[aw-failures] [aw] Failure Report 2026-05-07: Session Resume Regression + Schema Checker Max-Turns Escalation

[github-actions]

Executive Summary

Two distinct failure clusters were identified in the 6-hour lookback window (2026-05-07 ~00:20–06:30 UTC). Both failures are recurring, escalating patterns that began around May 3–4 and require intervention.

Cluster	Workflow	Failure Mode	Run ID	Streak	Severity
A	Multi-Device Docs Tester	Session resume error — no work executed	§25479583239	4 consecutive days	P1
B	Schema Consistency Checker	Max turns (60) exceeded — truncated mid-run	§25479424655	4/5 last scheduled runs	P1

---

Failure Clusters

Cluster A — Multi-Device Docs Tester: Session Resume Failure

Pattern: Every scheduled run since 2026-05-04 fails before executing any work. All 3 harness retries fail with the same error immediately.

Root cause: Error: No deferred tool marker found in the resumed session — the harness defers the session during initialization but cannot locate the deferred tool marker on resume. The agent exits with code 1 on all attempts.

Evidence from raw logs (run 25479583239):

Harness error trace

Error: No deferred tool marker found in the resumed session. Either the session was not
deferred, the marker is stale (tool already ran), or it exceeds the tail-scan window.
Provide a prompt to continue the conversation.

[claude-harness] attempt 4: process exit event exitCode=1
[claude-harness] attempt 4: process closed exitCode=1 duration=0s stdout=0B stderr=221B hasOutput=true
[claude-harness] attempt 4 failed: exitCode=1 isOverloadedError=false isRateLimitError=false
                  isMaxTurnsExit=false hasOutput=true retriesRemaining=0
[claude-harness] all 3 retries exhausted — giving up (exitCode=1)
[claude-harness] done: exitCode=1 totalDuration=3m 10s

Also: [DEBUG] Prior session exited uncleanly: 6655b237-5666-42c3-b6bf-2e1473e87727 (v2.1.126)

Failure history (last 30 days):

Run ID	Date	Conclusion
§25479583239	2026-05-07	failure
§25419750507	2026-05-06	failure
§25360837055	2026-05-05	failure
§25304397726	2026-05-04	failure
§25271781842	2026-05-03	success
§25245508593	2026-05-02	success

Audit-diff vs last success (25479583239 vs 25245508593): The failed run never reaches playwright tools or Google Android domains (accounts.google.com, android.clients.google.com, mtalk.google.com) — confirmed zero work executed.

---

Cluster B — Schema Consistency Checker: Max Turns Escalation

Pattern: Starting 2026-05-03, the workflow consistently exhausts its 60-turn limit mid-tool_use, hitting exactly 61 turns before the harness terminates the run.

Root cause: Reached maximum number of turns (60) — error_max_turns, terminal_reason: max_turns. The harness correctly identifies this as non-retriable.

Evidence from raw logs (run 25479424655):

Exit result JSON

{"type":"result","subtype":"error_max_turns","duration_ms":424089,"is_error":true,
 "num_turns":61,"stop_reason":"tool_use","terminal_reason":"max_turns",
 "errors":["Reached maximum number of turns (60)"]}

Cost & resource impact: $2.20 per failed run, 61 turns, 9.5 minutes, 4.57M tokens — all wasted as no safe-output is emitted before truncation.

Failure history (scheduled runs only):

Run #	Run ID	Date	Conclusion
242	§25479424655	2026-05-07	failure (max turns)
241	§25419613447	2026-05-06	success
240	§25360601823	2026-05-05	failure (max turns)
239	§25304247413	2026-05-04	failure (max turns)
237	§25271731933	2026-05-03	failure (max turns)
236	§25245322929	2026-05-02	success

Before May 3: 21 consecutive successful scheduled runs (Apr 8 – May 2). Zero max-turns failures prior to this streak.

---

Existing Issue Correlation

GitHub issue API was unavailable (403) during this run. Existing issue coverage could not be verified. Sub-issues created below should be checked for duplicates.

---

Proposed Fix Roadmap

Priority	Issue	Fix
P1	Multi-Device Docs Tester session resume	Investigate harness deferred-session logic; check for regression in Claude Code v2.1.126
P1	Schema Consistency Checker max turns	Raise --max-turns to 80–100, or decompose the checklist into batched passes

---

Sub-issues Created

• Sub-issue: Multi-Device Docs Tester — session resume failure (4-day streak)
• Sub-issue: Schema Consistency Checker — max turns regression (May 3 onset)

References:

• §25479583239 Multi-Device Docs Tester failure
• §25479424655 Schema Consistency Checker failure
• §25245508593 Multi-Device Docs Tester last success (for diff baseline)

Generated by [aw] Failure Investigator (6h) · ● 1M · ◷

• expires on May 14, 2026, 8:09 AM UTC

[github-actions]

This issue was automatically closed because it expired on 2026-05-14T08:09:12.828Z.

Closed by Workflow