feat(implement): pre-flight check skips closed/already-merged issues (#936)

.gitignore

@@ -7,6 +7,10 @@ tests/genai/.genai_cache/
 # Activity logs (local-only, generated by session_activity_logger)
 .claude/logs/
 
+# Headless run logs (drain-backlog.sh / triage-and-implement.sh)
+logs/drain/
+logs/scheduled/
+
 # Python
 venv/
 *.pyc

CHANGELOG.md

@@ -1,6 +1,8 @@
 ## [Unreleased]
 
 ### Added
+- **`/implement` pre-flight check — fast-fail on already-merged issues** (Issue #936): `/implement --issues #N` now exits immediately with `BLOCKED` + `exit 1` when the target issue is already CLOSED or when any of the last 20 commits references `#N`. Previously the full pipeline ran to completion before discovering the issue was already done, wasting up to 10 minutes per occurrence. Two independent checks fire: (1) issue state check via the already-fetched `gh issue view` response (free; skipped silently when `gh` is unavailable per AC #5), (2) recent-commit grep via `git log --oneline -n 20 --grep "#N\b"` (always runs, even without `gh`). Pass `--force` to bypass both checks for intentional re-implementation of a closed issue. Free-text invocations without an issue reference are unaffected. `docs/PIPELINE-MODES.md` Mode Selection table updated to document the pre-flight behavior and `--force` override. 4 new structural regression tests in `tests/unit/commands/test_implement_preflight_issue_check.py`.
+
 - **`install.sh --uninstall` — shell-only uninstall path** (Issue #951): Previously, the only documented removal path was `/sync --uninstall` (Python orchestrator), which requires a working Claude CLI. When Claude Code itself is broken or hooks are bricked, `/sync` cannot run. `install.sh` now accepts `--uninstall` (remove hooks and global registration), `--dry-run` (preview plan without mutating anything), and `--repos <list>` (also strip autonomous-dev hooks from named per-repo `.claude/settings.json` files). Two companion scripts — `plugins/autonomous-dev/scripts/uninstall_strip_repo_hooks.py` and `plugins/autonomous-dev/scripts/uninstall_unregister_plugin.py` — handle the per-repo hook stripping and plugin-registry cleanup respectively; both are self-contained (no imports from the plugin's lib tree) so they work even in a partially-broken install. A timestamped backup is created at `~/.claude/backups/uninstall-YYYYMMDD-HHMMSS/` before any mutation. `plugins/autonomous-dev/docs/TROUBLESHOOTING.md` gains an "Uninstalling autonomous-dev" section with a comparison table. `README.md` Uninstall section updated to document both paths. 15 new tests across `tests/unit/scripts/test_uninstall_strip_repo_hooks.py` (6), `tests/unit/scripts/test_uninstall_unregister_plugin.py` (5), and `tests/integration/test_install_sh_uninstall.py` (4).
 
 ### Security

docs/PIPELINE-MODES.md

@@ -21,6 +21,8 @@ covers:
 | **Batch (issues)** | `--issues <nums>` | Process GitHub issues with auto-worktree per issue |
 | **Resume** | `--resume <run_id>` | Recover from auto-compact / crash mid-pipeline |
 
+**Pre-flight check (Issue #936)**: When ARGUMENTS contains a single issue reference (`#N`), STEP 0 exits with `BLOCKED` and `exit 1` if the issue is already CLOSED or if a recent commit (last 20) references `#N`. This prevents burning pipeline time on already-merged work. Pass `--force` to skip the check (rare: intentional re-implementation of a closed issue). Free-text invocations (no `#N`) and `gh`-unavailable cases are unaffected.
+
 **Auto-detection**: If no mode flag is given, `/implement` scans the feature description for signals:
 - Fix keywords (`failing test`, `broken test`, `flaky test`) → suggests `--fix`
 - Light keywords or file paths (`*.md`, `*.json`, `*.yaml`, docs-only) not matching security patterns → suggests `--light`

