Review agent missed vouch-check missing `issues: write` permission (caught by qodo)

[fullsend-ai-retro]

What happened

On PR #32 (#32), the vouch-check workflow calls github.rest.issues.createComment() to post an explanatory comment when closing an unvouched contributor's PR. However, the workflow's permissions block only declares contents: read and pull-requests: write, omitting issues: write. This means the comment step will fail at runtime with a 403 error, leaving closed PRs with no explanation.

The qodo-code-review bot correctly identified this bug as finding #2 ("vouch-check lacks issues: write"), classifying it as a correctness issue. The fullsend-ai-review bot did not identify this finding despite reviewing the same workflow file and even noting other permission issues (unnecessary actions: write in stale.yml, unnecessary contents: read in vouch-check). The fullsend agent evaluated vouch-check permissions and found one expansion (contents: read) but missed the missing permission that would cause a runtime failure.

What could go better

The fullsend review agent's permission analysis was one-directional: it checked for unnecessary permissions (over-privileged) but did not verify that declared permissions satisfy all API calls in the code (under-privileged). The agent flagged contents: read as unused but failed to verify that issues: write was needed for createComment(). This is a consumer-producer analysis gap — the agent checked the permission manifest (producer) in isolation without cross-referencing it against the GitHub API calls (consumers) in the script.

Confidence: Medium. It's possible the review agent did consider this but rated it below threshold, or that the correctness sub-agent deprioritized it. However, the finding is absent from both the structured findings and the summary narrative, suggesting it was genuinely missed rather than filtered.

Proposed change

In the correctness sub-agent's review guidance (skills/pr-review/sub-agents/correctness.md or the code-review skill), add explicit instruction to verify API call permissions bidirectionally:

1. When reviewing GitHub Actions workflow files that use actions/github-script or direct API calls:

   • Check for over-privilege: Flag declared permissions not required by any API call (existing behavior).
   • Check for under-privilege: For each github.rest.* call in the script, verify the corresponding permission scope is declared. Common mappings: issues.createComment → issues: write, pulls.update → pull-requests: write, repos.getContent → contents: read.

2. Add this as an anchoring example in the permission review guidance: "A workflow declaring pull-requests: write but calling github.rest.issues.createComment() without issues: write will fail at runtime with HTTP 403."

Validation criteria

1. Re-run the review agent on PR docs: add CLAUDE.md, COMMITS.md, CONTRIBUTING.md, LICENSE, and repo workflows #32's diff (commit 7e7509f). The review should include a finding about the missing issues: write permission in vouch-check.yml in at least 3 out of 5 trials.
2. On a test PR with a correctly-permissioned workflow (all API calls covered by declared permissions), the review agent should not produce false-positive permission findings.
3. The review agent's permission analysis covers both directions: over-privilege (unnecessary permissions) and under-privilege (missing permissions for API calls).

---

Generated by retro agent from #32

[fullsend-ai-triage]

🤖 Finished Triage · ✅ Success · Started 10:37 PM UTC · Completed 10:41 PM UTC
Commit: e8381e3 · View workflow run →

[fullsend-ai-triage]

Triage Summary

Severity: Medium · Category: Feature (review agent enhancement)

Problem

The review agent's permission analysis for GitHub Actions workflows is one-directional — it flags unnecessary (over-privileged) permissions but does not verify that declared permissions cover all API calls in the workflow scripts. On PR #32, this caused the agent to miss that vouch-check.yml calls github.rest.issues.createComment() without declaring issues: write, a bug that would cause a runtime 403 error.

Root Cause

Neither the correctness sub-agent (skills/pr-review/sub-agents/correctness.md) nor the code-review skill (skills/code-review/SKILL.md) instruct the reviewer to cross-reference github.rest.* API calls against the workflow's permissions: block. The existing "Permission manifest changes" section focuses on evaluating excess capabilities, and the "Runtime mechanism checklist" checks producer-consumer paths generally but doesn't specifically address GitHub Actions permission scopes.

Recommended Fix

1. Correctness sub-agent (correctness.md): Extend the "Runtime mechanism checklist" or add a dedicated subsection for GitHub Actions permission verification. Instruct the reviewer to map each github.rest.* call to its required permission scope and verify it is declared.
2. Code-review skill (SKILL.md): Extend the "Permission manifest changes" section under Security to cover both directions — excess permissions AND missing permissions for API calls.
3. Include common API-to-permission mappings and the vouch-check example as an anchoring case.

Proposed Test Case

Validation:
- Re-run review agent on PR #32 diff (commit 7e7509f)
- Expected: finding about missing issues: write in vouch-check.yml
  in >= 3/5 trials

False-positive check:
- Review a workflow with correct permissions for all API calls
- Expected: no spurious permission findings

Related Issues

• Fix 4 unresolved bugs from PR #32: COMMITS.md, vouch-check.yml, stale.yml #44 — Fixes the actual vouch-check.yml bug (distinct: fixes the code, not the review agent)
• Review agent verdict contradicted its own analysis on PR #32: chose comment despite saying findings should be addressed before merge #45 — Review agent verdict inconsistency on PR docs: add CLAUDE.md, COMMITS.md, CONTRIBUTING.md, LICENSE, and repo workflows #32 (distinct: verdict selection, not analysis coverage)
• Add silent-failure severity escalation guidance to code-review skill and correctness sub-agent #33 — Silent-failure severity escalation (related but independent improvement)

---

Labels: This is a new feature request to enhance the review agent's permission analysis capabilities. Applying 'enhancement' and removing 'ready-for-triage' since triage is complete.