agent-harness-kit 0.6.0 → 0.8.0

package/src/templates/scripts/harness-report.mjs

@@ -20,7 +20,9 @@ const ROOT = process.cwd();
 const RESULTS_DIR = resolve(ROOT, ".harness/eval/results");
 const TELEMETRY = resolve(ROOT, ".harness/telemetry.jsonl");
 const NOW = Date.now();
-const SEVEN_DAYS = 7 * 24 * 60 * 60 * 1000;
+const ONE_DAY = 24 * 60 * 60 * 1000;
+const SEVEN_DAYS = 7 * ONE_DAY;
+const FOURTEEN_DAYS = 14 * ONE_DAY;
 
 async function readJsonl(path) {
   if (!existsSync(path)) return [];
@@ -61,6 +63,16 @@ function recent(rows, key = "ts") {
   });
 }
 
+// Rows aged 7–14 days. Used as the comparator for week-over-week deltas
+// so users can spot drift instead of staring at a single-week snapshot.
+function priorWeek(rows, key = "ts") {
+  return rows.filter((r) => {
+    const t = r[key] ? new Date(r[key]).getTime() : r._mtime ?? 0;
+    const age = NOW - t;
+    return age > SEVEN_DAYS && age <= FOURTEEN_DAYS;
+  });
+}
+
 function tokensOf(row) {
   return (row.grades ?? [])
     .filter((g) => g.dim === "efficiency")
@@ -172,16 +184,98 @@ function driftSignals(evalRows, telemetryRows) {
   }
 }
 
+// Aggregate eval rows by task into { passed, total, tokens }.
+function aggregateEvals(rows) {
+  const byTask = new Map();
+  for (const r of rows) {
+    const cur = byTask.get(r.taskId) ?? { passed: 0, total: 0, tokens: 0 };
+    cur.total++;
+    if (r.passed) cur.passed++;
+    cur.tokens += tokensOf(r);
+    byTask.set(r.taskId, cur);
+  }
+  return byTask;
+}
+
+// Render a single delta line. signMode controls icon meaning — for pass-rate,
+// up is good; for tokens, up is bad; for skill invocations, neutral.
+function fmtDelta(now, then, signMode = "neutral", unit = "") {
+  if (then === undefined) return `(new) ${now}${unit}`;
+  const diff = now - then;
+  if (diff === 0) return `${now}${unit} → ${then}${unit}  (=)`;
+  let arrow = diff > 0 ? "↑" : "↓";
+  // Color the arrow by "is this a regression?"
+  let marker = " ";
+  if (signMode === "good-up") marker = diff > 0 ? "+" : "-";
+  else if (signMode === "good-down") marker = diff > 0 ? "-" : "+";
+  return `${now}${unit} ← ${then}${unit}  (${arrow}${marker} ${Math.abs(diff)}${unit})`;
+}
+
+function weekOverWeek(evalRecent, evalPrior, telRecent, telPrior) {
+  console.log(`\n### Week-over-week (last 7d vs prior 7d)`);
+  const aRecent = aggregateEvals(evalRecent);
+  const aPrior = aggregateEvals(evalPrior);
+
+  if (aRecent.size === 0 && aPrior.size === 0) {
+    console.log("  (no eval data in either window — run `npm run harness:eval`)");
+  } else {
+    console.log("  task                    pass-rate (now ← prior)        avg-tokens (now ← prior)");
+    console.log("  ----------------------  ----------------------------   --------------------------");
+    const taskIds = new Set([...aRecent.keys(), ...aPrior.keys()]);
+    for (const t of [...taskIds].sort()) {
+      const now = aRecent.get(t);
+      const prior = aPrior.get(t);
+      const nowRate = now ? Math.round((now.passed / now.total) * 100) : null;
+      const priorRate = prior ? Math.round((prior.passed / prior.total) * 100) : null;
+      const nowTok = now && now.total > 0 ? Math.round(now.tokens / now.total) : 0;
+      const priorTok = prior && prior.total > 0 ? Math.round(prior.tokens / prior.total) : 0;
+      const rateCell = nowRate === null
+        ? "(absent now)"
+        : priorRate === null
+          ? `${nowRate}% (new)`
+          : `${nowRate}% ← ${priorRate}%  (${nowRate - priorRate >= 0 ? "+" : ""}${nowRate - priorRate})`;
+      const tokCell = nowTok === 0 && priorTok === 0
+        ? "—"
+        : `${nowTok} ← ${priorTok}  (${nowTok - priorTok >= 0 ? "+" : ""}${nowTok - priorTok})`;
+      console.log(
+        `  ${t.padEnd(22)}  ${rateCell.padEnd(30)} ${tokCell}`,
+      );
+    }
+  }
+
+  // Skill invocation deltas.
+  const recentBySkill = new Map();
+  for (const r of telRecent) recentBySkill.set(r.skill, (recentBySkill.get(r.skill) ?? 0) + 1);
+  const priorBySkill = new Map();
+  for (const r of telPrior) priorBySkill.set(r.skill, (priorBySkill.get(r.skill) ?? 0) + 1);
+
+  const allSkills = new Set([...recentBySkill.keys(), ...priorBySkill.keys()]);
+  if (allSkills.size > 0) {
+    console.log("\n  skill                          invocations (now ← prior)");
+    console.log("  -----------------------------  -------------------------------");
+    for (const s of [...allSkills].sort()) {
+      const n = recentBySkill.get(s) ?? 0;
+      const p = priorBySkill.get(s) ?? 0;
+      const d = n - p;
+      const cell = p === 0 ? `${n}  (new)` : `${n} ← ${p}  (${d >= 0 ? "+" : ""}${d})`;
+      console.log(`  ${s.padEnd(29)}  ${cell}`);
+    }
+  }
+}
+
 async function main() {
   const evalAll = await loadEvalResults();
   const telemetryAll = await readJsonl(TELEMETRY);
   const evalRows = recent(evalAll);
+  const evalPrior = priorWeek(evalAll);
   const telemetryRows = recent(telemetryAll);
+  const telemetryPrior = priorWeek(telemetryAll);
 