plugins/autonomous-dev/commands/implement.md

@@ -212,16 +212,39 @@ If ARGUMENTS contains an issue reference (`#NNN` or issue number), fetch the iss
 ```bash
 ISSUE_NUMBER=$(echo "ARGUMENTS" | grep -oE '#?([0-9]+)' | head -1 | tr -d '#')
 if [ -n "$ISSUE_NUMBER" ]; then
-  ISSUE_DATA=$(gh issue view "$ISSUE_NUMBER" --json title,body 2>/dev/null)
+  ISSUE_DATA=$(gh issue view "$ISSUE_NUMBER" --json title,body,state 2>/dev/null)
   if [ $? -eq 0 ]; then
     ISSUE_TITLE=$(echo "$ISSUE_DATA" | python3 -c "import sys,json; print(json.load(sys.stdin).get('title',''))")
     ISSUE_BODY=$(echo "$ISSUE_DATA" | python3 -c "import sys,json; print(json.load(sys.stdin).get('body',''))")
+    ISSUE_STATE=$(echo "$ISSUE_DATA" | python3 -c "import sys,json; print(json.load(sys.stdin).get('state',''))")
   fi
 fi
 ```
 
 Store `ISSUE_BODY` and `ISSUE_TITLE` as pipeline context. If `gh issue view` fails, proceed without issue body (ISSUE_BODY remains empty). Do NOT block the pipeline on fetch failure.
 
