Safe outputs returning "session not found" on 37m run

[dsyme]

This run failed to produce any safe outputs, with the MCP erroring with "session not found" errors.

I've seen this before on runs over 1h, but this one is 37m so am investigating it more specifically

https://github.com/dsyme/raft-lean-squad/actions/runs/24753119605

✓ safeoutputs-create_pull_request(body: ## Summary

Task 5 (Proof Assistance) + ..., branch: lean-squad/run-71-task5-task8-2475311960..., title: Progress wf-preservation proofs (PR31–PR...)

◆ The safeoutputs MCP server keeps failing with "session not found". This seems to be a persistent infrastructure issue, not a transient one. Let me try the noop tool to at least signal completion.

✓ safeoutputs-noop(message: Run 71 complete. Task 5: 5 theorems (PR3...)

✓ safeoutputs-report_incomplete(reason: safeoutputs MCP server returning "sessio...)

◆ The safeoutputs MCP server is returning "session not found" for all calls — this is a persistent infrastructure failure, not a transient issue.

[dsyme]

A debugging session didn't really discover underlying causes. Analysis below

---

Findings from debugging run https://github.com/dsyme/raft-lean-squad/actions/runs/24753119605 using the gh-aw debug flow:

1. This run did not fail at the GitHub Actions level.

• Overall run conclusion was success.
• All main jobs completed successfully.
• So the “failure” is functional, not workflow-status failure.

2. Root cause is a Safe Outputs MCP session failure near the end of the agent step.

• In the agent stdio log, write actions started failing repeatedly with:
  • session not found
• Failed operations included:
  • push_repo_memory
  • create_pull_request
  • noop
  • report_incomplete
• The agent had already completed real work and committed changes, but could not publish outputs (PR/status issue updates).

3. Secondary problem: very high token/turn cost.

• 96 turns, about 9.1M tokens, about 37 minutes.
• Audit classified this run as resource-heavy and partially reducible.
• This increases chance of late-session infra/tooling instability and raises cost.

What this means:

• The workflow appeared green, but the important end-of-run publishing actions failed.
• That can hide incomplete automation from maintainers.

Recommended fixes (in priority order):

1. Add explicit fallback behavior in prompt instructions in lean-squad.md.

• If any safeoutputs write call returns session not found (or repeated transport errors), immediately fall back to GitHub CLI write path in bash:
  • create/update status issue
  • open PR
  • comment with run summary
• This prevents losing completed work.

2. Add a completion guard so silent no-op write failures cannot pass as success.

• In the workflow conclusion logic, fail the run if:
  • agent made code changes or created commits
  • but no expected safe outputs were emitted
• This converts hidden publish failures into visible failures.

3. Reduce agentic churn in Lean Squad prompt.

• Move repetitive discovery steps into deterministic pre-agent shell steps.
• Tighten proof-attempt loops with a retry cap per theorem.
• Remove retry sleeps; do immediate fallback on infrastructure errors.

4. Update runtime/tooling to latest gh-aw stack.

• This error pattern is consistent with MCP session lifecycle issues.
• Recompile/update workflow artifacts after pulling latest gh-aw changes.

If you want, I can now draft a concrete patch for lean-squad.md that adds:

1. Safeoutputs failure fallback policy text,
2. explicit “publish-or-fail” completion criteria,
3. reduced retry/loop guidance for Lean proof

attempts.

[pelikhan]

Sounds like a timeout in the MCP server.

[pelikhan]

@copilot inspect the MCP gateway logs for timeouts

See Findings from debugging run https://github.com/dsyme/raft-lean-squad/actions/runs/24753119605 using the gh-aw debug flow: