnDisk(path: string): Json | undefined {
try {
const { size } = lstatSync(fsPath(path));
return parsePackageJson(size, () => readFileSync(fsPath(path)));
} catch {
return undefined;
}
}
/** The commands a package.json provides, or undefined if that can't be told. */
function commands(json: Json | undefined): string | undefined {
if (!json) return undefined;
// Other directories, such as the one for tests, provide no commands.
const directories = (json.directories as Json | null | undefined)?.bin;
if (!json.bin && !directories && !json.workspaces) return NO_COMMANDS;
const provided: Json = { ...json, directories };
return JSON.stringify(COMMAND_FIELDS.map((field) => provided[field] ?? null));
}
/**
* The attributes that make git write each file differently from what it stores. The
* pull request controls .gitattributes.
*/
function conversions(paths: string[]): Map<string, Set<string>> {
const found = new Map<string, Set<string>>();
if (paths.length === 0) return found;
// Checkout uses the index's .gitattributes, and the one on disk can be converted too.
// If the index has none, it uses the one on disk.
for (const source of [["--cached"], []]) {
const fields = git(
["check-attr", ...source, "-z", "--stdin", ENCODING, "ident", "filter"],
Buffer.from(paths.map((path) => `${path}\0`).join(""), "latin1"),
)
.toString("latin1")
.split("\0");
// "<path> <attribute> <value>", each ended by a NUL
for (let i = 0; i + 2 < fields.length; i += 3) {
const [path, attribute, value] = fields.slice(i) as [
string,
string,
string,
];
if (value === "unspecified" || value === "unset") continue;
// Git takes this as no conversion.
if (attribute === ENCODING && /^utf-?8$/i.test(value)) continue;
found.set(path, (found.get(path) ?? new Set()).add(attribute));
}
}
return found;
}
function differs(
path: string,
converted: Set<string> | undefined,
pr?: Entry,
base?: Entry,
): boolean {
const isSame = pr?.oid === base?.oid && pr?.mode === base?.mode;
// Even an unchanged file can then be one that npm or Bun can't read.
if (converted?.has(ENCODING)) return true;
if (rootName(path) !== PACKAGE_JSON) return !isSame;
if (!pr || isSame) return false;
const provided = [commands(readPackageJson(pr))];
// npm reads what is on disk, unless an earlier step removed it.
if (converted && hasRealParents(path) && onDisk(path)?.isFile()) {
provided.push(commands(readPackageJsonOnDisk(path)));
}
// Only one that provides commands matters, so most pull requests keep their own.
const allowed = [
NO_COMMANDS,
base ? commands(readPackageJson(base)) : NO_COMMANDS,
];
return provided.some((value) => !value || !allowed.includes(value));
}
/** The pull request's package.json with the base branch's commands. */
function withBaseCommands(
pr: Entry,
base: Entry | undefined,
): Buffer | undefined {
const json = readPackageJson(pr);
const baseJson = base ? readPackageJson(base) : {};
if (!json || !baseJson) return undefined;
for (const field of COMMAND_FIELDS) {
// Without a bin, a new package's name provides no command.
if (field === "name" && !base) continue;
if (baseJson[field] === undefined) delete json[field];
else json[field] = baseJson[field];
}
return Buffer.from(`${JSON.stringify(json, null, 2)}\n`);
}
/** Keeps the pull request's version readable, as for the other restored config. */
export function snapshotForReview(path: string, entry: Entry): void {
if (!isFile(entry) || rootName(path) === "node_modules") return;
try {
const content = git(["cat-file", "blob", entry.oid]);
const copy = fsPath(`.claude-pr/${path}`);
mkdirSync(fsPath(`.claude-pr/${dirname(path)}`), { recursive: true });
writeFileSync(copy, content);
ensureClaudePrExcludedFromGit();
} catch (error) {
warn("Could not keep a copy of tool config for review", error);
}
}
/**
* Replaces the pull request's tool config with the base branch's wherever they differ.
* Like restoreConfigFromBase, which must run first because it fetches the base branch,
* it changes the working tree and not the index. Only tracked files are replaced, so
* what an earlier step created stays. In a package.json only the commands are replaced.
*
* @returns The names that were replaced, from a fixed list. Callers that stage files
* must exclude them.
* @throws If the base branch's version can't be put in place. The message holds no
* file name, because the pull request controls those.
*/
export function restoreToolConfigFromBase(baseBranch: string): string[] {
try {
return restoreToolConfig(baseBranch);
} catch (error) {
warn("Could not replace tool config", error);
throw new Error(
"Could not use the trusted branch's version of the files that npm and Bun read",
);
}
}
function restoreToolConfig(baseBranch: string): string[] {
const pr = pullRequestEntries();
const base = baseEntries(baseBranch);
const all = [...new Set([...pr.keys(), ...base.keys()])];
// npx reads the packages in subdirectories only through the root's workspaces, and
// a pull request's own are replaced below.
const rootJson = base.get(PACKAGE_JSON);
const hasWorkspaces =
rootJson !== undefined &&
readPackageJson(rootJson)?.workspaces !== undefined;
const matters = (path: string): boolean =>
hasWorkspaces || rootName(path) !== PACKAGE_JSON || !path.includes("/");
const converted = conversions(
[...pr].flatMap(([path, entry]) =>
isFile(entry) && matters(path) && rootName(path) !== "node_modules"
? [path]
: [],
),
);
// Names that differ only in case can be the same file.
const differing = new Set(
all
.filter(
(path) =>
matters(path) &&
differs(path, converted.get(path), pr.get(path), base.get(path)),
)
.map(fold),
);
const candidates = all.filter((path) => differing.has(fold(path)));
// A directory where the pull request has a file or link came from an earlier step.
const replaced = candidates.filter((path) => {
const entry = pr.get(path);
const stats = entry && onDisk(path);
return (
stats &&
(!stats.isDirectory() || isSubmodule(entry)) &&
hasRealParents(path)
);
});
// All of these go before anything is written.
for (const path of replaced) {
snapshotForReview(path, pr.get(path)!);
remove(path);
}
const removed = new Set(replaced);
const toWrite = candidates.flatMap((path): Replacement[] => {
const entry = base.get(path);
const prEntry = pr.get(path);
// One that the pull request has but is not on disk was left out or deleted by the
// workflow. One that only the base has matters unless it's a package.json.
const wanted = prEntry
? removed.has(path)
: rootName(path) !== PACKAGE_JSON;
if (!wanted || onDisk(path)) return [];
const content =
prEntry && rootName(path) === PACKAGE_JSON
? withBaseCommands(prEntry, entry)
: undefined;
if (content) return [{ path, entry: prEntry!, content }];
return entry && !isSubmodule(entry) ? [{ path, entry, content }] : [];
});
const blobs = readBlobs(
toWrite.flatMap(({ entry, content }) => (content ? [] : [entry])),
);
const written: string[] = [];
for (const { path, entry, content = blobs.get(entry.oid) } of toWrite) {
if (!hasRealParents(path)) continue;
if (!content) throw new Error("missing object");
mkdirSync(fsPath(dirname(path)), { recursive: true });
if (isLink(entry)) symlinkSync(content, fsPath(path));
else {
const mode = entry.mode === "100755" ? 0o755 : 0o644;
writeFileSync(fsPath(path), content, { mode, flag: "wx" });
}
written.push(path);
}
const names = [
...new Set([...replaced, ...written].map((path) => rootName(path)!)),
];
if (names.length > 0) {
core.warning(
`The pull request and ${baseBranch} differ in what npm and Bun read (${names.join(", ")}), so that branch's version is used and Claude can't edit it. See docs/security.md.`,
);
}
return names;
}
/**
* Rules that stop a session from editing tool config, for a pull request whose author
* doesn't have write access. Other package.json edits are ordinary work. The rules
* cover subdirectories too, because a hook runs in the session's current directory.
*/
export const toolConfigDenyRules = (replaced: string[]): string[] => [
...TOOL_CONFIG.map((name) =>
name === "node_modules" ? "Edit(./node_modules/**)" : `Edit(./${name})`,
),
...(replaced.includes(PACKAGE_JSON) ? ["Edit(package.json)"] : []),
];
+50 -20
View File
@@ -10,6 +10,49 @@ export class WorkflowValidationSkipError extends Error {
}
}
type AppTokenExchangeErrorResponse = {
error?: {
message?: string;
details?: {
error_code?: string;
};
};
type?: string;
message?: string;
};
const WORKFLOW_VALIDATION_ERROR_CODES = new Set([
"workflow_not_found_on_default_branch",
]);
function getAppTokenExchangeErrorMessage(
responseJson: AppTokenExchangeErrorResponse,
): string {
return responseJson.error?.message ?? responseJson.message ?? "Unknown error";
}
function isWorkflowValidationError(
status: number,
responseJson: AppTokenExchangeErrorResponse,
): boolean {
const errorCode = responseJson.error?.details?.error_code;
if (
errorCode !== undefined &&
WORKFLOW_VALIDATION_ERROR_CODES.has(errorCode)
) {
return true;
}
if (status !== 401) {
return false;
}
const workflowValidationMessage = "workflow validation failed";
return [responseJson.message, responseJson.error?.message].some((message) =>
message?.toLowerCase().includes(workflowValidationMessage),
);
}
async function getOidcToken(): Promise<string> {
try {
const oidcToken = await core.getIDToken("claude-code-github-action");
@@ -80,25 +123,11 @@ async function exchangeForAppToken(
);
if (!response.ok) {
const responseJson = (await response.json()) as {
error?: {
message?: string;
details?: {
error_code?: string;
};
};
type?: string;
message?: string;
};
const responseJson =
(await response.json()) as AppTokenExchangeErrorResponse;
// Check for specific workflow validation error codes that should skip the action
const errorCode = responseJson.error?.details?.error_code;
if (errorCode === "workflow_not_found_on_default_branch") {
const message =
responseJson.message ??
responseJson.error?.message ??