+**Pre-flight: skip already-merged or already-addressed issues** (Issue #936):
+
+```bash
+# Pre-flight: skip if issue is already closed or recently referenced in commits.
+# Issue #936: prevent burning pipeline time on already-merged work.
+if [ -n "$ISSUE_NUMBER" ] && ! echo "ARGUMENTS" | grep -q -- "--force"; then
+  # Check 1: issue state (free — uses already-fetched ISSUE_STATE; AC #5: gh failures leave ISSUE_STATE empty, skip silently)
+  if [ -n "$ISSUE_STATE" ] && [ "$ISSUE_STATE" != "OPEN" ]; then
+    echo "BLOCKED: Issue #${ISSUE_NUMBER} is ${ISSUE_STATE}, not OPEN. Use --force to override (rare: re-implement intentionally)."
+    exit 1
+  fi
+  # Check 2: recent commit references (always runs, even if gh unavailable; AC #5)
+  RECENT_FIX=$(git log --oneline -n 20 --grep "#${ISSUE_NUMBER}\b" 2>/dev/null | head -3)
+  if [ -n "$RECENT_FIX" ]; then
+    echo "BLOCKED: Issue #${ISSUE_NUMBER} appears already addressed in recent commits:"
+    echo "$RECENT_FIX"
+    echo "Use --force to override."
+    exit 1
+  fi
+fi
+```
+
 Activate pipeline state:
 ```bash
 # Garbage-collect stale state files from prior crashed runs (Issue #1048)

scripts/drain-all.sh

@@ -0,0 +1,160 @@
+#!/usr/bin/env bash
+# Serial single-issue drain of the open-issue backlog.
+#
+# Lists open issues, excludes epics + already-PR'd, runs /implement N for each
+# in headless mode with full harness. Resumable: re-running picks up where it
+# left off by skipping issues already attempted.
+#
+# Usage:
+#   bash scripts/drain-all.sh                 # drain everything
+#   bash scripts/drain-all.sh --phase1        # only the top auto-improvement cluster (~36 issues)
+#   bash scripts/drain-all.sh --dry           # print queue, don't run
+#   bash scripts/drain-all.sh --label LABEL   # filter by GH label
+#   bash scripts/drain-all.sh --limit 10      # stop after N issues
+#
+# Background-friendly: run with `nohup bash scripts/drain-all.sh &` and tail
+# logs/drain-all/progress.log to watch.
+
+set -euo pipefail
+cd "$(dirname "$0")/.."
+
+LOG_DIR="logs/drain-all"
+STATE_DIR="$LOG_DIR/state"
+PROGRESS_LOG="$LOG_DIR/progress.log"
+mkdir -p "$STATE_DIR"
+
+REPO="akaszubski/autonomous-dev"
+DRY=0
+PHASE1=0
+LABEL_FILTER=""
+LIMIT=0
+
+while [[ $# -gt 0 ]]; do
+  case "$1" in
+    --dry) DRY=1; shift ;;
+    --phase1) PHASE1=1; shift ;;
+    --label) LABEL_FILTER="$2"; shift 2 ;;
+    --limit) LIMIT="$2"; shift 2 ;;
+    --repo) REPO="$2"; shift 2 ;;
+    -h|--help) sed -n '1,20p' "$0"; exit 0 ;;
+    *) echo "unknown flag: $1"; exit 2 ;;
+  esac
+done
+
+log() { echo "[$(date '+%Y-%m-%d %H:%M:%S')] $*" | tee -a "$PROGRESS_LOG"; }
+
+# Build the queue.
+QUEUE_FILE="$LOG_DIR/queue.txt"
+
+if (( PHASE1 )); then
+  log "Phase 1: using /triage top cluster"
+  python3 -c "
+import sys, os, json
+for _p in ('.claude/lib', 'plugins/autonomous-dev/lib', os.path.expanduser('~/.claude/lib')):
+    if os.path.isdir(_p):
+        sys.path.insert(0, _p); break
+from issue_triage_analyzer import main
+import io, contextlib
+buf = io.StringIO()
+with contextlib.redirect_stdout(buf):
+    main(['--auto-improvement', '--repo', '$REPO', '--json'])
+data = json.loads(buf.getvalue())
+clusters = sorted(data, key=lambda c: -c['rank_score'])
+print('\n'.join(str(n) for n in clusters[0]['issue_numbers']))
+" > "$QUEUE_FILE"
+else
+  log "Full drain: listing all open issues (excluding epics + PR-linked)"
+  GH_ARGS=(--repo "$REPO" --state open --limit 300 --json number,labels,closedByPullRequestsReferences)
+  if [[ -n "$LABEL_FILTER" ]]; then GH_ARGS+=(--label "$LABEL_FILTER"); fi
