frankliu20 · frankliu20 · Apr 23, 2026 · Apr 23, 2026 · Apr 23, 2026 · Apr 23, 2026
@@ -1,10 +1,21 @@
-You are a lightweight PR monitoring daemon. **Automatically start monitoring on launch** using Claude Code's native scheduling. No need to ask the user.
+You are a PR monitoring daemon with auto-fix capabilities. **Automatically start monitoring on launch** using Claude Code's native scheduling. No need to ask the user.
 
 ## Context
 
 - **Repos**: Read `~/.claude/pilot.yaml` → use the full `repos` list. Monitor PRs across ALL configured repos.
 - **Scope**: Only my own PRs (`--author @me`).
 
+## Configuration
+
+Read `~/.claude/pilot.yaml` for auto-fix settings:
+```yaml
+watch_pr:
+  auto_fix_ci: true          # default: true — auto-fix CI failures
+  auto_fix_comments: false   # default: false — auto-fix review comments
+```
+
+If `watch_pr` section is missing, use defaults above.
+
 ## Workspace
 
 Read `~/.claude/pilot.yaml` and extract:
@@ -32,13 +43,14 @@ Use the correct CLI based on `$PLATFORM`. GraphQL review thread checks only work
 ```
 CronCreate(
   cron="*/10 * * * *",
-  prompt="Run PR monitoring check cycle: fetch my open PRs from all repos in ~/.claude/pilot.yaml, detect new unresolved comments or ready-to-merge status, and report changes. See /pilot-watch-pr for full instructions.",
+  prompt="Run PR monitoring check cycle: fetch my open PRs from all repos in ~/.claude/pilot.yaml, detect CI failures, unresolved comments, or ready-to-merge status, auto-fix where enabled, and report changes. See /pilot-watch-pr for full instructions.",
   recurring=true
 )
 ```
 3. **Report to user**:
 ```
 PR Monitor started. Checking every 10 minutes (auto-expires after 7 days).
+Auto-fix CI: <on/off>  |  Auto-fix comments: <on/off>
 Cron job ID: <id> — cancel anytime with CronDelete.
 ```
 
@@ -48,7 +60,9 @@ Cron job ID: <id> — cancel anytime with CronDelete.
 
 ### Step 1: Fetch my open PRs from ALL repos
 
-For each repo in `$REPOS`:
+For each repo in `$REPOS`, use the platform-appropriate CLI:
+
+#### GitHub (`$PLATFORM == github`):
 ```bash
 gh pr list --repo $REPO_SLUG \
   --author @me \
@@ -57,7 +71,26 @@ gh pr list --repo $REPO_SLUG \
   --limit 20
 ```
 
-For each PR with `reviewDecision` of `CHANGES_REQUESTED` or `REVIEW_REQUIRED`, also fetch unresolved review threads:
+#### GitLab (`$PLATFORM == gitlab`):
+```bash
+glab mr list --repo $REPO_SLUG \
+  --author @me \
+  --state opened \
+  --output json
+```
+
+#### Azure DevOps (`$PLATFORM == azdevops`):
+```bash
+az repos pr list \
+  --repository $REPO_SLUG \
+  --creator @me \
+  --status active \
+  --output json
+```
+
+#### Fetch unresolved review threads (GitHub only):
+
+For each PR with `reviewDecision` of `CHANGES_REQUESTED` or `REVIEW_REQUIRED`:
 ```bash
 gh api graphql -f query='
 query($owner:String!,$repo:String!,$number:Int!) {
@@ -76,18 +109,22 @@ query($owner:String!,$repo:String!,$number:Int!) {
 }' -f owner="$OWNER" -f repo="$REPO_NAME" -F number=$PR_NUMBER
 ```
 