"Workflow validation failed";
if (isWorkflowValidationError(response.status, responseJson)) {
const message = getAppTokenExchangeErrorMessage(responseJson);
core.warning(`Skipping action due to workflow validation: ${message}`);
console.log(
"Action skipped due to workflow validation error. This is expected when adding Claude Code workflows to new repositories or on PRs with workflow changes. If you're seeing this, your workflow will begin working once you merge your PR.",
@@ -106,10 +135,11 @@ async function exchangeForAppToken(
throw new WorkflowValidationSkipError(message);
}
const message = getAppTokenExchangeErrorMessage(responseJson);
console.error(
`App token exchange failed: ${response.status} ${response.statusText} - ${responseJson?.error?.message ?? "Unknown error"}`,
`App token exchange failed: ${response.status} ${response.statusText} - ${message}`,
);
throw new Error(`${responseJson?.error?.message ?? "Unknown error"}`);
throw new Error(message);
}
const appTokenData = (await response.json()) as {
+19 -5
View File
@@ -1,14 +1,23 @@
// Types for GitHub GraphQL query responses
// GitHub's GraphQL `author`/`actor` fields resolve to null when the underlying
// account has been deleted (the "ghost" user). Any field typed as
// `GitHubAuthor | null` can therefore be null at runtime and must be guarded.
// `__typename` distinguishes an App/bot actor from a human. GraphQL's
// `Actor.login` returns the bare name for bots ("dependabot"), unlike REST which
// appends a suffix ("dependabot[bot]"), so the typename is the only reliable bot
// signal on this data. See `resolveActorName` in `utils/actor-filter.ts`.
export type GitHubAuthor = {
login: string;
name?: string;
__typename?: string;
};
export type GitHubComment = {
id: string;
databaseId: string;
body: string;
author: GitHubAuthor;
author: GitHubAuthor | null;
createdAt: string;
updatedAt?: string;
lastEditedAt?: string;
@@ -18,6 +27,7 @@ export type GitHubComment = {
export type GitHubReviewComment = GitHubComment & {
path: string;
line: number | null;
diffHunk?: string | null;
};
export type GitHubCommit = {
@@ -39,7 +49,7 @@ export type GitHubFile = {
export type GitHubReview = {
id: string;
databaseId: string;
author: GitHubAuthor;
author: GitHubAuthor | null;
body: string;
state: string;
submittedAt: string;
@@ -53,7 +63,7 @@ export type GitHubReview = {
export type GitHubPullRequest = {
title: string;
body: string;
author: GitHubAuthor;
author: GitHubAuthor | null;
baseRefName: string;
headRefName: string;
headRefOid: string;
@@ -81,9 +91,13 @@ export type GitHubPullRequest = {
commit: GitHubCommit;
}>;
};
// GitHub's GraphQL `files` field resolves to null when the PR's diff is too
// large for GitHub to compute (very large PRs). `changedFiles` is also
// misreported as 0 in that case, so the null must be guarded and treated as
// "file list unavailable" rather than "no files changed".
files: {
nodes: GitHubFile[];
};
} | null;
comments: {
nodes: GitHubComment[];
};
@@ -95,7 +109,7 @@ export type GitHubPullRequest = {
export type GitHubIssue = {
title: string;
body: string;
author: GitHubAuthor;
author: GitHubAuthor | null;
createdAt: string;
updatedAt?: string;
lastEditedAt?: string;
+25
View File
@@ -11,6 +11,31 @@ export function parseActorFilter(filterString: string): string[] {
.filter((actor) => actor.length > 0);
}
/**
* Resolves the name to match actor filter patterns against.
*
* GitHub's GraphQL API returns the bare login for App actors ("dependabot"),
* whereas REST and the GitHub UI use a "[bot]" suffix ("dependabot[bot]"). Users
* write filter patterns in the suffixed form, both the documented "*[bot]"
* wildcard and exact entries like "renovate[bot]", so GraphQL bot logins are
* normalized to that form before matching. Without this no "[bot]" pattern can
* ever match, because the suffix is simply absent from the data.
*
* @param author - Comment author; null for deleted ("ghost") accounts
* @returns Actor name, "[bot]"-suffixed for App actors
*/
export function resolveActorName(
author: { login: string; __typename?: string } | null | undefined,
): string {
if (!author) return "ghost";
if (author.__typename === "Bot" && !author.login.endsWith("[bot]")) {
return `${author.login}[bot]`;
}
return author.login;
}
/**
* Checks if an actor matches a pattern
* Supports wildcards: "*[bot]" matches all bots, "dependabot[bot]" matches specific
+183 -36
View File
@@ -14,6 +14,45 @@ const HTML_IMG_REGEX = new RegExp(
"gi",
);
const SIGNED_URL_REGEX =
/https:\/\/private-user-images\.githubusercontent\.com\/[^"]+\?jwt=[^"]+/g;
// GitHub identifies an uploaded asset by a GUID that appears both in the
// user-attachment URL and in the signed download URL rendered in body_html.
const ASSET_GUID_REGEX =
/[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}/i;
function extractAssetGuid(url: string): string | undefined {
return url.match(ASSET_GUID_REGEX)?.[0]?.toLowerCase();
}
function isSameText(a: string | null | undefined, b: string): boolean {
return a?.replaceAll("\r\n", "\n") === b.replaceAll("\r\n", "\n");
}
const SIGNED_URL_HOST = "private-user-images.githubusercontent.com";
// Signed download URLs have the shape /<owner-id>/<asset-id>-<guid>.<ext>.
// The GUID must come from the resolved filename, not from anywhere in the raw
// string, so text that merely embeds a GUID cannot claim another asset.
const SIGNED_URL_PATH_REGEX =
/^\/[^/]+\/[^/]*-([0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12})(?:\.[a-z0-9]+)?$/i;
const DEFAULT_IMAGE_DOWNLOAD_TIMEOUT_MS = 30_000;
function extractSignedUrlAssetGuid(signedUrl: string): string | undefined {
let parsed: URL;
try {
parsed = new URL(signedUrl);
} catch {
return undefined;
}
if (parsed.host !== SIGNED_URL_HOST) {
return undefined;
}
return parsed.pathname.match(SIGNED_URL_PATH_REGEX)?.[1]?.toLowerCase();
}
type IssueComment = {
type: "issue_comment";
id: string;
@@ -52,13 +91,19 @@ export type CommentWithImages =
| IssueBody
| PullRequestBody;
type ImageDownloadOptions = {
timeoutMs?: number;
};
export async function downloadCommentImages(
octokits: Octokits,
owner: string,
repo: string,
comments: CommentWithImages[],
options: ImageDownloadOptions = {},
): Promise<Map<string, string>> {
const urlToPathMap = new Map<string, string>();
const timeoutMs = options.timeoutMs ?? DEFAULT_IMAGE_DOWNLOAD_TIMEOUT_MS;
const downloadsDir = "/tmp/github-images";
await fs.mkdir(downloadsDir, { recursive: true });
@@ -66,6 +111,7 @@ export async function downloadCommentImages(
const commentsWithImages: Array<{
comment: CommentWithImages;
urls: string[];
id: string;
}> = [];
for (const comment of comments) {
@@ -81,21 +127,22 @@ export async function downloadCommentImages(
const urls = [...new Set([...markdownUrls, ...htmlUrls])];
if (urls.length > 0) {
commentsWithImages.push({ comment, urls });
const id =
comment.type === "issue_body"
? comment.issueNumber
: comment.type === "pr_body"
? comment.pullNumber
: comment.id;
commentsWithImages.push({ comment, urls, id });
console.log(`Found ${urls.length} image(s) in ${comment.type} ${id}`);
}
}
// Process each comment with images
for (const { comment, urls } of commentsWithImages) {
for (const { comment, urls, id } of commentsWithImages) {
try {
let bodyHtml: string | undefined;
let currentBody: string | null | undefined;
// Get the HTML version based on comment type
switch (comment.type) {
@@ -109,6 +156,7 @@ export async function downloadCommentImages(
},
});
bodyHtml = response.data.body_html;
currentBody = response.data.body;
break;
}
case "review_comment": {
@@ -121,6 +169,7 @@ export async function downloadCommentImages(
},
});
bodyHtml = response.data.body_html;
currentBody = response.data.body;
break;
}
case "review_body": {
@@ -134,6 +183,7 @@ export async function downloadCommentImages(
},
});
bodyHtml = response.data.body_html;
currentBody = response.data.body;
break;
}
case "issue_body": {
@@ -146,6 +196,7 @@ export async function downloadCommentImages(
},
});
bodyHtml = response.data.body_html;
currentBody = response.data.body;
break;
}
case "pr_body": {
@@ -159,55 +210,73 @@ export async function downloadCommentImages(
});
// Type here seems to be wrong
bodyHtml = (response.data as any).body_html;
currentBody = response.data.body;
break;
}
}
// body_html is rendered from GitHub's current text, which may have been
// edited since this run's copy was taken.