   console.log("=== agent-harness-kit report ===");
   console.log(`Generated: ${new Date().toISOString()}`);
   summarizeEvals(evalRows);
   summarizeTelemetry(telemetryRows);
+  weekOverWeek(evalRows, evalPrior, telemetryRows, telemetryPrior);
   driftSignals(evalRows, telemetryRows);
   console.log("");
 }

package/src/templates/scripts/notify-on-block.sh.hbs

@@ -0,0 +1,73 @@
+#!/usr/bin/env bash
+# Notification hook — OS-native notification when Claude wants attention.
+# macOS osascript / Linux notify-send / Windows skip.
+# Never blocks. Always exits 0. Opt-out: AHK_DISABLE_NOTIFY=1.
+set -eo pipefail
+
+INPUT=$(cat)
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+have_jq() {
+  [ "${AHK_DISABLE_JQ:-}" = "1" ] && return 1
+  command -v jq >/dev/null 2>&1
+}
+have_jp() {
+  have_jq && return 0
+  command -v node >/dev/null 2>&1 && [ -f "$SCRIPT_DIR/_lib/json-pick.mjs" ] && return 0
+  return 1
+}
+jp() {
+  if have_jq; then
+    if [ -n "$2" ]; then jq -r "$1" "$2"; else jq -r "$1"; fi
+  else
+    if [ -n "$2" ]; then
+      node "$SCRIPT_DIR/_lib/json-pick.mjs" "$1" "$2"
+    else
+      node "$SCRIPT_DIR/_lib/json-pick.mjs" "$1"
+    fi
+  fi
+}
+
+if [ "${AHK_DISABLE_NOTIFY:-}" = "1" ]; then
+  exit 0
+fi
+
+TYPE=""
+TITLE=""
+BODY=""
+if have_jp; then
+  TYPE=$(echo "$INPUT"  | jp '.notification.type  // empty')
+  TITLE=$(echo "$INPUT" | jp '.notification.title // empty')
+  BODY=$(echo "$INPUT"  | jp '.notification.body  // empty')
+fi
+
+[ -z "$TITLE" ] && TITLE="Claude Code"
+if [ -n "$TYPE" ]; then
+  BODY="[$TYPE] ${BODY}"
+fi
+[ -z "$BODY" ] && BODY="Claude Code wants your attention."
+
+mkdir -p .harness
+TS=$(date -u +%Y-%m-%dT%H:%M:%SZ)
+ESCAPED_TITLE=${TITLE//\"/\\\"}
+ESCAPED_BODY=${BODY//\"/\\\"}
+printf '{"ts":"%s","hook":"Notification","type":"%s","title":"%s","body":"%s"}\n' \
+  "$TS" "$TYPE" "$ESCAPED_TITLE" "$ESCAPED_BODY" >> .harness/telemetry.jsonl
+
+OS_KIND=$(uname -s 2>/dev/null || echo "Unknown")
+case "$OS_KIND" in
+  Darwin)
+    OSA_TITLE=${TITLE//\"/\\\"}
+    OSA_BODY=${BODY//\"/\\\"}
+    osascript -e "display notification \"$OSA_BODY\" with title \"$OSA_TITLE\"" >/dev/null 2>&1 || true
+    ;;
+  Linux)
+    if command -v notify-send >/dev/null 2>&1; then
+      notify-send -a "Claude Code" "$TITLE" "$BODY" >/dev/null 2>&1 || true
+    fi
+    ;;
+  *)
+    :
+    ;;
+esac
+
+exit 0

