[aw-failures] [aw-fix] P1: PR Sous Chef — add_comment returns 401 Bad Credentials on PR #48593, failing Process Safe Outputs

[github-actions]

Problem

Fix the add_comment credential path in PR Sous Chef — it is 0-for-3 on PR #48593 in the last 68 minutes, and it fails only after a different write (update_pull_request / dismiss_pull_request_review) already succeeded against the same PR moments earlier.

Affected workflow / runs

• Workflow: .github/workflows/pr-sous-chef.lock.yml ("PR Sous Chef")
• Failing job: safe_outputs → step "Process Safe Outputs" (agent job itself succeeds every time)
• Target: PR Fix DefaultCopilotVersion toolcache bypass caused by compat.json drift #48593 (open, same-org, not a fork — so this is not a cross-fork permission issue)
• Runs (chronological):
  • §30384256680 — 2026-07-28T17:54:26Z — add_comment failed 401 (request id 1340:1AEE4B:72FE2D:186887D:6A68ECD1)
  • §30386933161 — 2026-07-28T18:26:44Z — add_comment failed 401 (request id 20C3:33FC32:A056B5:220844C:6A68F464)
  • §30389499577 — 2026-07-28T19:02:45Z — add_comment failed 401 (request id 1C28:2ABB8:EA167:3221C5:6A68FCD5)

Evidence

All three runs show the identical shape:

POST /repos/github/gh-aw/issues/48593/comments - 401 with id <req-id> in ~40-65ms
##[error]Failed to add comment: Bad credentials - https://docs.github.com/rest

In all three runs, a different safe output writing to the same PR (either update_pull_request or dismiss_pull_request_review) completed successfully in the same job execution, immediately before or after the failing add_comment call. This rules out a blanket token-revocation or repo-permission problem — the failure is specific to the add_comment code path/endpoint (issues/{n}/comments), not to authentication in general for this job.

Probable root cause

The add_comment handler likely mints or caches a GitHub App installation token separately from the other safe-output handlers (e.g., a stale/short-lived token reused across a multi-message batch, or a token scoped without issues:write while pull_requests:write is granted). A 401 (not 403) specifically points to an invalid/expired credential rather than a permissions/policy denial — distinguish this from the already-tracked #47753 (branch-from-base rejected by verified-signatures rule, a 422/permission-shaped failure) and #47659 (resolve_pull_request_review_thread failing) which are different safe-output code paths with different error shapes.

Proposed remediation

1. Audit the safe-outputs processor's token acquisition for add_comment vs. update_pull_request/dismiss_pull_request_review — confirm whether they use the same cached installation token instance or independently mint one.
2. If tokens are cached across the batch, verify freshness/expiry is checked before each API call, not just once at job start.
3. Add a retry-with-fresh-token on 401 for add_comment (matching the retry-exempt-status-codes config already present: 400,401,403,404,422 — note 401 is currently in the no-retry list, which may be masking a legitimately transient token-refresh case; confirm this is intentional).
4. Add a regression test that runs add_comment after another mutating safe output in the same batch against the same target PR.

Success criteria

• PR Sous Chef's add_comment step succeeds when run in the same batch as update_pull_request/dismiss_pull_request_review against the same PR, across at least 3 consecutive scheduled runs.
• No new 401 Bad Credentials errors for add_comment in safe_outputs job logs for PR Sous Chef.
  Related to [aw-failures] [aw] Failure Investigator (6h) - Issue Group #46171

Generated by 🔍 [aw] Failure Investigator (6h) · age00 · 140.3 AIC · ⌖ 26 AIC · ⊞ 5.2K · ◷

• expires on Aug 4, 2026, 11:23 AM UTC-08:00

---

Problem

Fix the add_comment credential path in PR Sous Chef — it is 0-for-3 on PR #48593 in the last 68 minutes, and it fails only after a different write (update_pull_request / dismiss_pull_request_review) already succeeded against the same PR moments earlier.