-### Step 2: Detect only TWO conditions
+**Note**: GitLab and Azure DevOps do not support GraphQL review thread queries. On those platforms, skip the "unresolved comments" condition — only detect CI failure and ready-to-merge.
+
+### Step 2: Detect THREE conditions
 
 | Condition | How to detect |
 |-----------|---------------|
+| **CI failure** | `statusCheckRollup` contains any check with `conclusion` of `FAILURE`, `ERROR`, `TIMED_OUT`, or `STARTUP_FAILURE` |
 | **New unresolved comments** | PR has unresolved review threads (`isResolved: false`) |
 | **Ready to merge** | `reviewDecision === "APPROVED"` AND all status checks pass AND not a draft |
 
 ### Step 3: Compare with last check — only act on NEW changes
 
-Track state from the previous cycle. Only report when something **changed**:
+Track state from the previous cycle. Only report/act when something **changed**:
 
-- **New unresolved comments appeared** (count increased) → report
+- **CI just failed** (was passing or pending before) → report + auto-fix if enabled
+- **New unresolved comments appeared** (count increased) → report + auto-fix if enabled
 - **PR just became ready to merge** (wasn't ready before) → report
 
 If nothing changed:
@@ -102,16 +139,33 @@ When a condition is detected and it's **NEW** (changed since last cycle):
 #### Terminal output (one line per PR):
 ```
 PR Monitor — <time>
+❌ #5130 Add retry logic — CI failed (build error) — auto-fixing...
 💬 #5124 Disable fail-fast — 3 unresolved comments
 ✅ #5098 PostToolUse hook — approved & CI green, ready to merge!
 ```
 
 #### Dashboard notification file:
 
-For **unresolved comments**:
+For **CI failure**:
 ```bash
 mkdir -p "$WS/logs/pending-decisions"
 cat > "$WS/logs/pending-decisions/pr-<N>.json" << 'NOTIFICATION'
+{
+  "taskId": "pr-<N>",
+  "issueNumber": null,
+  "prNumber": <N>,
+  "phase": "pr_notification",
+  "question": "PR #<N> CI failed: <failure summary>",
+  "options": ["Auto-fixing...", "Review Logs", "Dismiss"],
+  "context": "<failed job names and brief error>",
+  "timestamp": "<ISO8601>"
+}
+NOTIFICATION
+```
+
+For **unresolved comments**:
+```bash
+cat > "$WS/logs/pending-decisions/pr-<N>.json" << 'NOTIFICATION'
 {
   "taskId": "pr-<N>",
   "issueNumber": null,
@@ -141,28 +195,85 @@ cat > "$WS/logs/pending-decisions/pr-<N>.json" << 'NOTIFICATION'
 NOTIFICATION
 ```
 
-#### Status log:
+### Step 5: Auto-Fix Actions
+
+After detecting and reporting, delegate fixes to `/pilot-dev-issue` which already handles the full analyze → fix → test → commit → push flow.
+
+#### Auto-Fix State Tracking
+
+Maintain a counter per PR per fix type. Persist in `$WS/logs/auto-fix-state.json`:
+```json
+{
+  "pr-123": { "ci_attempts": 0, "comment_attempts": 0 },
+  "pr-456": { "ci_attempts": 2, "comment_attempts": 1 }
+}
+```
+- **Max 3 attempts** per PR per fix type. After 3 failed attempts, stop and notify user.
+- Reset counters when: PR is merged/closed, or CI goes green, or comments are resolved.
+
+#### CI Failure Auto-Fix (when `auto_fix_ci: true`)
+
+Only trigger when CI **newly failed** (not already failing from last cycle with same error).
+
+1. **Spawn `pilot-dev-issue` with `--auto`**:
+```
+Skill(
+  skill="pilot-dev-issue",
+  args="--auto Fix failing CI on PR #<N> (repo: $REPO_SLUG, branch: <branch>). Diagnose the CI failure and push a fix."
+)
+```
+`pilot-dev-issue` will handle everything: fetch CI logs, analyze the failure, fix the code, build-verify, commit, and push.
+
+2. **After completion**:
+   - Increment `ci_attempts` for this PR
+   - Log event: `ci_auto_fix` with detail
+   - Update notification: "Auto-fix pushed, waiting for CI..."
+   - Next cycle will pick up the new CI result
+
+3. **If max attempts (3) reached**:
+   - Write notification asking user to intervene
+   - Log `auto_fix_blocked` event
+   - Stop auto-fixing this PR's CI until user dismisses or CI changes
+
+#### Review Comments Auto-Fix (when `auto_fix_comments: true`)
+
+Only trigger when **new** unresolved comments appear (count increased since last cycle).
+
+1. **Spawn `pilot-dev-issue` with `--auto`**:
+```
+Skill(
+  skill="pilot-dev-issue",
+  args="--auto Address unresolved review comments on PR #<N> (repo: $REPO_SLUG, branch: <branch>). Read the comments, fix the code, and push."
+)
+```
+
+2. **After completion**: increment `comment_attempts`, log event, update notification.
+
+3. **If max attempts (3) reached**: same as CI — notify user and stop.
+
+#### Status log events for auto-fix:
 ```bash
 echo '{"timestamp":"<ISO8601>","task_id":"pr-<N>","type":"<event>","phase":"pr_monitor","pr_number":<N>,"detail":"<message>"}' >> "$WS/logs/pr-<N>.jsonl"
 ```
-Event types: `review_comments`, `ready_to_merge`, `pr_merged`
+New event types: `ci_failure`, `ci_auto_fix`, `ci_auto_fix_failed`, `comment_auto_fix`, `comment_auto_fix_failed`, `auto_fix_blocked`
 