+  gh issue list "${GH_ARGS[@]}" \
+    --jq '.[] | select((.labels | map(.name) | contains(["epic"])) | not) | select((.closedByPullRequestsReferences | length) == 0) | .number' \
+    | sort -n > "$QUEUE_FILE"
+fi
+
+TOTAL=$(wc -l < "$QUEUE_FILE" | tr -d ' ')
+log "Queue: $TOTAL issues"
+
+if (( LIMIT > 0 )) && (( LIMIT < TOTAL )); then
+  head -n "$LIMIT" "$QUEUE_FILE" > "$QUEUE_FILE.limited" && mv "$QUEUE_FILE.limited" "$QUEUE_FILE"
+  TOTAL="$LIMIT"
+  log "Limited to first $LIMIT"
+fi
+
+if (( DRY )); then
+  log "DRY run — queue contents:"
+  cat "$QUEUE_FILE"
+  exit 0
+fi
+
+# Pre-flight checks.
+if ! git diff --quiet || ! git diff --cached --quiet; then
+  log "ABORT: working tree dirty"
+  git status --short
+  exit 1
+fi
+
+if ! command -v claude >/dev/null; then
+  log "ABORT: claude CLI not on PATH"
+  exit 1
+fi
+
+# Drain loop.
+i=0
+DONE=0
+FAILED=0
+SKIPPED=0
+
+while read -r n; do
+  i=$((i+1))
+  marker="$STATE_DIR/done-$n"
+  fail_marker="$STATE_DIR/failed-$n"
+
+  if [[ -f "$marker" ]]; then
+    SKIPPED=$((SKIPPED+1))
+    log "[$i/$TOTAL] #$n SKIP (already done)"
+    continue
+  fi
+  if [[ -f "$fail_marker" ]]; then
+    SKIPPED=$((SKIPPED+1))
+    log "[$i/$TOTAL] #$n SKIP (previously failed; rm $fail_marker to retry)"
+    continue
+  fi
+
+  log "[$i/$TOTAL] #$n START"
+  ts="$(date +%Y%m%d-%H%M%S)"
+  events="$LOG_DIR/${ts}-${n}.events.json"
+
+  # Refresh state: working tree must be clean before each iteration.
+  if ! git diff --quiet || ! git diff --cached --quiet; then
+    log "  abort: tree became dirty between iterations — stopping"
+    exit 1
+  fi
+
+  rm -f /tmp/implement_pipeline_state.json /tmp/implement_pipeline_state.lock 2>/dev/null || true
+
+  if claude \
+      --print \
+      --permission-mode acceptEdits \
+      --output-format stream-json \
+      --include-hook-events \
+      --verbose \
+      --name "drain-all-${n}" \
+      --setting-sources user,project,local \
+      "/implement $n" \
+      > "$events" 2>>"$PROGRESS_LOG"; then
+    touch "$marker"
+    DONE=$((DONE+1))
+    log "[$i/$TOTAL] #$n OK"
+  else
+    touch "$fail_marker"
+    FAILED=$((FAILED+1))
+    log "[$i/$TOTAL] #$n FAIL (events: $events)"
+  fi
+
+  log "  progress: $DONE done, $FAILED failed, $SKIPPED skipped, $((TOTAL - i)) remaining"
+done < "$QUEUE_FILE"
+
+log "==="
+log "Drain complete. done=$DONE failed=$FAILED skipped=$SKIPPED total=$TOTAL"
+log "Failed markers in $STATE_DIR/failed-*"
+log "Retry failures by removing markers: rm $STATE_DIR/failed-*"

scripts/drain-backlog.sh

@@ -0,0 +1,131 @@
+#!/usr/bin/env bash
+# Headless drainer for /implement --batch — uses the full harness.
+#
+# Each batch runs in its own `claude -p` process: fresh context, full hook stack,
+# all agents, all enforcement. No --bare, no dangerously-skip-permissions.
+# Hooks decide what's allowed; we just don't pause for permission prompts on edits.
+#
+# Usage:
+#   bash scripts/drain-backlog.sh              # run all batches in order
+#   bash scripts/drain-backlog.sh --dry        # print plan only
+#   bash scripts/drain-backlog.sh --from 5     # resume from batch index 5
+#   bash scripts/drain-backlog.sh --only m0    # run a single named batch
+#   bash scripts/drain-backlog.sh --budget 25  # USD cap per batch (default 50)
+#
+# Output:
+#   logs/drain/<ts>-<idx>-<name>.log         human-readable session log
+#   logs/drain/<ts>-<idx>-<name>.events.json stream-json event stream
+#   logs/drain/state.json                    resume marker
+
+set -euo pipefail
+
