Spec a CI-green gate with bounded fix-and-retry loop for iteration acceptance

[Copilot]

Today an iteration is marked ✅ Accepted as soon as the sandbox-computed metric improves, but the agent sandbox often can't install project toolchains (tsc, cargo, pytest, …) so self-evaluation rubber-stamps broken commits onto the long-running branch with no revert or retry. Acceptance needs to be gated on real CI on the pushed HEAD, with fix-and-retry preferred over revert.

This PR is the spec change in workflows/autoloop.md only. The shared failure-signature extractor is deferred to sibling issue #34 (scheduler extraction).

Step 5 split into 5a / 5b / 5c

• 5a — Push and wait for CI: commit, push, ensure draft PR exists, then gh pr checks --watch and reduce all check-runs to a single success / failure / pending via the awk pipeline from the issue.
• 5b — Fix loop: on red, fetch logs, build a normalized failure signature, no-progress guard (pause if signature repeats), 5-attempt budget, 60-min wall-clock cap. Loops back to 5a after each fix push — never reverts.
• 5c — Accept: only on green; existing accept actions plus fix-attempt count when > 0.

status=$(gh pr checks "$PR" --json conclusion,state -q '.[] | (.conclusion // .state // "")' \
  | awk '
      BEGIN { r = "success" }
      /^(FAILURE|CANCELLED|TIMED_OUT|ACTION_REQUIRED|STARTUP_FAILURE|STALE)$/ { r = "failure" }
      /^(PENDING|QUEUED|IN_PROGRESS|WAITING|REQUESTED)$/ { if (r == "success") r = "pending" }
      END { print r }')

pending is explicitly handled: never accepted, re-runs the watch step, falls through to ci-timeout if it persists past the wall-clock cap.

Machine State vocabulary

• pause_reason: documents ci-fix-exhausted: <signature>, stuck in CI fix loop: <signature>, ci-timeout.
• recent_statuses: adds ci-fix-exhausted as the coarse bucket for any CI-gate failure mode (fine-grained reason lives in pause_reason). Existing parser accepts arbitrary status words — no scheduler code change needed.
• Iteration History entry format: optional CI fix attempts field.

Coordination

New subsection notes that PR-health-keeper / Evergreen workflows can pick up paused Autoloop PRs via the structured pause_reason field — same handoff as for human-authored PRs.

Preserved unchanged

The "metric did not improve" and "evaluation could not run" branches of Step 5 are untouched.

[mrjf]

@copilot resolve the merge conflicts in this pull request

[Copilot]

@copilot resolve the merge conflicts in this pull request

Merged origin/main and resolved the conflicts in workflows/autoloop.md in 83b791f. Adopted main's "program issue" terminology (the steering-issue/experiment-log collapse from #41) while preserving all the CI-gate spec additions (5a/5b/5c, fix-attempt count, PR-health-keeper coordination subsection). All 109 tests still pass.

Warning

Firewall rules blocked me from connecting to one or more addresses (expand for details)

I tried to connect to the following addresses, but was blocked by firewall rules:

• https://api.github.com/repos/bogus.invalid/bogus/issues
  • Triggering command: /usr/bin/python /usr/bin/python /home/REDACTED/work/autoloop/autoloop/workflows/scripts/autoloop_scheduler.py (http block)

If you need me to access, download, or install something from one of these locations, you can either:

• Configure Actions setup steps to set up my environment, which run before the firewall is enabled
• Add the appropriate URLs or hosts to the custom allowlist in this repository's Copilot coding agent settings (admins only)