package/src/templates/scripts/pre-compact.sh.hbs

@@ -0,0 +1,121 @@
+#!/usr/bin/env bash
+# PreCompact hook — write a small snapshot of state to
+# .harness/compaction-snapshot.json BEFORE the context compactor runs.
+# The companion SessionStart hook (matcher: compact) reads this snapshot
+# back and re-injects the salient fields so the post-compaction model
+# knows which feature it was working on, which branch, and how dirty
+# the tree was.
+#
+# This is the kit's answer to the "I lost everything after compaction"
+# failure mode that recurs in long sessions. Pair with:
+#   - SessionStart matcher compact → re-inject
+#   - PostCompact (not implemented; SessionStart does the work)
+#
+# Snapshot contents:
+#   {
+#     "compacted_at": "2026-05-16T19:00:00Z",
+#     "branch": "main",
+#     "sha": "abc1234",
+#     "uncommitted": 7,
+#     "feature": "auth-endpoint — POST /auth/login",
+#     "trigger": "manual|auto",
+#     "estimated_tokens_removed": 5000
+#   }
+#
+# The hook NEVER blocks (exit 0 always). PreCompact can technically block
+# compaction but doing so defeats the entire point.
+set -eo pipefail
+
+INPUT=$(cat)
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+have_jq() {
+  [ "${AHK_DISABLE_JQ:-}" = "1" ] && return 1
+  command -v jq >/dev/null 2>&1
+}
+have_jp() {
+  have_jq && return 0
+  command -v node >/dev/null 2>&1 && [ -f "$SCRIPT_DIR/_lib/json-pick.mjs" ] && return 0
+  return 1
+}
+jp() {
+  if have_jq; then
+    if [ -n "$2" ]; then jq -r "$1" "$2"; else jq -r "$1"; fi
+  else
+    if [ -n "$2" ]; then
+      node "$SCRIPT_DIR/_lib/json-pick.mjs" "$1" "$2"
+    else
+      node "$SCRIPT_DIR/_lib/json-pick.mjs" "$1"
+    fi
+  fi
+}
+
+TRIGGER=""
+TOKENS=""
+if have_jp; then
+  TRIGGER=$(echo "$INPUT" | jp '.trigger // "auto"' 2>/dev/null || true)
+  TOKENS=$(echo "$INPUT" | jp '.estimated_tokens_removed // 0' 2>/dev/null || true)
+fi
+
+mkdir -p .harness
+
+TS=$(date -u +%Y-%m-%dT%H:%M:%SZ)
+BR="(no-git)"
+SHA="(no-git)"
+COUNT=0
+if command -v git >/dev/null 2>&1 && git rev-parse --git-dir >/dev/null 2>&1; then
+  BR=$(git branch --show-current 2>/dev/null || echo "(detached)")
+  SHA=$(git rev-parse --short HEAD 2>/dev/null || echo "(none)")
+  COUNT=$(git status --short 2>/dev/null | wc -l | tr -d ' ')
+fi
+
+FEAT=""
+if [ -f feature_list.json ]; then
+  if have_jq; then
+    FEAT=$(jq -r 'first(.features[] | select(.passes == false)) | "\(.id) — \(.title)"' \
+      feature_list.json 2>/dev/null || true)
+  elif command -v node >/dev/null 2>&1; then
+    FEAT=$(node -e "
+      const f = JSON.parse(require('fs').readFileSync('feature_list.json','utf8'));
+      const o = (f.features || []).find(x => x.passes === false);
+      if (o) process.stdout.write(o.id + ' — ' + o.title);
+    " 2>/dev/null || true)
+  fi
+fi
+
+# Compose JSON via Node when available — handles escaping right.
+if command -v node >/dev/null 2>&1; then
+  node -e "
+    const fs = require('fs');
+    const snap = {
+      compacted_at: '$TS',
+      branch: '$BR',
+      sha: '$SHA',
+      uncommitted: parseInt('$COUNT', 10) || 0,
+      feature: process.argv[1] || '',
+      trigger: '$TRIGGER' || 'auto',
+      estimated_tokens_removed: parseInt('$TOKENS', 10) || 0
+    };
+    fs.writeFileSync('.harness/compaction-snapshot.json', JSON.stringify(snap, null, 2) + '\n');
+  " "$FEAT"
+elif have_jq; then
+  jq -n --arg ts "$TS" --arg br "$BR" --arg sha "$SHA" \
+        --argjson cnt "$COUNT" --arg feat "$FEAT" \
+        --arg trig "${TRIGGER:-auto}" --argjson tok "${TOKENS:-0}" \
+    '{compacted_at: $ts, branch: $br, sha: $sha, uncommitted: $cnt,
+      feature: $feat, trigger: $trig, estimated_tokens_removed: $tok}' \
+    > .harness/compaction-snapshot.json
+else
+  # No JSON tool available — write a minimal record. SessionStart compact
+  # branch reads fields individually so partial records still work.
+  cat > .harness/compaction-snapshot.json <<EOF
+{
+  "compacted_at": "$TS",
+  "branch": "$BR",
+  "sha": "$SHA",
+  "uncommitted": $COUNT,
+  "feature": "$FEAT",
+  "trigger": "${TRIGGER:-auto}"
+}
+EOF
+fi
+exit 0

