[kyle-sexton]

No related issue: root-caused from the address-feedback stops on #9536, #9479 and #9483 today; this is the fix.

Summary

pr-unblock's address-feedback fixed VALID findings, then feedback-thread.sh record refused to record the reply: "the working tree changes .bash_profile, which is outside every kept thread's scope". The lane's Bash sandbox masks dotfiles in the checkout with /dev/null, git ls-files --others lists that device as untracked, and is_stub excuses only empty regular files. With no reply recorded, the reply job stopped feedback-invalid (no replies beside a patch, publish-feedback.mjs:404-406) or feedback-incomplete (:422-425), and a person had to finish every lane PR with a review finding.

PR	run	stop
#9536	40050576425	feedback-invalid
#9479	40005930064	feedback-invalid
#9483 r1	40014113360	feedback-incomplete (2 of 3)
#9483 r3	40048712108	feedback-incomplete (1 of 4)

Each run's reply quotes the .bash_profile refusal. The feedback-uncertain stops (#9483 r2, #9418) were correct escalations and are unchanged.

Fix

• plugins/source-control/skills/pull-request/scripts/feedback-thread.sh: the untracked loop skips a character device that is not a symlink, the same rule scripts/check-affected.sh:130-133 uses since fix(ci): skip device nodes when check-affected stages uncommitted work #8520. -L is tested first because -c follows symlinks, so an untracked symlink to /dev/null is still scope-checked and refused.
• feedback-thread.test.sh: a bwrap case with /dev/null mounted over .bash_profile (fails before the fix, passes after) and a symlink case that must still refuse.
• .changes/source-control/: patch fragment.

Verification

• bash plugins/source-control/skills/pull-request/scripts/feedback-thread.test.sh: 0 failures; case 79 fails on origin/main's script (exit 1) and passes with the fix.
• shellcheck on both files: clean.

Related

#9536, #9479, #9483, #8520

🤖 Generated with Claude Code

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-24T02:46:34.554867Z	42f9633	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[github-actions]

claude-security-review reviewed 42f9633eb7fdeb0809afcb0f514d752b6f7f7027: 0 finding(s), 0 as inline comments.

No security issues found in PR #9551.

The change makes the untracked-file scope check in feedback-thread.sh record skip character devices that are not symlinks. The skip is [[ -L $workspace/$path || ! -c $workspace/$path ]] || continue at plugins/source-control/skills/pull-request/scripts/feedback-thread.sh:183.

• Scope-check bypass: this is the only control touched. Writing a payload into a path that evades the check would need a character device in the checkout, and creating one needs root (mknod) or a sandbox-level bind mount. An agent or PR author working with file-write access can't do either. A device node also carries no file content, so skipping it can't hide a code change.
• Symlink handling: -L is tested before -c, and -c follows symlinks. An untracked symlink to /dev/null, or to any other device, is still scope-checked and refused. The new test covers this case.
• Path handling: $path comes from git ls-files -z and is always quoted. No new injection, path-traversal or secret-handling surface is added.
• Other files: the test file and the .changes fragment are inert for this review.
• Instruction surface and manifests: no standing instruction was deleted or softened, and no plugin manifest changed, so there are no manifest <field> lines.
• PR context: the PR description contains no embedded directives.

critical-findings: 0

[github-actions]

claude-review reviewed 42f9633eb7fdeb0809afcb0f514d752b6f7f7027: 0 finding(s), 0 as inline comments.

No findings on this pull request.

What I checked

• feedback-thread.sh:183 skips an untracked path only when it is a character device and not a symlink. The -L test runs first, which matters because -c follows symlinks, so an untracked symlink to /dev/null is still scope-checked and refused. The new line sits before is_stub, so stub handling is unchanged.
• The skip applies only in the untracked-file loop of the VALID path. Tracked changes from git diff --name-only are still checked. Creating a character device takes root, so a reply can't use the skip to hide a real change.
• workspace is set and required at the top of the script (line 42), so the $workspace/$path reference is safe. Paths come from ls-files -z, so quoting is correct.
• The tests cover both sides. One bwrap case mounts /dev/null over .bash_profile and expects exit 0. It is skipped when bwrap is unavailable. A second case uses an untracked symlink to /dev/null and expects exit 1.
• The changelog fragment uses bump: patch, which fits a bug fix with no interface change.
• There is no closing issue (the PR says so), no manifest change, and no hold class: no cost, no trust-gate files, and no CRITICAL security finding.

I read the diff and the surrounding script. I did not run the tests, so the author's claims of 0 failures and a clean shellcheck are unverified by me.