if (!isSameText(currentBody, comment.body)) {
console.warn(
`Skipping images in ${comment.type} ${id}: its text on GitHub differs from the text this run uses.`,
);
continue;
}
if (!bodyHtml) {
const id =
comment.type === "issue_body"
? comment.issueNumber
: comment.type === "pr_body"
? comment.pullNumber
: comment.id;
console.warn(`No HTML body found for ${comment.type} ${id}`);
continue;
}
// Extract signed URLs from HTML
const signedUrlRegex =
/https:\/\/private-user-images\.githubusercontent\.com\/[^"]+\?jwt=[^"]+/g;
const signedUrls = bodyHtml.match(signedUrlRegex) || [];
const signedUrls = bodyHtml.match(SIGNED_URL_REGEX) || [];
// Index the signed URLs by the asset GUID they reference. The signed
// URLs come from a separate render of the body, so their order and
// count are not guaranteed to line up with the URLs extracted from the
// markdown; pairing by asset identifier keeps each download tied to the
// URL it actually belongs to.
const signedUrlByGuid = new Map<string, string>();
for (const signedUrl of signedUrls) {
const guid = extractSignedUrlAssetGuid(signedUrl);
if (guid && !signedUrlByGuid.has(guid)) {
signedUrlByGuid.set(guid, signedUrl);
}
}
// Download each image
for (let i = 0; i < Math.min(signedUrls.length, urls.length); i++) {
const signedUrl = signedUrls[i];
const originalUrl = urls[i];
if (!signedUrl || !originalUrl) {
continue;
}
for (const [i, originalUrl] of urls.entries()) {
// Check if we've already downloaded this URL
if (urlToPathMap.has(originalUrl)) {
continue;
}
const fileExtension = getImageExtension(originalUrl);
const filename = `image-${Date.now()}-${i}${fileExtension}`;
const localPath = path.join(downloadsDir, filename);
const guid = extractAssetGuid(originalUrl);
const signedUrl = guid ? signedUrlByGuid.get(guid) : undefined;
if (!signedUrl) {
console.warn(
`No matching signed URL found for ${originalUrl}, skipping`,
);
continue;
}
try {
console.log(`Downloading ${originalUrl}...`);
const imageResponse = await fetch(signedUrl);
if (!imageResponse.ok) {
throw new Error(
`HTTP ${imageResponse.status}: ${imageResponse.statusText}`,
);
}
const buffer = await fetchImage(signedUrl, timeoutMs);
const arrayBuffer = await imageResponse.arrayBuffer();
const buffer = Buffer.from(arrayBuffer);
// GitHub user-attachment URLs (/user-attachments/assets/<uuid>) carry
// no file extension, so the URL-based guess silently falls back to
// ".png". When the bytes are actually JPEG/GIF/WebP, the saved file is
// mislabeled and the Read tool sends a base64 image with the wrong
// media_type, which the Anthropic API rejects (400 invalid_request).
// Detect the real type from the magic bytes and only fall back to the
// URL extension when the signature is unrecognized.
const fileExtension =
detectImageExtensionFromBuffer(buffer) ??
getImageExtension(originalUrl);
const filename = `image-${Date.now()}-${i}${fileExtension}`;
const localPath = path.join(downloadsDir, filename);
await fs.writeFile(localPath, buffer);
console.log(`✓ Saved: ${localPath}`);
@@ -218,12 +287,6 @@ export async function downloadCommentImages(
}
}
} catch (error) {
const id =
comment.type === "issue_body"
? comment.issueNumber
: comment.type === "pr_body"
? comment.pullNumber
: comment.id;
console.error(
`Failed to process images for ${comment.type} ${id}:`,
error,
@@ -234,6 +297,37 @@ export async function downloadCommentImages(
return urlToPathMap;
}
async function fetchImage(url: string, timeoutMs: number): Promise<Buffer> {
const controller = new AbortController();
let timeoutHandle: ReturnType<typeof setTimeout> | undefined;
const timeoutPromise = new Promise<never>((_, reject) => {
timeoutHandle = setTimeout(() => {
controller.abort();
reject(new Error(`Image download timed out after ${timeoutMs}ms`));
}, timeoutMs);
});
try {
const response = await Promise.race([
fetch(url, { signal: controller.signal }),
timeoutPromise,
]);
if (!response.ok) {
throw new Error(`HTTP ${response.status}: ${response.statusText}`);
}
const arrayBuffer = await Promise.race([
response.arrayBuffer(),
timeoutPromise,
]);
return Buffer.from(arrayBuffer);
} finally {
if (timeoutHandle !== undefined) {
clearTimeout(timeoutHandle);
}
}
}
function getImageExtension(url: string): string {
const urlParts = url.split("/");
const filename = urlParts[urlParts.length - 1];
@@ -244,3 +338,56 @@ function getImageExtension(url: string): string {
const match = filename.match(/\.(png|jpg|jpeg|gif|webp|svg)$/i);
return match ? match[0] : ".png";
}
/**
* Determine an image's file extension from its magic bytes, independent of the
* (often extensionless) source URL. Returns undefined when the signature is not
* a format we can confidently identify, so the caller can fall back to the
* URL-based extension. Covers the raster formats the Anthropic API accepts.
*/
function detectImageExtensionFromBuffer(buffer: Buffer): string | undefined {
// PNG: 89 50 4E 47 0D 0A 1A 0A
if (
buffer.length >= 8 &&
buffer[0] === 0x89 &&
buffer[1] === 0x50 &&
buffer[2] === 0x4e &&
buffer[3] === 0x47
) {
return ".png";
}
// JPEG: FF D8 FF
if (
buffer.length >= 3 &&
buffer[0] === 0xff &&
buffer[1] === 0xd8 &&
buffer[2] === 0xff
) {
return ".jpg";
}
// GIF: "GIF8" (47 49 46 38)
if (
buffer.length >= 6 &&
buffer[0] === 0x47 &&
buffer[1] === 0x49 &&
buffer[2] === 0x46 &&
buffer[3] === 0x38
) {
return ".gif";
}
// WebP: "RIFF" (52 49 46 46) .... "WEBP" (57 45 42 50) at offset 8
if (
buffer.length >= 12 &&
buffer[0] === 0x52 &&
buffer[1] === 0x49 &&
buffer[2] === 0x46 &&
buffer[3] === 0x46 &&
buffer[8] === 0x57 &&
buffer[9] === 0x45 &&
buffer[10] === 0x42 &&
buffer[11] === 0x50
) {
return ".webp";
}
return undefined;
}
+73 -11
View File
@@ -10,7 +10,13 @@ export function stripInvisibleCharacters(content: string): string {
}
export function stripMarkdownImageAltText(content: string): string {
return content.replace(/!\[[^\]]*\]\(/g, "![](");
// Inline images: ![alt](url) -> ![](url)
content = content.replace(/!\[[^\]]*\]\(/g, "![](");
// Reference-style images: ![alt][ref] -> ![][ref] (keep the label, drop the
// alt text, which is otherwise a hidden-instruction channel just like the
// inline form above).
content = content.replace(/!\[[^\]]*\](\[[^\]]*\])/g, "![]$1");
return content;
}
export function stripMarkdownLinkTitles(content: string): string {
@@ -20,15 +26,23 @@ export function stripMarkdownLinkTitles(content: string): string {
}
export function stripHiddenAttributes(content: string): string {
content = content.replace(/\salt\s*=\s*["'][^"']*["']/gi, "");
// Quoted values are matched per quote type so that a value containing the
// other quote character (e.g. an apostrophe inside a double-quoted value)
// does not terminate the match early and mangle surrounding content (#1366).