package/src/templates/scripts/pre-push.sh

@@ -4,17 +4,42 @@
 # Lives in scripts/ so it ships with the repo; install via install-git-hooks.sh.
 set -eo pipefail
 
+# Resolve script dir so we can find _lib/json-pick.mjs (Node fallback for jq).
+# Without this fallback, `jq` missing on a fresh CI image silently disabled
+# the baseline-monotonic guard — a known audit hole.
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+have_jq() {
+  [ "${AHK_DISABLE_JQ:-}" = "1" ] && return 1
+  command -v jq >/dev/null 2>&1
+}
+have_jp() {
+  have_jq && return 0
+  command -v node >/dev/null 2>&1 && [ -f "$SCRIPT_DIR/_lib/json-pick.mjs" ] && return 0
+  return 1
+}
+jp() {
+  if have_jq; then
+    if [ -n "$2" ]; then jq -r "$1" "$2"; else jq -r "$1"; fi
+  else
+    if [ -n "$2" ]; then
+      node "$SCRIPT_DIR/_lib/json-pick.mjs" "$1" "$2"
+    else
+      node "$SCRIPT_DIR/_lib/json-pick.mjs" "$1"
+    fi
+  fi
+}
+
 # Baseline monotonic guard. .harness/structural-baseline.json is decreasing-
 # only — fixes REMOVE entries; no path should ADD them. Catches the "mask
 # violations by baselining them" anti-pattern before code leaves the machine.
 # Runs first because a grown baseline silently masks structural-test failures.
 BASELINE_FILE=".harness/structural-baseline.json"
 if [ -f "$BASELINE_FILE" ] \
-   && command -v jq >/dev/null 2>&1 \
+   && have_jp \
    && git rev-parse --verify HEAD >/dev/null 2>&1 \
    && git cat-file -e "HEAD:$BASELINE_FILE" 2>/dev/null; then
-  CURRENT_COUNT=$(jq 'length' "$BASELINE_FILE" 2>/dev/null || echo 0)
-  HEAD_COUNT=$(git show "HEAD:$BASELINE_FILE" 2>/dev/null | jq 'length' 2>/dev/null || echo 0)
+  CURRENT_COUNT=$(jp 'length' "$BASELINE_FILE" 2>/dev/null || echo 0)
+  HEAD_COUNT=$(git show "HEAD:$BASELINE_FILE" 2>/dev/null | jp 'length' 2>/dev/null || echo 0)
   if [ "$CURRENT_COUNT" -gt "$HEAD_COUNT" ]; then
     {
       echo

package/src/templates/scripts/precompletion-checklist.sh.hbs

@@ -12,9 +12,43 @@ set -e
 
 INPUT=$(cat)
 