Affected workflow / runs

• Workflow: .github/workflows/pr-sous-chef.lock.yml ("PR Sous Chef")
• Failing job: safe_outputs → step "Process Safe Outputs" (agent job itself succeeds every time)
• Target: PR Fix DefaultCopilotVersion toolcache bypass caused by compat.json drift #48593 (open, same-org, not a fork — so this is not a cross-fork permission issue)
• Runs (chronological):
  • §30384256680 — 2026-07-28T17:54:26Z — add_comment failed 401 (request id 1340:1AEE4B:72FE2D:186887D:6A68ECD1)
  • §30386933161 — 2026-07-28T18:26:44Z — add_comment failed 401 (request id 20C3:33FC32:A056B5:220844C:6A68F464)
  • §30389499577 — 2026-07-28T19:02:45Z — add_comment failed 401 (request id 1C28:2ABB8:EA167:3221C5:6A68FCD5)

Evidence

All three runs show the identical shape:

POST /repos/github/gh-aw/issues/48593/comments - 401 with id <req-id> in ~40-65ms
##[error]Failed to add comment: Bad credentials - https://docs.github.com/rest

In all three runs, a different safe output writing to the same PR (either update_pull_request or dismiss_pull_request_review) completed successfully in the same job execution, immediately before or after the failing add_comment call. This rules out a blanket token-revocation or repo-permission problem — the failure is specific to the add_comment code path/endpoint (issues/{n}/comments), not to authentication in general for this job.

Update — still failing 2+ hours after filing (2026-07-28, later same day)

Not fixed. Two more add_comment 401 failures hit the identical safe_outputs → "Process Safe Outputs" code path, well after this issue was opened:

• §30400044871 — 2026-07-28T21:17:26Z — Process Safe Outputs failed: Failed to add comment: Bad credentials
• §30401927723 — 2026-07-28T21:45:07Z — Process Safe Outputs failed: Failed to add comment: Bad credentials (agent turn count 144 — high, but the failure is the same credential error, not a new symptom)

Fix priority stands at P1 — this is now confirmed recurring across 5 runs over ~4 hours, not a one-off. Prioritize the token-freshness check in proposed remediation step 2 first; it's the most likely culprit given the failure keeps recurring hours after the initial report.

Probable root cause

The add_comment handler likely mints or caches a GitHub App installation token separately from the other safe-output handlers (e.g., a stale/short-lived token reused across a multi-message batch, or a token scoped without issues:write while pull_requests:write is granted). A 401 (not 403) specifically points to an invalid/expired credential rather than a permissions/policy denial — distinguish this from the already-tracked #47753 (branch-from-base rejected by verified-signatures rule, a 422/permission-shaped failure) and #47659 (resolve_pull_request_review_thread failing) which are different safe-output code paths with different error shapes.

Proposed remediation

1. Audit the safe-outputs processor's token acquisition for add_comment vs. update_pull_request/dismiss_pull_request_review — confirm whether they use the same cached installation token instance or independently mint one.
2. If tokens are cached across the batch, verify freshness/expiry is checked before each API call, not just once at job start.
3. Add a retry-with-fresh-token on 401 for add_comment (matching the retry-exempt-status-codes config already present: 400,401,403,404,422 — note 401 is currently in the no-retry list, which may be masking a legitimately transient token-refresh case; confirm this is intentional).
4. Add a regression test that runs add_comment after another mutating safe output in the same batch against the same target PR.

Success criteria

• PR Sous Chef's add_comment step succeeds when run in the same batch as update_pull_request/dismiss_pull_request_review against the same PR, across at least 3 consecutive scheduled runs.
• No new 401 Bad Credentials errors for add_comment in safe_outputs job logs for PR Sous Chef.
  Related to [aw-failures] [aw] Failure Investigator (6h) - Issue Group #46171

Generated by 🔍 [aw] Failure Investigator (6h) · age00 · 140.3 AIC · ⌖ 26 AIC · ⊞ 5.2K · ◷

• expires on Aug 4, 2026, 11:23 AM UTC-08:00

Generated by 🔍 [aw] Failure Investigator (6h) · age00 · 121.4 AIC · ⌖ 20.5 AIC · ⊞ 5.2K · ◷

[github-actions]

🍪 Issue Monster selected this for Copilot

I've identified this issue as a good candidate for automated resolution and requested assignment to the Copilot coding agent.

If assignment succeeds, the Copilot coding agent will analyze the issue and create a pull request with the fix.

Om nom nom! 🍪

🍪 Om nom nom by Issue Monster · gpt54 · 2.74 AIC · ⌖ 5.86 AIC · ⊞ 6.8K · ◷