content = content.replace(/\salt\s*=\s*"[^"]*"/gi, "");
content = content.replace(/\salt\s*=\s*'[^']*'/gi, "");
content = content.replace(/\salt\s*=\s*[^\s>]+/gi, "");
content = content.replace(/\stitle\s*=\s*["'][^"']*["']/gi, "");
content = content.replace(/\stitle\s*=\s*"[^"]*"/gi, "");
content = content.replace(/\stitle\s*=\s*'[^']*'/gi, "");
content = content.replace(/\stitle\s*=\s*[^\s>]+/gi, "");
content = content.replace(/\saria-label\s*=\s*["'][^"']*["']/gi, "");
content = content.replace(/\saria-label\s*=\s*"[^"]*"/gi, "");
content = content.replace(/\saria-label\s*=\s*'[^']*'/gi, "");
content = content.replace(/\saria-label\s*=\s*[^\s>]+/gi, "");
content = content.replace(/\sdata-[a-zA-Z0-9-]+\s*=\s*["'][^"']*["']/gi, "");
content = content.replace(/\sdata-[a-zA-Z0-9-]+\s*=\s*"[^"]*"/gi, "");
content = content.replace(/\sdata-[a-zA-Z0-9-]+\s*=\s*'[^']*'/gi, "");
content = content.replace(/\sdata-[a-zA-Z0-9-]+\s*=\s*[^\s>]+/gi, "");
content = content.replace(/\splaceholder\s*=\s*["'][^"']*["']/gi, "");
content = content.replace(/\splaceholder\s*=\s*"[^"]*"/gi, "");
content = content.replace(/\splaceholder\s*=\s*'[^']*'/gi, "");
content = content.replace(/\splaceholder\s*=\s*[^\s>]+/gi, "");
return content;
}
@@ -62,34 +76,82 @@ export function sanitizeContent(content: string): string {
return content;
}
/**
* Redact well-known credential formats (GitHub, Anthropic, AWS, Slack, JWTs)
* from arbitrary text. Callers don't need to know which vendor a value belongs to.
*
* Vendor-prefixed formats are matched without a leading word boundary: the
* prefix already anchors them, and runtime output frequently puts a word
* character directly against the value (e.g. an ANSI color code ending in `m`,
* or a serialized JSON escape such as `\n`).
*/
export function redactSecrets(content: string): string {
content = redactGitHubTokens(content);
// Anthropic API keys: sk-ant-...
content = content.replace(
/sk-ant-[A-Za-z0-9_-]{20,}/g,
"[REDACTED_ANTHROPIC_KEY]",
);
// AWS access key ids: AKIA/ASIA followed by 16 uppercase alphanumerics. All
// uppercase alphanumeric, so keep a leading boundary to avoid matching inside
// larger blobs; also treat a JSON escape or ANSI color code as a boundary.
content = content.replace(
/(?:\b|(?<=\\(?:[nrtbf"\\/]|u[0-9a-fA-F]{4}))|(?<=\[[0-9;]*m))(?:AKIA|ASIA)[A-Z0-9]{16}\b/g,
"[REDACTED_AWS_KEY_ID]",
);
// Slack tokens: xoxb-, xoxp-, xoxa-, xoxs-, xoxr-
content = content.replace(
/xox[abpsr]-[A-Za-z0-9-]{10,}/g,
"[REDACTED_SLACK_TOKEN]",
);
// JWT-shaped strings: three base64url segments, the first two starting
// with eyJ (base64 of `{"`).
content = content.replace(
/eyJ[A-Za-z0-9_-]{10,2000}\.eyJ[A-Za-z0-9_-]{10,4000}\.[A-Za-z0-9_-]{10,2000}\b/g,
"[REDACTED_JWT]",
);
return content;
}
export function redactGitHubTokens(content: string): string {
// GitHub Personal Access Tokens (classic): ghp_XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX (40 chars)
content = content.replace(
/\bghp_[A-Za-z0-9]{36}\b/g,
/ghp_[A-Za-z0-9]{36}\b/g,
"[REDACTED_GITHUB_TOKEN]",
);
// GitHub OAuth tokens: gho_XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX (40 chars)
content = content.replace(
/\bgho_[A-Za-z0-9]{36}\b/g,
/gho_[A-Za-z0-9]{36}\b/g,
"[REDACTED_GITHUB_TOKEN]",
);
// GitHub user-to-server tokens: ghu_XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX (40 chars)
content = content.replace(
/ghu_[A-Za-z0-9]{36}\b/g,
"[REDACTED_GITHUB_TOKEN]",
);
// GitHub installation tokens: ghs_XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX (40 chars)
content = content.replace(
/\bghs_[A-Za-z0-9]{36}\b/g,
/ghs_[A-Za-z0-9]{36}\b/g,
"[REDACTED_GITHUB_TOKEN]",
);
// GitHub refresh tokens: ghr_XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX (40 chars)
content = content.replace(
/\bghr_[A-Za-z0-9]{36}\b/g,
/ghr_[A-Za-z0-9]{36}\b/g,
"[REDACTED_GITHUB_TOKEN]",
);
// GitHub fine-grained personal access tokens: github_pat_XXXXXXXXXX (up to 255 chars)
content = content.replace(
/\bgithub_pat_[A-Za-z0-9_]{11,221}\b/g,
/github_pat_[A-Za-z0-9_]{11,221}\b/g,
"[REDACTED_GITHUB_TOKEN]",
);
+8
View File
@@ -8,6 +8,14 @@
import type { Octokit } from "@octokit/rest";
import type { GitHubContext } from "../context";
/** Whether two GitHub logins name the same account. Logins are case-insensitive. */
export function isSameLogin(
a: string | null | undefined,
b: string | null | undefined,
): boolean {
return !!a && !!b && a.toLowerCase() === b.toLowerCase();
}
function isAllowedBot(actor: string, allowedBots: string): boolean {
const trimmed = allowedBots.trim();
if (trimmed === "*") return true;
+78
View File
@@ -0,0 +1,78 @@
import type { Octokit } from "@octokit/rest";
import { retryWithBackoff } from "../../utils/retry";
import { toLookupError, type LookupError } from "./untrusted-pr";
import {
isEntityContext,
isWorkflowRunEvent,
type GitHubContext,
} from "../context";
type Repository = { id?: number } | null | undefined;
// A deleted fork has no repository.
const isSameRepository = (head: Repository, base: Repository): boolean =>
typeof head?.id === "number" && head.id === base?.id;
// Only "true" allows forks, so a mistyped or empty value refuses them.
export const forksAllowed = (): boolean =>
(process.env.ALLOW_FORK_PRS ?? "true").trim().toLowerCase() === "true";
/**
* Whether allow_fork_prs is off and the event says that it is about a fork's pull
* request, or follows a workflow run of a fork's code. Needs no token. A comment on
* a pull request's conversation doesn't say: see isCommentOnRefusedForkPr.
*/
export function isRefusedForkEvent(context: GitHubContext): boolean {
if (forksAllowed()) return false;
if (isWorkflowRunEvent(context)) {
const { workflow_run: run, repository } = context.payload;
return !isSameRepository(run?.head_repository, repository);
}
if (isEntityContext(context) && "pull_request" in context.payload) {
const { head, base } = context.payload.pull_request;
return !isSameRepository(head.repo, base.repo);
}
return false;
}
/**
* Whether allow_fork_prs is off and the event is a comment on the conversation of a
* fork's pull request. Throws if that can't be found out.
*/
export async function isCommentOnRefusedForkPr(
octokit: Octokit,
context: GitHubContext,
): Promise<boolean> {
if (
forksAllowed() ||
!isEntityContext(context) ||
!context.isPR ||
"pull_request" in context.payload
) {
return false;
}
const { owner, repo } = context.repository;
const lookUp = async () => {
try {
return await octokit.pulls.get({
owner,
repo,
pull_number: context.entityNumber,
});
} catch (error) {
throw toLookupError(error);
}
};
const { data } = await retryWithBackoff(lookUp, {
initialDelayMs: 1000,
shouldRetry: (error) => (error as LookupError).retryable,
}).catch((error: LookupError) => {
// Only the bounded text from toLookupError reaches the log.
throw new Error(
`Could not find out whether pull request #${context.entityNumber} is from a fork (${error.message}). The token needs pull-requests: read.`,
);
});
return !isSameRepository(data.head.repo, data.base.repo);
}
+114 -3
View File
@@ -1,5 +1,9 @@
import * as core from "@actions/core";
import type { ParsedGitHubContext } from "../context";
import {
isPullRequestEvent,
isWorkflowRunEvent,
type GitHubContext,
} from "../context";
import type { Octokit } from "@octokit/rest";
/**
@@ -24,6 +28,49 @@ function isAllowedBot(actor: string, allowedBots: string): boolean {
return allowedList.includes(normalizedActor);
}
/**
* Collect the actors whose repository access should be checked. This is
* normally just the workflow actor (GITHUB_ACTOR). For workflow_run events
* the actor that started the upstream run is checked as well when it
* differs, since that is the account the run originates from. GitHub credits a
* run to whoever caused it, so a maintainer reopening a fork's pull request
* starts a run of the fork's code; the fork's owner is checked too.
*
* Returns null when the run's origin cannot be determined.