+# Resolve the directory this hook lives in (used to find _lib/json-pick.mjs).
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+
+# have_jq — env-overridable probe. AHK_DISABLE_JQ=1 forces the Node fallback,
+# used by tests to exercise the jq-less code path on machines that have jq
+# installed locally.
+have_jq() {
+  [ "${AHK_DISABLE_JQ:-}" = "1" ] && return 1
+  command -v jq >/dev/null 2>&1
+}
+# jp — JSON picker. Uses `jq` when available, else falls back to a bundled
+# Node script with a jq-subset implementation. Keeps hooks portable on
+# minimal CI / Windows where jq is not installed by default. Without this
+# fallback, the entire pre-completion check used to be a silent no-op.
+jp() {
+  if have_jq; then
+    if [ -n "$2" ]; then jq -r "$1" "$2"; else jq -r "$1"; fi
+  else
+    if [ -n "$2" ]; then
+      node "$SCRIPT_DIR/_lib/json-pick.mjs" "$1" "$2"
+    else
+      node "$SCRIPT_DIR/_lib/json-pick.mjs" "$1"
+    fi
+  fi
+}
+# Probe: do we have either jq or the Node fallback? Node is always
+# present (kit's `engines` field requires >=20), so this is just an explicit
+# probe and a fail-loud branch if even node is missing.
+have_jp() {
+  have_jq && return 0
+  command -v node >/dev/null 2>&1 && [ -f "$SCRIPT_DIR/_lib/json-pick.mjs" ] && return 0
+  return 1
+}
+
 # CRITICAL: avoid infinite loops. If the hook already ran, do not block again.
-if command -v jq >/dev/null 2>&1; then
-  if [ "$(echo "$INPUT" | jq -r '.stop_hook_active // false')" = "true" ]; then
+if have_jp; then
+  if [ "$(echo "$INPUT" | jp '.stop_hook_active // false')" = "true" ]; then
     exit 0
   fi
 fi
@@ -59,12 +93,17 @@ elif [ -f pyproject.toml ] && command -v ruff >/dev/null 2>&1; then
   run_check ruff ruff check . || true
 fi
 
