… and PID-based lock

…sult block

…es after rebase

… running)

…ding

…ariable in Step 4

… and new label state machine

…ude, timeout kill, detailed logging)

…p comments, auto-rebase, remove status board

…ushing to any branch

…, dedup logic

…om timeout kills

…t found

…-fix in Step 3b

…o 60s

- Add --output-format stream-json --verbose to capture structured output
- Extract session_id from first JSON message and log with each iteration
- Change process poll interval from 5s to 60s to reduce unnecessary CPU usage

…tion log

…k PRs

Each EXIT point in the skill now logs a [pr-automation:exit] line with
action, PR number, and reason. The daemon script parses this line and
duration_ms from the session log after each Claude run, producing a
single-line summary in the main log. Falls back to ELAPSED if
duration_ms is unavailable.

Worst-case exit detection latency was 60s, causing up to 90s gap
between iterations (60s poll + 30s sleep).  is negligible
overhead, so 10s is a reasonable balance.

LOW issues are non-blocking and low priority — skip them during automated
fix runs. Add a note in the verification report explaining why they are skipped.

Add [pr-automation:skip] log markers to all 'find next PR' paths in the
skill (ci_running, ci_failure_dedup, merge_unknown, conflict_dedup), and
update the daemon script to collect all exit/skip markers instead of only
the last one, printing multi-action summaries when a session processes
multiple PRs.

…ation section

- Fix skill table: replace cron/Chinese description with accurate English
- Fix run mode: daemon loop, not cron job
- Fix label name: bot:needs-fix -> bot:ready-to-fix
- Remove operational 'how to block' detail (covered in pr-automation.md)

- Add NEEDS_HUMAN_REVIEW gate: PRs with >50 files changed or matching
  CRITICAL_PATH_PATTERN skip auto-merge and get bot:ready-to-merge instead
- Add bot:ready-to-merge label for PRs that are code-clean but need
  human confirmation before merging (distinct from bot:needs-human-review)
- Clear CRITICAL_PATH_PATTERN (empty by default, define when needed)
- Apply NEEDS_HUMAN_REVIEW check in both APPROVED path (Step 7) and
  post-fix path (Step 3b)
- Sync IS_CRITICAL_PATH detection in pr-review to use CRITICAL_PATH_PATTERN
- Update docs/conventions/pr-automation.md and AGENTS.md with new label

- Move REPO_DIR definition before LOG_FILE so it can be referenced
- Default LOG_FILE to <repo>/logs/pr-automation-{date}.log (daily rotation)
- Move PID_FILE to <repo>/logs/pr-automation-daemon.pid
- Auto-create logs/ directory with mkdir -p on startup

- Replace repo-relative logs/ with ~/Library/Logs/AionUi (macOS convention)
- Add LOG_DIR env var to override the log directory at startup
- LOG_FILE env var still overrides the full path if needed
- PID file moves alongside logs under LOG_DIR

…Step 6

- Detect isCrossRepository and maintainerCanModify in pre-flight checks
- For fork PRs with canModify=true: use gh pr checkout + push to fork remote
- For fork PRs with canModify=false (fallback):
  - Create bot/fix-pr-N branch on main repo with fork code + fixes
  - Open replacement PR with bot:done label and auto-merge enabled
  - Immediately close original fork PR with explanatory comment
- Update pr-automation Step 3b to detect fork fallback (PR state=CLOSED)
  and exit early instead of running merge gate on closed PR

…ed PRs

- Add Step 0 to wake up snoozed PRs when new commits are detected
- Add bot:ci-waiting to Step 3 skip conditions
- Step 4 ci_failure_dedup now swaps bot:reviewing → bot:ci-waiting instead of plain skip
- Clarify that CI failure check covers any job (required or not)
- Clarify that CI running check covers required jobs only
- Align pr-review Step 2 failure detection to cover all jobs, not just required ones
- Update docs/conventions/pr-automation.md and AGENTS.md with new label

…rmal queue exhausted

Previously Step 0 always checked all bot:ci-waiting PRs upfront, making
N API calls per session even when other eligible PRs existed. Now the
wake-up check only runs as a fallback when no eligible PR is found in
the normal queue, reducing unnecessary API overhead in the common case.

…early exit

Previously, a BEHIND branch caused an early EXIT in Step 4.5, wasting
one full CI cycle before review even started.

Now review runs first regardless of BEHIND status. update-branch is
triggered only when the PR is approved (or fixed) and ready to merge,
reducing CI wait cycles from 2 to 1 for the CONDITIONAL path.

…_LOOKBACK env var

Default remains 7 days. Override by setting PR_DAYS_LOOKBACK=N before starting
the daemon (e.g. PR_DAYS_LOOKBACK=30 ./scripts/pr-automation.sh).

In the update-branch path (APPROVED+BEHIND or fix+BEHIND), the PR ends
up with no bot: labels after triggering update-branch. The next session
would pick it up as a fresh PR and run a full review again.

Add a pre-check at the start of Step 6: if an existing pr-review-bot
comment is newer than the latest commit, load its cached conclusion and
skip directly to Step 7 instead of running pr-review again.

…eview check

When checking if a cached review is still valid, the previous logic used
the last commit time regardless of whether it was an update-branch merge
commit. GitHub's update-branch API creates a merge commit *after* the
review comment is posted, causing a false 'new commits since review'
detection and triggering an unnecessary re-review.

Fix: filter out commits whose messageHeadline matches
'^Merge branch '<base>' into ' before taking the last committedDate.
These commits are automatically generated when syncing the PR branch
with base and contain no new author code.

…automation-v2

…erge flow

Disabled 'Require branches to be up to date' in GitHub branch protection
(strict: false), so BEHIND branches can be merged directly via --squash --auto.
Removed all three BEHIND check blocks from the skill.

- pr-review: LOW-only issues now emit CONCLUSION=APPROVED in automation
  mode (was CONDITIONAL), preventing a wasteful fix session since pr-fix
  skips LOW issues entirely
- pr-fix: add early abort when all issues are LOW after filtering, guard
  against empty commit / stuck bot:fixing label when called after a
  LOW-only review
- pr-fix: unify commit message wording in Mandatory Rules to match Step 6
  template ('Review follow-up for #N')

…detection

- Step 2: codecov/patch and codecov/project are informational: true in
  codecov.yml and must not trigger CI_FAILED conclusion or block review;
  excluded from 情形 1 full-pass check and 情形 3 failure check
- Step 9: add codecov/patch FAILURE as LOW-level test coverage signal in
  the testing review dimension (patch target is 50%, informational only)

…ure detection

- Step 3b re-check CI: Any job FAILURE/CANCELLED → non-informational only
- Step 4 CI check: both the full-pass condition and the failure-trigger row
  now exclude codecov/* (informational: true in codecov.yml)