*/
function getActorsToCheck(context: GitHubContext): string[] | null {
const actors = [context.actor];
if (isWorkflowRunEvent(context)) {
const run = context.payload.workflow_run;
const runActor = run?.actor?.login;
if (runActor && !actors.includes(runActor)) {
core.info(
`workflow_run was started by ${runActor}; checking permissions for that actor as well`,
);
actors.push(runActor);
}
const headRepository = run?.head_repository;
if (!headRepository?.full_name) {
core.warning("workflow_run does not say which repository it ran from");
return null;
}
const isFork =
headRepository.full_name.toLowerCase() !==
context.repository.full_name.toLowerCase();
const forkOwner = headRepository.full_name.split("/")[0]!;
if (isFork && !actors.includes(forkOwner)) {
core.info(
`workflow_run ran code from a fork owned by ${forkOwner}; checking permissions for that account as well`,
);
actors.push(forkOwner);
}
}
return actors;
}
/**
* Check if the actor has write permissions to the repository
* @param octokit - The Octokit REST client
@@ -34,11 +81,75 @@ function isAllowedBot(actor: string, allowedBots: string): boolean {
*/
export async function checkWritePermissions(
octokit: Octokit,
context: ParsedGitHubContext,
context: GitHubContext,
allowedNonWriteUsers?: string,
githubTokenProvided?: boolean,
): Promise<boolean> {
const { repository, actor } = context;
const actors = getActorsToCheck(context);
if (!actors) return false;
for (const actor of actors) {
const allowed = await checkActorWritePermissions(
octokit,
context,
actor,
allowedNonWriteUsers,
githubTokenProvided,
);
if (!allowed) return false;
}
return true;
}
/**
* When a bot acts on a pull request, the bot is the actor even though someone
* else wrote the pull request. Bots are not checked for repository access, so an
* allowed bot acting on an outsider's fork would otherwise start a run for them.
* Call this once a run is going to start, so that bot activity with nothing to
* do still skips quietly.
*/
export async function checkBotActorPullRequestAuthor(
octokit: Octokit,
context: GitHubContext,
allowedNonWriteUsers?: string,
githubTokenProvided?: boolean,
): Promise<boolean> {
if (!isPullRequestEvent(context) || !context.actor.endsWith("[bot]")) {
return true;
}
const author = context.payload.pull_request.user?.login;
// A bot's own pull request is left to the check that bots are allowed.
if (author === context.actor) return true;
let allowed = false;
if (author?.endsWith("[bot]")) {
allowed = isAllowedBot(author, context.inputs.allowedBots ?? "");
} else if (author) {
allowed = await checkActorWritePermissions(
octokit,
context,
author,
allowedNonWriteUsers,
githubTokenProvided,
);
}
if (!allowed) {
core.warning(
`${context.actor} acted on a pull request by ${author ?? "an unknown author"}, who is not allowed to start a run`,
);
}
return allowed;
}
async function checkActorWritePermissions(
octokit: Octokit,
context: GitHubContext,
actor: string,
allowedNonWriteUsers?: string,
githubTokenProvided?: boolean,
): Promise<boolean> {
const { repository } = context;
const allowedBots = context.inputs.allowedBots ?? "";
try {
+82 -20
View File
@@ -1,6 +1,5 @@
#!/usr/bin/env bun
import * as core from "@actions/core";
import {
isIssuesEvent,
isIssuesAssignedEvent,
@@ -10,35 +9,86 @@ import {
isPullRequestReviewCommentEvent,
} from "../context";
import type { ParsedGitHubContext } from "../context";
import { isSameLogin } from "./actor";
/**
* Whether the text of a comment or review event comes from the person who was checked.
* Deleting or editing someone else's fires these events with the moderator as the actor.
*/
function isTextFromActor(context: ParsedGitHubContext): boolean {
if (
(isIssueCommentEvent(context) ||
isPullRequestReviewCommentEvent(context)) &&
(context.eventAction === "deleted" ||
!isSameLogin(context.payload.comment.user?.login, context.actor))
) {
console.log("Comment was not written by the actor, or was deleted");
return false;
}
if (
isPullRequestReviewEvent(context) &&
!isSameLogin(context.payload.review.user?.login, context.actor)
) {
console.log("Review was written by someone other than the actor");
return false;
}
return true;
}
/**
* Whether a comment or review event comes from moderation, so that its text must not
* start a run. A new comment or review is always its author's own, and some apps have
* one login as its author and another as the actor.
*/
export function isModeration(context: ParsedGitHubContext): boolean {
const isNew =
context.eventAction === "created" || context.eventAction === "submitted";
return !isNew && !isTextFromActor(context);
}
export function checkContainsTrigger(context: ParsedGitHubContext): boolean {
const {
inputs: { assigneeTrigger, labelTrigger, triggerPhrase, prompt },
} = context;
// If prompt is provided, always trigger
if (!isTextFromActor(context)) return false;
// A prompt replaces the trigger phrase, but a configured label or assignee
// still decides which label or assignment event starts a run. Otherwise any
// triage action on an issue would start one.
if (prompt) {
if (
isIssuesEvent(context) &&
context.eventAction === "labeled" &&
labelTrigger &&
!matchesLabelTrigger(context, labelTrigger)
) {
console.log(`Issue label is not the trigger label '${labelTrigger}'`);
return false;
}
if (
isIssuesAssignedEvent(context) &&
assigneeTrigger &&
!matchesAssigneeTrigger(context, assigneeTrigger)
) {
console.log(`Issue was not assigned to '${assigneeTrigger}'`);
return false;
}
console.log(`Prompt provided, triggering action`);
return true;
}
// Check for assignee trigger
if (isIssuesAssignedEvent(context)) {
// Remove @ symbol from assignee_trigger if present
let triggerUser = assigneeTrigger.replace(/^@/, "");
const assigneeUsername = context.payload.assignee?.login || "";
if (triggerUser && assigneeUsername === triggerUser) {
console.log(`Issue assigned to trigger user '${triggerUser}'`);
if (matchesAssigneeTrigger(context, assigneeTrigger)) {
console.log(`Issue assigned to trigger user '${assigneeTrigger}'`);
return true;
}
}
// Check for label trigger
if (isIssuesEvent(context) && context.eventAction === "labeled") {
const labelName = (context.payload as any).label?.name || "";
if (labelTrigger && labelName === labelTrigger) {
if (matchesLabelTrigger(context, labelTrigger)) {
console.log(`Issue labeled with trigger label '${labelTrigger}'`);
return true;
}
@@ -122,9 +172,7 @@ export function checkContainsTrigger(context: ParsedGitHubContext): boolean {
isIssueCommentEvent(context) ||
isPullRequestReviewCommentEvent(context)
) {
const commentBody = isIssueCommentEvent(context)
? context.payload.comment.body
: context.payload.comment.body;
const commentBody = context.payload.comment.body;
// Check for exact match with word boundaries or punctuation
const regex = new RegExp(
`(^|\\s)${escapeRegExp(triggerPhrase)}([\\s.,!?;:]|$)`,
@@ -141,12 +189,26 @@ export function checkContainsTrigger(context: ParsedGitHubContext): boolean {
return false;
}
function matchesLabelTrigger(
context: ParsedGitHubContext,
labelTrigger: string,
): boolean {
const labelName = (context.payload as any).label?.name || "";
return (
!!labelTrigger && labelName.toLowerCase() === labelTrigger.toLowerCase()
);
}
function matchesAssigneeTrigger(
context: ParsedGitHubContext,
assigneeTrigger: string,
): boolean {
// Remove @ symbol from assignee_trigger if present
const triggerUser = assigneeTrigger.replace(/^@/, "");
const assigneeUsername = (context.payload as any).assignee?.login || "";
return !!triggerUser && assigneeUsername === triggerUser;
}
export function escapeRegExp(string: string) {
return string.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
}
export async function checkTriggerAction(context: ParsedGitHubContext) {
const containsTrigger = checkContainsTrigger(context);
core.setOutput("contains_trigger", containsTrigger.toString());
return containsTrigger;
}
+221
View File
@@ -0,0 +1,221 @@
import * as core from "@actions/core";
import type { Octokit } from "@octokit/rest";
import {
parseAllowedTools,
stripShellComments,
} from "../../modes/agent/parse-tools";
import { retryWithBackoff } from "../../utils/retry";
import type { ParsedGitHubContext } from "../context";
/**
* Commands that commonly execute code from the checkout: package scripts, test
* files, build files and config loaded by interpreters. On a pull request from
* an author without write access that code is attacker-controlled, while the
* session holds the job's tokens. Deny rules take precedence over allow rules,
* so these stay blocked even when claude_args allows them.
*
* The list cannot cover every command, for example a repository's own scripts.
* Rules in claude_args that this list does not cover are logged as a warning.
* Every entry has the form Bash(<command>:*), which isDenied relies on.