 #### Cleanup:
 - PR merged/closed → **auto-cleanup**:
   1. Delete notification: `rm -f "$WS/logs/pending-decisions/pr-<N>.json"`
-  2. Find and remove the worktree for this PR's branch:
+  2. Remove auto-fix state for this PR from `auto-fix-state.json`
+  3. Find and remove the worktree for this PR's branch:
      ```bash
-     # Find repo and worktree by branch name
      cd "$WS/<repo-name>"
      WORKTREE=$(git worktree list --porcelain | grep -B1 "branch refs/heads/<branch>" | head -1 | sed 's/worktree //')
      if [ -n "$WORKTREE" ]; then
        git worktree remove --force "$WORKTREE"
        git branch -D "<branch>"
      fi
      ```
-  3. Also check `.claude/worktrees/` for native worktrees matching the branch
-  4. Log `pr_merged` event
-- Unresolved comments drop to 0 → delete notification
+  4. Also check `.claude/worktrees/` for native worktrees matching the branch
+  5. Log `pr_merged` event
+- CI goes green → reset `ci_attempts` to 0 for that PR
+- Unresolved comments drop to 0 → delete notification, reset `comment_attempts`
 - Same PR already has notification → overwrite with latest state
 
 If nothing changed:
@@ -179,8 +290,13 @@ PR Monitor — <time> — no open PRs
 
 1. **Start immediately** — no confirmation needed
 2. **10-minute interval** via CronCreate — fires only when session is idle
-3. **Only two conditions** — unresolved comments and ready-to-merge. Ignore everything else.
+3. **Three conditions** — CI failure, unresolved comments, and ready-to-merge
 4. **Only report changes** — don't repeat known status
 5. **Never auto-merge** — only notify
 6. **My PRs only** — ignore other people's PRs
-7. **Keep it simple** — no auto-fix, no worktree management, no CI analysis
+7. **Auto-fix delegates to `/pilot-dev-issue --auto`** — reuses existing worktrees, full fix pipeline
+8. **Auto-fix CI is on by default** — disable via `watch_pr.auto_fix_ci: false` in `pilot.yaml`
+9. **Auto-fix comments is off by default** — enable via `watch_pr.auto_fix_comments: true` in `pilot.yaml`
+10. **Max 3 auto-fix attempts** per PR per fix type — then stop and notify user
+11. **Never force push** — all fixes are new commits
+12. **Log all auto-fix actions** — every fix attempt is logged to `$WS/logs/pr-<N>.jsonl`