Monitor CI status on clayde-opened PRs and self-fix failing pipelines

[ClaydeCode]

Problem

When clayde opens a PR and the GitHub Actions pipeline fails, clayde currently does nothing — it only reacts to human review comments. A failing CI is effectively a blocking review signal that should be handled autonomously.

Desired behaviour

During the pr_open phase, in addition to checking for PR reviews, clayde should:

1. Fetch the CI status for the PR's head commit (GitHub Checks API)
2. If any required check has failed and clayde has not yet attempted a fix for that run:
   • Invoke Claude with the failing job logs + current branch state
   • Claude analyzes the failure, pushes a fix commit to the branch
   • Clayde posts an issue comment summarising what failed and what was changed
3. Re-check on the next cycle — if CI is still failing after a fix attempt, retry (up to a configurable max, e.g. 3 attempts per PR)
4. If CI is green (or no required checks exist), continue existing review-monitoring logic unchanged

Implementation notes

• New helper in github.py: get_check_runs(g, owner, repo, ref) — wraps the Checks API to return failed check runs for a given commit SHA
• New Jinja2 prompt template fix_ci.j2 — includes: issue context, PR branch, list of failed checks with their log URLs, instruction to fetch logs via gh run view --log-failed and push a fix
• State: track last_ci_fix_attempt_sha in issue state — only attempt a fix once per head SHA to avoid infinite loops on unfixable failures
• Max attempts: ci_fix_max_attempts setting (default 3), tracked as ci_fix_attempts in state; stop attempting and notify operator if exceeded
• Notification: ntfy alert when max attempts exceeded ("Clayde: CI still failing after N attempts")
• Scope: only act on required checks (same checks that block merging), ignore informational ones

[ClaydeCode]

Implemented autonomous CI monitoring and self-fixing for clayde-opened PRs (#86). Added get_check_runs() and get_required_check_names() to github.py (Checks API + branch-protection-required-check filtering), a new fix_ci task + fix_ci.j2 prompt that has Claude read failing job logs via gh run view --log-failed and push a fix, and orchestrator._handle_ci_fix() which — when a PR is open and there's no new human activity — detects failed required checks and invokes the fix task. Added a ci_fix_max_attempts setting (default 3) with per-SHA dedup (last_ci_fix_attempt_sha) and a per-PR attempt cap (ci_fix_attempts); the attempt is recorded before invoking Claude for loop-safety, and the operator is notified once via ntfy when the budget is exhausted (ci_fix_exhausted_notified). Green/pending CI falls through to normal review monitoring unchanged. Updated README, CLAUDE.md, and config.env.template. Added 30 tests; full suite passes (371). Opened PR #87 (#87).