*/
export const UNTRUSTED_PR_DISALLOWED_TOOLS = [
// JavaScript
"Bash(bun:*)",
"Bash(bunx:*)",
"Bash(deno:*)",
"Bash(jest:*)",
"Bash(node:*)",
"Bash(npm:*)",
"Bash(npx:*)",
"Bash(pnpm:*)",
"Bash(ts-node:*)",
"Bash(tsx:*)",
"Bash(vitest:*)",
"Bash(yarn:*)",
// Python
"Bash(hatch:*)",
"Bash(nox:*)",
"Bash(pip:*)",
"Bash(pip3:*)",
"Bash(pipenv:*)",
"Bash(poetry:*)",
"Bash(pytest:*)",
"Bash(python:*)",
"Bash(python3:*)",
"Bash(tox:*)",
"Bash(uv:*)",
"Bash(uvx:*)",
// Other languages and build tools
"Bash(bundle:*)",
"Bash(cargo:*)",
"Bash(cmake:*)",
"Bash(composer:*)",
"Bash(docker:*)",
"Bash(dotnet:*)",
"Bash(go:*)",
"Bash(gradle:*)",
"Bash(./gradlew:*)",
"Bash(java:*)",
"Bash(just:*)",
"Bash(make:*)",
"Bash(mvn:*)",
"Bash(./mvnw:*)",
"Bash(perl:*)",
"Bash(php:*)",
"Bash(rake:*)",
"Bash(rspec:*)",
"Bash(ruby:*)",
"Bash(task:*)",
// Shells
"Bash(bash:*)",
"Bash(dash:*)",
"Bash(sh:*)",
"Bash(zsh:*)",
// The simplest ways to point git at a program. Other git commands can still
// run programs, so a broad rule like Bash(git:*) stays unsafe and is warned about.
"Bash(git config:*)",
"Bash(git -c:*)",
];
const SKIPS_PERMISSION_CHECKS =
/--(allow-)?dangerously-skip-permissions|--permission-mode[=\s]+["']?bypassPermissions/;
function getPrAuthor(context: ParsedGitHubContext): string | undefined {
const { payload } = context;
if ("pull_request" in payload) {
return payload.pull_request.user?.login;
}
// On issue_comment events a pull request arrives as an issue, whose user is
// the pull request's author.
if ("issue" in payload) {
return payload.issue.user?.login;
}
return undefined;
}
export type LookupError = Error & { retryable: boolean };
/**
* Reduces a failed API call to its HTTP status, so that nothing else from the
* error can reach the workflow log, and decides whether to try again. A 404 or a
* plain 403 won't change, but GitHub also uses 403 for rate
* limits.
*/
export function toLookupError(error: unknown): LookupError {
const { status, response } = error as {
status?: number;
response?: { headers?: Record<string, string | number | undefined> };
};
const headers = response?.headers;
const rateLimited =
status === 429 ||
(status === 403 &&
(headers?.["retry-after"] !== undefined ||
String(headers?.["x-ratelimit-remaining"]) === "0"));
return Object.assign(
new Error(status === undefined ? "no response" : `HTTP ${status}`),
{
retryable: status === undefined || status >= 500 || rateLimited,
},
);
}
/**
* Whether the pull request's author has write access to the repository. This
* is about who wrote the code, not who triggered the run, so it is deliberately
* stricter than checkWritePermissions: bots are not trusted, allowed_bots and
* allowed_non_write_users do not apply, and any failure counts as untrusted.
*/
async function isPrAuthorTrusted(
octokit: Octokit,
context: ParsedGitHubContext,
): Promise<boolean> {
const author = getPrAuthor(context);
if (!author || author.endsWith("[bot]")) {
return false;
}
try {
const response = await retryWithBackoff(
async () => {
try {
return await octokit.repos.getCollaboratorPermissionLevel({
owner: context.repository.owner,
repo: context.repository.repo,
username: author,
});
} catch (error) {
throw toLookupError(error);
}
},
{
initialDelayMs: 1000,
shouldRetry: (error) => (error as LookupError).retryable,
},
);
const { permission } = response.data;
return permission === "admin" || permission === "write";
} catch (error) {
// Only the bounded text from toLookupError reaches the log.
core.warning(
`Could not check repository access for pull request author ${author}, treating them as untrusted: ${(error as Error).message}`,
);
return false;
}
}
/** Whether a Bash allow rule only permits commands that the deny list blocks. */
function isDenied(rule: string): boolean {
const command = rule.slice("Bash(".length, -")".length);
return UNTRUSTED_PR_DISALLOWED_TOOLS.some((denied) => {
const prefix = denied.slice("Bash(".length, -":*)".length);
return (
command === prefix ||
command.startsWith(`${prefix} `) ||
command.startsWith(`${prefix}:`)
);
});
}
/**
* Returns the tools to deny when the session works on a pull request from an
* author without write access, or an empty list when no restriction applies.
* On by default: only restrict_untrusted_prs: false turns it off, so a mistyped
* value keeps the restriction.
*/
export async function getUntrustedPrDisallowedTools(
octokit: Octokit,
context: ParsedGitHubContext,
): Promise<string[]> {
const disabled =
process.env.RESTRICT_UNTRUSTED_PRS?.trim().toLowerCase() === "false";
if (disabled || !context.isPR) {
return [];
}
if (await isPrAuthorTrusted(octokit, context)) {
return [];
}
core.warning(
`The author of pull request #${context.entityNumber} is not confirmed to have write access, so commands that run code from the checkout are denied. Set restrict_untrusted_prs to "false" to turn this off.`,
);
// These warnings report a count and fixed text, not the content of claude_args,
// which is never printed to the workflow log.
const claudeArgs = process.env.CLAUDE_ARGS || "";
const uncovered = parseAllowedTools(claudeArgs).filter(
(tool) => tool === "Bash" || (tool.startsWith("Bash(") && !isDenied(tool)),
);
if (uncovered.length > 0) {
core.warning(
`${uncovered.length} Bash rule(s) in claude_args are not on the deny list and stay allowed. Check that they cannot run code from the pull request.`,
);
}
if (SKIPS_PERMISSION_CHECKS.test(stripShellComments(claudeArgs))) {
core.warning(
"claude_args turns off permission checks, so commands that are not on the deny list run without approval.",
);
}
return UNTRUSTED_PR_DISALLOWED_TOOLS;
}
+22
View File
@@ -0,0 +1,22 @@
/**
* Decides whether a file has to be committed as a base64 blob instead of being
* inlined in the Git tree as UTF-8 text.
*
* Inlining is only safe for content that survives a UTF-8 decode untouched;
* anything else gets its invalid bytes replaced during the decode, which
* silently corrupts the committed file. A NUL byte is treated as binary for the
* same reason Git does it: no text file carries one, and it is the cheapest
* signal available.
*/
export function isBinaryContent(content: Buffer): boolean {
if (content.includes(0)) {
return true;
}
try {
new TextDecoder("utf-8", { fatal: true }).decode(content);
return false;
} catch {
return true;
}
}
+19
View File
@@ -0,0 +1,19 @@
import type { Octokit } from "@octokit/rest";
type ActionsClient = Octokit["actions"];
export type WorkflowRunsParams = Parameters<
ActionsClient["listWorkflowRunsForRepo"]
>[0];
export type WorkflowJobsParams = Parameters<
ActionsClient["listJobsForWorkflowRun"]
>[0];
export function listWorkflowRuns(client: Octokit, params: WorkflowRunsParams) {
return client.paginate(client.actions.listWorkflowRunsForRepo, params);
}
export function listWorkflowJobs(client: Octokit, params: WorkflowJobsParams) {
return client.paginate(client.actions.listJobsForWorkflowRun, params);
}
+63 -30
View File
@@ -6,6 +6,10 @@ import { z } from "zod";
import { GITHUB_API_URL } from "../github/api/config";
import { mkdir, writeFile } from "fs/promises";
import { Octokit } from "@octokit/rest";
import {
listWorkflowJobs,
listWorkflowRuns,
} from "./github-actions-pagination";
const REPO_OWNER = process.env.REPO_OWNER;
const REPO_NAME = process.env.REPO_NAME;
@@ -13,11 +17,18 @@ const PR_NUMBER = process.env.PR_NUMBER;
const GITHUB_TOKEN = process.env.GITHUB_TOKEN;
const RUNNER_TEMP = process.env.RUNNER_TEMP || "/tmp";
if (!REPO_OWNER || !REPO_NAME || !PR_NUMBER || !GITHUB_TOKEN) {
console.error(
"[GitHub CI Server] Error: REPO_OWNER, REPO_NAME, PR_NUMBER, and GITHUB_TOKEN environment variables are required",
);
process.exit(1);
// Job logs are fetched by ID from GitHub-hosted storage; bound the request so a
// stalled fetch can't hang this MCP call forever. Mirrors the timeout added to
// fetchImage() in src/github/utils/image-downloader.ts (#1625).
