[aw-failures] [aw] Failure Investigator (6h) - Issue Group

[github-actions]

[aw] Failure Investigator (6h)

Parent issue for grouping related issues from [aw] Failure Investigator (6h).

Sub-issues are automatically linked below (max 64 per parent).

Workflow: [aw] Failure Investigator (6h)

• expires on May 8, 2026, 1:30 AM UTC

---

Update — 2026-04-30 ~19:00 UTC (run §25184286958)

Failures in last 6h window (39 runs, 4 true failures)

Cluster	Runs	Status	Tracking
Design Decision Gate GitHub MCP drops (~100s timeout)	25179263531, 25181104179, 25177070075	Recurring P1	New sub-issue created
Smoke Claude safe_outputs resolve_thread failure	25181816514	New P2	New sub-issue created
Design Decision Gate gate rejection (expected)	25181238397	Expected behavior	No action needed

Root Causes Confirmed

DDG MCP drops: All three runs show identical signature — GitHub MCP HTTP connection drops at ~100–110s uptime with Terminal connection error 1/3 → 2/3 → exit 1. The DDG agent takes 8–13 minutes total, so the MCP connection always drops before the agent finishes. The ~100s timeout appears to be a hard limit in the GitHub MCP server connection.

Smoke Claude: Agent completed all 19 smoke tests successfully. The resolve_pull_request_review_thread safe output fails because thread PRRT_kwDOPc1QR85-0kYa on PR #29360 was already resolved. 1 of 11 safe outputs failed, causing the safe_outputs step to return failure.

No stale issues closed

Open issues reviewed — #29355, #29279, #29275 are per-run auto-expiring reports; #29231/#29233 remain active tracking. No actionable closures warranted.

Generated by [aw] Failure Investigator (6h) · ● 397.8K · ◷

---

[aw] Failure Investigator (6h)

Parent issue for grouping related issues from [aw] Failure Investigator (6h).

Sub-issues are automatically linked below (max 64 per parent).

Workflow: [aw] Failure Investigator (6h)

• expires on May 8, 2026, 1:30 AM UTC

---

Update — 2026-04-30 ~19:00 UTC (run §25184286958)

Failures in last 6h window (39 runs, 4 true failures)

Cluster	Runs	Status	Tracking
Design Decision Gate GitHub MCP drops (~100s timeout)	25179263531, 25181104179, 25177070075	Recurring P1	New sub-issue created
Smoke Claude safe_outputs resolve_thread failure	25181816514	New P2	New sub-issue created
Design Decision Gate gate rejection (expected)	25181238397	Expected behavior	No action needed

Root Causes Confirmed

DDG MCP drops: All three runs show identical signature — GitHub MCP HTTP connection drops at ~100–110s uptime with Terminal connection error 1/3 → 2/3 → exit 1. The DDG agent takes 8–13 minutes total, so the MCP connection always drops before the agent finishes. The ~100s timeout appears to be a hard limit in the GitHub MCP server connection.

Smoke Claude: Agent completed all 19 smoke tests successfully. The resolve_pull_request_review_thread safe output fails because thread PRRT_kwDOPc1QR85-0kYa on PR #29360 was already resolved. 1 of 11 safe outputs failed, causing the safe_outputs step to return failure.

No stale issues closed

Open issues reviewed — #29355, #29279, #29275 are per-run auto-expiring reports; #29231/#29233 remain active tracking. No actionable closures warranted.

Generated by [aw] Failure Investigator (6h) · ● 397.8K

---

Update — 2026-05-01 ~01:30 UTC (run §25197698525)

Failures in last 6h window (21 runs, 2 true failures)

Cluster	Runs	Status	Tracking
DDG error_max_turns on complex PRs (bash safeoutputs not authorized)	§25196075860, §25196318154	New P1 — distinct from prior MCP-drop pattern	New sub-issue created

Root Cause Confirmed — Different from prior DDG MCP-drop failures

Both new DDG failures are caused by error_max_turns, NOT an active GitHub MCP connection drop. Evidence from agent-stdio.log:

terminal_reason: max_turns
errors: ["Reached maximum number of turns (12)"]
num_turns: 13

The GitHub MCP "drop" messages in the failure reporter appear only during post-session cleanup (after the agent exited). This distinguishes these from runs 25177070075/25178770914/etc. where the GitHub MCP dropped mid-session.

What actually happened (both runs):

• Agent successfully analyzed the PR diff on complex PRs (docs/variable changes) — 13 turns consumed
• ADR draft written and pushed to PR branch via MCP push_to_pull_request_branch (2 successful calls)
• Agent attempted bash safeoutputs add_comment to post the PR review comment → permission denied (not pre-authorized in DDG workflow)
• Agent hit max_turns (12) trying to find an alternative path
• Pre-step file /tmp/gh-aw/agent/adr-prefetch-summary.json was absent, costing extra analysis turns (was present in Apr 21 runs)

Successful DDG runs in same window — 4 runs all completed in 4–5 turns on simple fix PRs. No max_turns issue.

Historical context

Prior sub-issue #27470 (closed completed Apr 21) fixed max_turns from 5 → 7. Limit is now 12, but complex PRs with large diffs + missing pre-step file need 13+ turns. The adr-prefetch-summary.json pre-step appears to have regressed since April 21.

No stale issues closed

• [aw] Design Decision Gate 🏗️ failed #29355 (DDG auto-failure tracker): still open and actively receiving new failure comments — not stale
• [aw-failures] Smoke suite: 4 engines failing on every run for 5+ days — no root-cause fix deployed #29231/[aw-failures] Fix: Gemini API key invalid — smoke-gemini fails on every run (5+ days) #29233 (smoke suite / Gemini key): no smoke failures in this 6h window; issues remain valid
• [aw-failures] [aw] Failure Investigator (6h) - Issue Group #29232 (this issue): updated with this run's findings

References:

• §25196075860
• §25196318154
• §25197698525

Generated by [aw] Failure Investigator (6h)

Note

🔒 Integrity filter blocked 3 items

The following items were blocked because they don't meet the GitHub integrity level.

• question: why is COPILOT_GITHUB_TOKEN still required when using an external provider? #29379 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• docs: COPILOT_PROVIDER_* variables not documented and strict-mode allowlist not updated #29378 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".
• Replace repo-memory patch size limit with a rate limit over time #29310 list_issues: has lower integrity than agent requires. The agent cannot read data with integrity below "approved".

To allow these resources, lower min-integrity in your GitHub frontmatter:

tools:
  github:
    min-integrity: approved  # merged | approved | unapproved | none

Generated by [aw] Failure Investigator (6h) · ● 697.5K · ◷