+cd "$(dirname "$0")/.."
+
+LOG_DIR="logs/drain"
+STATE_FILE="$LOG_DIR/state.json"
+mkdir -p "$LOG_DIR"
+
+# Ordered batches: name:csv-issues. Security first, big epics excluded.
+BATCHES=(
+  "security:958,1040,1111,918,984,978,1018"
+  "ordering:935,957,936,940,973,1090,1106,1107,1109,1110"
+  "hook-fps:1001,1002,1031,1032,1038,917,1055"
+  "coordinator:980,987,988,990,1010,1095"
+  "plan-critic:903,1006,1073,920,873"
+  "install:945,952,894,1036,895,977,1108"
+  "m0:1045,1046,1047,1048"
+  "m2:1058,1059,1060,1014,1015,961,962,963"
+  "m3:974,975,976,1025,1027"
+  "m4m5:964,1022,1026,1044"
+  "periodic:1099,1100"
+  "intent:1070,1072,1074,1076,1078,1079"
+  "refactor:1000,1013,1016,1034"
+  "doc-drift:900,1061,915"
+  "misc:924,931,932,981,983,1009,955,956,959,965,912,913,914,916"
+  "new-feat:909,910,911,979"
+)
+
+DRY=0
+FROM=0
+ONLY=""
+while [[ $# -gt 0 ]]; do
+  case "$1" in
+    --dry) DRY=1; shift ;;
+    --from) FROM="$2"; shift 2 ;;
+    --only) ONLY="$2"; shift 2 ;;
+    --budget) shift 2 ;;  # accepted but ignored — subscription auth, no $ cap
+    -h|--help) sed -n '1,30p' "$0"; exit 0 ;;
+    *) echo "unknown flag: $1"; exit 2 ;;
+  esac
+done
+
+# Pre-flight: working tree must be clean (harness commits per batch).
+if [[ -z "${SKIP_GIT_CHECK:-}" ]] && ! git diff --quiet || ! git diff --cached --quiet; then
+  echo "ERROR: working tree dirty. Commit/stash first, or set SKIP_GIT_CHECK=1."
+  git status --short
+  exit 1
+fi
+
+# Pre-flight: plugin must be loaded.
+if ! claude plugin list 2>/dev/null | grep -q autonomous-dev; then
+  echo "WARN: autonomous-dev plugin not in 'claude plugin list'. Hooks may not fire."
+fi
+
+run_batch() {
+  local idx="$1" name="$2" issues="$3"
+  local ts; ts="$(date +%Y%m%d-%H%M%S)"
+  local base="$LOG_DIR/${ts}-${idx}-${name}"
+  local log="${base}.log"
+  local events="${base}.events.json"
+
+  echo "[$idx/${#BATCHES[@]}] batch=$name issues=$issues -> $log"
+
+  # Clear any stale pipeline state from a prior crashed run.
+  rm -f /tmp/implement_pipeline_state.json /tmp/implement_pipeline_state.lock 2>/dev/null || true
+
+  # Run the harness via headless claude. Hooks + agents + gates all active.
+  # --permission-mode acceptEdits: auto-approve file edits BUT hook decisions still block.
+  # --output-format stream-json + --include-hook-events: full diagnostic trail.
+  claude \
+    --print \
+    --permission-mode acceptEdits \
+    --output-format stream-json \
+    --include-hook-events \
+    --verbose \
+    --name "drain-${name}-${ts}" \
+    --setting-sources user,project,local \
+    "/implement --batch $issues" \
+    > "$events" 2> "$log" || {
+      echo "  batch $name FAILED — see $log"
+      jq -n --arg name "$name" --arg ts "$ts" --arg log "$log" --arg status failed \
+        '{ts:$ts,name:$name,status:$status,log:$log}' \
+        >> "$LOG_DIR/state.json"
+      return 0  # don't abort the queue
+    }
+
+  # Extract last result line for summary.
+  jq -n --arg name "$name" --arg ts "$ts" --arg log "$log" --arg status ok \
+    '{ts:$ts,name:$name,status:$status,log:$log}' \
+    >> "$LOG_DIR/state.json"
+  echo "  ok"
+}
+
+i=0
+for entry in "${BATCHES[@]}"; do
+  i=$((i+1))
+  name="${entry%%:*}"
+  issues="${entry#*:}"
+
+  if [[ -n "$ONLY" && "$ONLY" != "$name" ]]; then continue; fi
+  if (( i <= FROM )); then continue; fi
+
+  if (( DRY )); then
+    echo "[$i/${#BATCHES[@]}] DRY: claude -p '/implement --batch $issues'"
+    continue
+  fi
+
+  run_batch "$i" "$name" "$issues"
+done
+
+echo
+echo "Done. Logs: $LOG_DIR/    State: $STATE_FILE"