const DOWNLOAD_JOB_LOG_TIMEOUT_MS = 30_000;
if (import.meta.main) {
if (!REPO_OWNER || !REPO_NAME || !PR_NUMBER || !GITHUB_TOKEN) {
console.error(
"[GitHub CI Server] Error: REPO_OWNER, REPO_NAME, PR_NUMBER, and GITHUB_TOKEN environment variables are required",
);
process.exit(1);
}
}
const server = new McpServer({
@@ -66,7 +77,7 @@ server.tool(
});
const headSha = prData.head.sha;
const { data: runsData } = await client.actions.listWorkflowRunsForRepo({
const runs = await listWorkflowRuns(client, {
owner: REPO_OWNER!,
repo: REPO_NAME!,
head_sha: headSha,
@@ -74,7 +85,6 @@ server.tool(
});
// Process runs to create summary
const runs = runsData.workflow_runs || [];
const summary = {
total_runs: runs.length,
failed: 0,
@@ -148,13 +158,13 @@ server.tool(
});
// Get jobs for this workflow run
const { data: jobsData } = await client.actions.listJobsForWorkflowRun({
const jobs = await listWorkflowJobs(client, {
owner: REPO_OWNER!,
repo: REPO_NAME!,
run_id,
});
const processedJobs = jobsData.jobs.map((job: any) => {
const processedJobs = jobs.map((job: any) => {
// Extract failed steps
const failedSteps = (job.steps || [])
.filter((step: any) => step.conclusion === "failure")
@@ -202,6 +212,40 @@ server.tool(
},
);
export async function downloadJobLog(
client: Octokit,
params: { owner: string; repo: string; job_id: number },
runnerTemp: string,
timeoutMs: number = DOWNLOAD_JOB_LOG_TIMEOUT_MS,
): Promise<{ path: string; size_bytes: number }> {
const controller = new AbortController();
const timeoutHandle = setTimeout(() => controller.abort(), timeoutMs);
try {
const response = await client.actions.downloadJobLogsForWorkflowRun({
owner: params.owner,
repo: params.repo,
job_id: params.job_id,
request: { signal: controller.signal },
});
const logsText = response.data as unknown as string;
const logsDir = `${runnerTemp}/github-ci-logs`;
await mkdir(logsDir, { recursive: true });
const logPath = `${logsDir}/job-${params.job_id}.log`;
await writeFile(logPath, logsText, "utf-8");
return {
path: logPath,
size_bytes: Buffer.byteLength(logsText, "utf-8"),
};
} finally {
clearTimeout(timeoutHandle);
}
}
server.tool(
"download_job_log",
"Download job logs to disk",
@@ -215,24 +259,11 @@ server.tool(
baseUrl: GITHUB_API_URL,
});
const response = await client.actions.downloadJobLogsForWorkflowRun({
owner: REPO_OWNER!,
repo: REPO_NAME!,
job_id,
});
const logsText = response.data as unknown as string;
const logsDir = `${RUNNER_TEMP}/github-ci-logs`;
await mkdir(logsDir, { recursive: true });
const logPath = `${logsDir}/job-${job_id}.log`;
await writeFile(logPath, logsText, "utf-8");
const result = {
path: logPath,
size_bytes: Buffer.byteLength(logsText, "utf-8"),
};
const result = await downloadJobLog(
client,
{ owner: REPO_OWNER!, repo: REPO_NAME!, job_id },
RUNNER_TEMP,
);
return {
content: [
@@ -274,6 +305,8 @@ async function runServer() {
}
}
runServer().catch(() => {
process.exit(1);
});
if (import.meta.main) {
runServer().catch(() => {
process.exit(1);
});
}
+2 -2
View File
@@ -6,7 +6,7 @@ import { z } from "zod";
import { GITHUB_API_URL } from "../github/api/config";
import { Octokit } from "@octokit/rest";
import { updateClaudeComment } from "../github/operations/comments/update-claude-comment";
import { sanitizeContent } from "../github/utils/sanitizer";
import { redactSecrets, sanitizeContent } from "../github/utils/sanitizer";
// Get repository information from environment variables
const REPO_OWNER = process.env.REPO_OWNER;
@@ -55,7 +55,7 @@ server.tool(
const isPullRequestReviewComment =
eventName === "pull_request_review_comment";
const sanitizedBody = sanitizeContent(body);
const sanitizedBody = redactSecrets(sanitizeContent(body));
const result = await updateClaudeComment(octokit, {
owner,
+24
View File
@@ -0,0 +1,24 @@
import { z } from "zod";
/** Raw shape passed to `server.tool` for commit_files. */
export const commitFilesInputSchema = {
files: z
.array(z.string())
.describe(
'Array of file paths relative to repository root (e.g. ["src/main.js", "README.md"]). All files must exist locally.',
),
message: z.string().describe("Commit message"),
};
/** Raw shape passed to `server.tool` for delete_files. */
export const deleteFilesInputSchema = {
paths: z
.array(z.string())
.describe(
'Array of file paths to delete relative to repository root (e.g. ["src/old-file.js", "docs/deprecated.md"])',
),
message: z.string().describe("Commit message"),
};
export const commitFilesPayloadSchema = z.object(commitFilesInputSchema);
export const deleteFilesPayloadSchema = z.object(deleteFilesInputSchema);
+48 -151
View File
@@ -2,14 +2,19 @@
// GitHub File Operations MCP Server
import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
import { z } from "zod";
import { readFile, stat } from "fs/promises";
import { resolve } from "path";
import { constants } from "fs";
import fetch from "node-fetch";
import { GITHUB_API_URL } from "../github/api/config";
import { retryWithBackoff } from "../utils/retry";
import { isBinaryContent } from "./binary-detection";
import { validatePathWithinRepo } from "./path-validation";
import { updateGitReference } from "./update-git-reference";
import { encodeBranchNameForUrl } from "../github/operations/comments/common";
import {
commitFilesInputSchema,
deleteFilesInputSchema,
} from "./github-file-ops-schemas";
type GitHubRef = {
object: {
@@ -62,7 +67,9 @@ async function getOrCreateBranchRef(
githubToken: string,
): Promise<string> {
// Try to get the branch reference
const refUrl = `${GITHUB_API_URL}/repos/${owner}/${repo}/git/refs/heads/${branch}`;
// git/ref is an exact lookup. git/refs returns an array of prefix matches when the
// branch doesn't exist, so pr-5 would find pr-50.
const refUrl = `${GITHUB_API_URL}/repos/${owner}/${repo}/git/ref/heads/${encodeBranchNameForUrl(branch)}`;
const refResponse = await fetch(refUrl, {
headers: {
Accept: "application/vnd.github+json",
@@ -83,7 +90,7 @@ async function getOrCreateBranchRef(
const baseBranch = process.env.BASE_BRANCH!;
// Get the SHA of the base branch
const baseRefUrl = `${GITHUB_API_URL}/repos/${owner}/${repo}/git/refs/heads/${baseBranch}`;
const baseRefUrl = `${GITHUB_API_URL}/repos/${owner}/${repo}/git/ref/heads/${encodeBranchNameForUrl(baseBranch)}`;
const baseRefResponse = await fetch(baseRefUrl, {
headers: {
Accept: "application/vnd.github+json",
@@ -115,7 +122,7 @@ async function getOrCreateBranchRef(
const defaultBranch = repoData.default_branch;
// Try default branch
const defaultRefUrl = `${GITHUB_API_URL}/repos/${owner}/${repo}/git/refs/heads/${defaultBranch}`;
const defaultRefUrl = `${GITHUB_API_URL}/repos/${owner}/${repo}/git/ref/heads/${encodeBranchNameForUrl(defaultBranch)}`;
const defaultRefResponse = await fetch(defaultRefUrl, {
headers: {
Accept: "application/vnd.github+json",
@@ -196,14 +203,7 @@ async function getFileMode(filePath: string): Promise<string> {
server.tool(
"commit_files",
"Commit one or more files to a repository in a single commit (this will commit them atomically in the remote repository)",
{
files: z
.array(z.string())
.describe(
'Array of file paths relative to repository root (e.g. ["src/main.js", "README.md"]). All files must exist locally.',
),
message: z.string().describe("Commit message"),
},
commitFilesInputSchema,
async ({ files, message }) => {
const owner = REPO_OWNER;
const repo = REPO_NAME;
@@ -258,17 +258,14 @@ server.tool(
// Get the proper file mode based on file permissions
const fileMode = await getFileMode(fullPath);
// Check if file is binary (images, etc.)
const isBinaryFile =
/\.(png|jpg|jpeg|gif|webp|ico|pdf|zip|tar|gz|exe|bin|woff|woff2|ttf|eot)$/i.test(
relativePath,
);
// Check if the file is binary by inspecting its contents. An
// extension allowlist used to decide this, which corrupted every
// binary type that wasn't on the list.
