Repository navigation
Description
Problem Statement
The Multi-Device Docs Tester workflow has failed on every scheduled run since 2026-05-04 (4 consecutive days). No actual testing work ever executes — the harness defers the session during initialization, then fails to resume it on all 3 retry attempts with exit code 1.
Parent report: [aw] Failure Report 2026-05-07: Session Resume Regression + Schema Checker Max-Turns Escalation
Affected Runs
| Run ID | Date | Note |
|---|---|---|
| §25479583239 | 2026-05-07 | Latest |
| §25419750507 | 2026-05-06 | |
| §25360837055 | 2026-05-05 | |
| §25304397726 | 2026-05-04 | First failure |
Last successful run: §25271781842 (2026-05-03)
Root Cause
Harness session-resume error in Claude Code v2.1.126:
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] all 3 retries exhausted — giving up (exitCode=1)
[claude-harness] done: exitCode=1 totalDuration=3m 10s
Additional signal: [DEBUG] Prior session exited uncleanly: 6655b237-5666-42c3-b6bf-2e1473e87727 (v2.1.126)
The workflow defers its session during pre-agent initialization, but when the harness attempts to resume, the deferred-tool marker is missing or expired. All 3 retries hit the same error instantly (duration=0s stderr=221B).
Audit-diff confirms the failed run never reaches the actual test phase — no playwright MCP tool calls, no Android/Google domain traffic, compared to the last successful baseline §25245508593 which contacted accounts.google.com, android.clients.google.com, and mtalk.google.com.
Probable Root Cause
A regression around or after 2026-05-03 in:
- Harness deferred-session state management — tail-scan window too short, or marker not persisted correctly across job boundaries, OR
- A change to the workflow initialization sequence that altered when/how the session is deferred
The "Prior session exited uncleanly" message points to a lingering stale session ID that may conflict with each new run's resume attempt.
Proposed Remediation
- Investigate harness deferred-session logic in Claude Code v2.1.126 — specifically the tail-scan window and marker persistence between the pre-agent and agent jobs
- Check git history around 2026-05-03 for changes to
multi-device-docs-tester.lock.ymlor shared harness initialization code - Short-term workaround: Modify the workflow to avoid session deferral, or add a fallback that starts a fresh session when the deferred marker is missing
- Clean up stale session state for ID
6655b237-5666-42c3-b6bf-2e1473e87727
Success Criteria
- Multi-Device Docs Tester produces a successful conclusion on the next scheduled run
- No
No deferred tool marker founderrors in harness logs - Playwright and Android device testing phase reached (Google domains visible in firewall logs)
References:
- Failed run (latest): §25479583239
- Last successful run (diff baseline): §25245508593
- Engine: Claude Code v2.1.126, claude-sonnet-4-6
Generated by [aw] Failure Investigator (6h) · ● 1M · ◷
- expires on May 14, 2026, 8:09 AM UTC
Activity
github-actions commentedon May 8, 2026
🍪 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! 🍪