[FEATURE] CI failure auto-diagnosis for evolution PRs

[acipat]

Feature Description

Problem Statement

The evolution pipeline can open pull requests but cannot continue working on them once CI reports a failure. Agent-generated PRs sit open when a single CI shard fails — there is no automated step that reads the failing job log, extracts the concrete error (pytest traceback, lint violation, type error), and turns it into an actionable fix or child issue.

Proposed Solution

Add a CI failure auto-diagnosis stage that:

1. Polls the fork's open PRs via the GitHub Actions API (gh run list, gh run view --log-failed).
2. Fetches the failing job log and extracts the concrete error class and message.
3. For small safe fixes (e.g., a lint violation, a missing import), pushes a follow-up commit to the PR branch.
4. For larger failures, creates a focused child issue linking the PR and the extracted error, so the next implementation cycle can pick it up.

This differs from the broader evaluation harness (#491) because it targets real CI outcomes for real code changes, not synthetic benchmarks. It directly unblocks the merge stage by converting stuck PR states into actionable next steps.

Value Proposition

• Impact: 0.9
• Effort: 0.5
• Priority Score: 1.44

Research Evidence

• OpenHands / SWE-agent "verify-before-submit" workflows (shift verification left)
• GitHub Actions API log-parsing practices (gh run view --log-failed)

Implementation Plan

1. New skill evolution-ci-diagnosis + cron job entry.
2. Step 1: list open PRs and their latest CI run status.
3. Step 2: fetch failing log shard, extract error via a bounded regex/parser.
4. Step 3: classify failure as "trivial-fixable" vs "needs-child-issue".
5. Step 4: push fix commit OR create child issue with extracted context.
6. Reuse existing terminal/web toolsets — no new core model tools.

Success Criteria

• Open PRs with failed CI are automatically triaged within one cron cycle.
• Trivial CI failures (lint, import) get auto-fixed commits.
• Complex failures produce a child issue with the extracted error context.
• No increase in manual triage burden for maintainers.

[Da-Mikey]

Accepted — PR #600 implements CI failure auto-diagnosis for evolution PRs. evolution_ci_diagnosis.py polls open PRs every 30 minutes, fetches failing CI logs, classifies failures, and either auto-fixes (trivial lint) or creates child issues (complex failures).

[Lexus2016]

Blocked at integration review: the PR #600 implementation cannot actually read CI status.

What is wrong: scripts/evolution_ci_diagnosis.py calls "gh pr view --json recentCheckRuns" to discover the latest CI run. The gh CLI in the target environment does not support the "recentCheckRuns" JSON field, so fetch_latest_ci_run() always returns None and the script silently performs no diagnosis/fix/issue creation.

Definition of done:

1. Use a supported method to get CI status for a PR: either parse the text output of "gh pr checks --repo Lexus2016/hermes-agent-evolution", or query the GitHub REST API (repos/Lexus2016/hermes-agent-evolution/commits//check-runs and /actions/runs) with urllib/httpx.
2. Provide a unit test with an injected runner that returns realistic check-run JSON and asserts that a failed run is detected and its log is fetched/classified.
3. Keep the cron YAML and reporting behavior; only fix the retrieval path.

Re-open a fresh PR referencing #577 once the above is implemented.