const fileContent = await readFile(fullPath);
if (isBinaryFile) {
if (isBinaryContent(fileContent)) {
// For binary files, create a blob first using the Blobs API
const binaryContent = await readFile(fullPath);
// Create blob using Blobs API (supports encoding parameter)
// (supports the encoding parameter)
const blobUrl = `${GITHUB_API_URL}/repos/${owner}/${repo}/git/blobs`;
const blobResponse = await fetch(blobUrl, {
method: "POST",
@@ -279,7 +276,7 @@ server.tool(
"Content-Type": "application/json",
},
body: JSON.stringify({
content: binaryContent.toString("base64"),
content: fileContent.toString("base64"),
encoding: "base64",
}),
});
@@ -302,12 +299,11 @@ server.tool(
};
} else {
// For text files, include content directly in tree
const content = await readFile(fullPath, "utf-8");
return {
path: relativePath,
mode: fileMode,
type: "blob",
content: content,
content: fileContent.toString("utf-8"),
};
}
}),
@@ -365,57 +361,13 @@ server.tool(
const newCommitData = (await newCommitResponse.json()) as GitHubNewCommit;
// 6. Update the reference to point to the new commit
const updateRefUrl = `${GITHUB_API_URL}/repos/${owner}/${repo}/git/refs/heads/${branch}`;
// We're seeing intermittent 403 "Resource not accessible by integration" errors
// on certain repos when updating git references. These appear to be transient
// GitHub API issues that succeed on retry.
await retryWithBackoff(
async () => {
const updateRefResponse = await fetch(updateRefUrl, {
method: "PATCH",
headers: {
Accept: "application/vnd.github+json",
Authorization: `Bearer ${githubToken}`,
"X-GitHub-Api-Version": "2022-11-28",
"Content-Type": "application/json",
},
body: JSON.stringify({
sha: newCommitData.sha,
force: false,
}),
});
if (!updateRefResponse.ok) {
const errorText = await updateRefResponse.text();
// Provide a more helpful error message for 403 permission errors
if (updateRefResponse.status === 403) {
const permissionError = new Error(
`Permission denied: Unable to push commits to branch '${branch}'. ` +
`Please rebase your branch from the main/master branch to allow Claude to commit.\n\n` +
`Original error: ${errorText}`,
);
throw permissionError;
}
// For other errors, use the original message
const error = new Error(
`Failed to update reference: ${updateRefResponse.status} - ${errorText}`,
);
// For non-403 errors, fail immediately without retry
console.error("Non-retryable error:", updateRefResponse.status);
throw error;
}
},
{
maxAttempts: 3,
initialDelayMs: 1000, // Start with 1 second delay
maxDelayMs: 5000, // Max 5 seconds delay
backoffFactor: 2, // Double the delay each time
},
);
await updateGitReference({
owner,
repo,
branch,
sha: newCommitData.sha,
githubToken,
});
const simplifiedResult = {
commit: {
@@ -461,14 +413,7 @@ server.tool(
server.tool(
"delete_files",
"Delete one or more files from a repository in a single commit",
{
paths: z
.array(z.string())
.describe(
'Array of file paths to delete relative to repository root (e.g. ["src/old-file.js", "docs/deprecated.md"])',
),
message: z.string().describe("Commit message"),
},
deleteFilesInputSchema,
async ({ paths, message }) => {
const owner = REPO_OWNER;
const repo = REPO_NAME;
@@ -479,21 +424,18 @@ server.tool(
throw new Error("GITHUB_TOKEN environment variable is required");
}
// Convert absolute paths to relative if they match CWD
const cwd = process.cwd();
const processedPaths = paths.map((filePath) => {
if (filePath.startsWith("/")) {
if (filePath.startsWith(cwd)) {
// Strip CWD from absolute path
return filePath.slice(cwd.length + 1);
} else {
throw new Error(
`Path '${filePath}' must be relative to repository root or within current working directory`,
);
}
}
return filePath;
});
// Validate all paths are within the repository root and normalize them to
// repo-relative paths for the git tree entries. This mirrors the validation
// already performed by the commit_files tool and rejects path traversal
// ("../") and symlinked escapes as defense-in-depth.
const resolvedRepoDir = resolve(REPO_DIR);
const processedPaths = await Promise.all(
paths.map(async (filePath) => {
await validatePathWithinRepo(filePath, REPO_DIR);
const normalizedPath = resolve(resolvedRepoDir, filePath);
return normalizedPath.slice(resolvedRepoDir.length + 1);
}),
);
// 1. Get the branch reference (create if doesn't exist)
const baseSha = await getOrCreateBranchRef(
@@ -580,58 +522,13 @@ server.tool(
const newCommitData = (await newCommitResponse.json()) as GitHubNewCommit;
// 6. Update the reference to point to the new commit
const updateRefUrl = `${GITHUB_API_URL}/repos/${owner}/${repo}/git/refs/heads/${branch}`;
// We're seeing intermittent 403 "Resource not accessible by integration" errors
// on certain repos when updating git references. These appear to be transient
// GitHub API issues that succeed on retry.
await retryWithBackoff(
async () => {
const updateRefResponse = await fetch(updateRefUrl, {
method: "PATCH",
headers: {
Accept: "application/vnd.github+json",
Authorization: `Bearer ${githubToken}`,
"X-GitHub-Api-Version": "2022-11-28",
"Content-Type": "application/json",
},
body: JSON.stringify({
sha: newCommitData.sha,
force: false,
}),
});
if (!updateRefResponse.ok) {
const errorText = await updateRefResponse.text();
// Provide a more helpful error message for 403 permission errors
if (updateRefResponse.status === 403) {
console.log("Received 403 error, will retry...");
const permissionError = new Error(
`Permission denied: Unable to push commits to branch '${branch}'. ` +
`Please rebase your branch from the main/master branch to allow Claude to commit.\n\n` +
`Original error: ${errorText}`,
);
throw permissionError;
}
// For other errors, use the original message
const error = new Error(
`Failed to update reference: ${updateRefResponse.status} - ${errorText}`,
);
// For non-403 errors, fail immediately without retry
console.error("Non-retryable error:", updateRefResponse.status);
throw error;
}
},
{
maxAttempts: 3,
initialDelayMs: 1000, // Start with 1 second delay
maxDelayMs: 5000, // Max 5 seconds delay
backoffFactor: 2, // Double the delay each time
},
);
await updateGitReference({
owner,
repo,
branch,
sha: newCommitData.sha,
githubToken,
});
const simplifiedResult = {
commit: {
+14 -3
View File
@@ -4,7 +4,8 @@ import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js"
import { appendFileSync } from "fs";
import { z } from "zod";
import { createOctokit } from "../github/api/client";
import { sanitizeContent } from "../github/utils/sanitizer";
import { redactSecrets, sanitizeContent } from "../github/utils/sanitizer";
import { removeBufferedComment } from "./inline-comment-buffer";
// Get repository and PR information from environment variables
const REPO_OWNER = process.env.REPO_OWNER;
@@ -97,8 +98,8 @@ server.tool(
const repo = REPO_NAME;
const pull_number = parseInt(PR_NUMBER, 10);
// Sanitize the comment body to remove any potential GitHub tokens
const sanitizedBody = sanitizeContent(body);
// Sanitize the comment body to remove potential prompt injections and redact secrets
const sanitizedBody = redactSecrets(sanitizeContent(body));
// Validate that either line or both startLine and line are provided
if (!line && !startLine) {
@@ -180,6 +181,16 @@ server.tool(
const result = await octokit.rest.pulls.createReviewComment(params);
// The comment is now live. Drop any buffered copy of it so the
// post-session replay step cannot post it a second time (the model often
// re-issues a buffered call with confirmed=true after the buffer reply).
if (CLASSIFY_ENABLED) {
removeBufferedComment(
{ path, line, startLine, body: sanitizedBody },
BUFFER_PATH,
);
}
return {
content: [
{
+54
View File
@@ -0,0 +1,54 @@
import { existsSync, readFileSync, writeFileSync } from "fs";
export type BufferedCommentMatch = {
path: string;
line?: number;
startLine?: number;
body: string;
};
/**
* Remove any buffered inline comment that matches an already-posted comment.
*
* When a comment is posted live (confirmed=true), an earlier buffered copy of
* the same comment must be dropped so the post-session replay step does not
* post it a second time. The model frequently re-issues a buffered call with
* confirmed=true after reading the "Set confirmed=true to post immediately"
* reply; previously the original buffered entry was left behind and replayed,
* producing duplicate inline comments.
*
* Entries are matched on path, line, startLine and body. Lines that cannot be
* parsed are kept untouched.
*/
export function removeBufferedComment(
match: BufferedCommentMatch,
bufferPath: string,
): void {
if (!existsSync(bufferPath)) {
return;
}
const remaining = readFileSync(bufferPath, "utf8")
.split("\n")
.filter((line) => line.trim() !== "")
.filter((line) => {
let entry: BufferedCommentMatch;
try {
entry = JSON.parse(line);
} catch {
// Keep anything we cannot parse rather than silently dropping it.
return true;
}
const isSameComment =
entry.path === match.path &&
entry.line === match.line &&
entry.startLine === match.startLine &&
entry.body === match.body;
return !isSameComment;
});
writeFileSync(
bufferPath,
remaining.length > 0 ? remaining.join("\n") + "\n" : "",
);
}

Some files were not shown because too many files have changed in this diff Show More