-# CLAUDE.md instruction cap. HumanLayer measurement: agents stop following
-# CLAUDE.md reliably beyond ~150-200 bullets/numbered items. Treat the file
-# as a table of contents; promote details to docs/ or @-imports.
-if [ -f harness.config.json ] && command -v jq >/dev/null 2>&1; then
-  CMD_PATH=$(jq -r '.claudeMd.path // "CLAUDE.md"' harness.config.json)
-  CMD_CAP=$(jq -r '.claudeMd.maxInstructions // 200' harness.config.json)
+# CLAUDE.md size caps. Two complementary signals:
+#   - maxInstructions (default 200): bullet/numbered-item count. Suits
+#     ASCII-heavy English where a bullet ≈ a fixed token weight.
+#   - maxTokens (default 0 = off): approximate token cap. Catches drift
+#     in non-ASCII content (Vietnamese, CJK, etc.) where 200 bullets
+#     may carry 2–3× more tokens than the HumanLayer baseline measured.
+# Both checks fire independently — exceed either → block.
+if [ -f harness.config.json ] && have_jp; then
+  CMD_PATH=$(jp '.claudeMd.path // "CLAUDE.md"' harness.config.json)
+  CMD_CAP=$(jp '.claudeMd.maxInstructions // 200' harness.config.json)
+  CMD_TOK_CAP=$(jp '.claudeMd.maxTokens // 0' harness.config.json)
   if [ -f "$CMD_PATH" ] && [ "$CMD_CAP" -gt 0 ] 2>/dev/null; then
     CMD_COUNT=$(grep -cE '^[[:space:]]*([-*]|[0-9]+\.)[[:space:]]' "$CMD_PATH" 2>/dev/null || echo 0)
     if [ "$CMD_COUNT" -gt "$CMD_CAP" ]; then
@@ -85,6 +124,24 @@ if [ -f harness.config.json ] && command -v jq >/dev/null 2>&1; then
       echo "claude-md-cap" >> "$TMPDIR_HOOK/failed.list"
     fi
   fi
+  if [ -f "$CMD_PATH" ] && [ "$CMD_TOK_CAP" -gt 0 ] 2>/dev/null \
+     && command -v node >/dev/null 2>&1 \
+     && [ -f "$SCRIPT_DIR/_lib/approx-tokens.mjs" ]; then
+    CMD_TOK=$(node "$SCRIPT_DIR/_lib/approx-tokens.mjs" "$CMD_PATH" 2>/dev/null || echo 0)
+    if [ "$CMD_TOK" -gt "$CMD_TOK_CAP" ]; then
+      {
+        echo "$CMD_PATH approximate token count: $CMD_TOK (cap: $CMD_TOK_CAP)"
+        echo
+        echo "Heuristic token cap — set because instruction count alone misses"
+        echo "drift in non-ASCII content (Vietnamese, CJK) where a bullet can"
+        echo "carry 2-3x more tokens than the HumanLayer baseline measured."
+        echo
+        echo "Adjust the cap (with justification) in harness.config.json:"
+        echo "  .claudeMd.maxTokens"
+      } > "$TMPDIR_HOOK/claude-md-tokens.out"
+      echo "claude-md-tokens" >> "$TMPDIR_HOOK/failed.list"
+    fi
+  fi
 fi
 
 # Multi-layer review trigger. When uncommitted/staged/untracked changes touch
@@ -94,7 +151,7 @@ fi
 # with a mechanical count off `harness.config.json` `domains[].layers` /
 # `.root`. Fires once per stop; the loop guard (`stop_hook_active`) lets the
 # next stop succeed after the agent has read the recommendation.
-if [ -f harness.config.json ] && command -v jq >/dev/null 2>&1 && command -v git >/dev/null 2>&1; then
+if [ -f harness.config.json ] && have_jp && command -v git >/dev/null 2>&1; then
   CHANGED=$(
     {
       git diff --name-only 2>/dev/null || true
@@ -103,19 +160,19 @@ if [ -f harness.config.json ] && command -v jq >/dev/null 2>&1 && command -v git
     } | sort -u
   )
   if [ -n "$CHANGED" ]; then
-    NUM_DOMAINS=$(jq '.domains | length' harness.config.json 2>/dev/null || echo 0)
+    NUM_DOMAINS=$(jp '.domains | length' harness.config.json 2>/dev/null || echo 0)
     MULTI_OUT="$TMPDIR_HOOK/multi-layer-review.out"
     : > "$MULTI_OUT"
     MULTI_HIT=0
     i=0
     while [ "$i" -lt "$NUM_DOMAINS" ]; do
-      ROOT=$(jq -r ".domains[$i].root" harness.config.json)
-      DOMAIN=$(jq -r ".domains[$i].name" harness.config.json)
+      ROOT=$(jp ".domains[$i].root" harness.config.json)
+      DOMAIN=$(jp ".domains[$i].name" harness.config.json)
       # Optional layerDirPattern — supports conventions where the layer
       # directory is not literally `{layer}`. Example: a Rust workspace
       # with crates named `unibot-types`, `unibot-crypto`, ... uses
       # `"layerDirPattern": "unibot-{layer}"`. Defaults to `{layer}`.
-      LAYER_PATTERN=$(jq -r ".domains[$i].layerDirPattern // \"{layer}\"" harness.config.json)
+      LAYER_PATTERN=$(jp ".domains[$i].layerDirPattern // \"{layer}\"" harness.config.json)
       TOUCHED_COUNT=0
       TOUCHED_NAMES=""
       while IFS= read -r layer; do
@@ -125,7 +182,7 @@ if [ -f harness.config.json ] && command -v jq >/dev/null 2>&1 && command -v git
           TOUCHED_COUNT=$((TOUCHED_COUNT + 1))
           TOUCHED_NAMES="$TOUCHED_NAMES $layer"
         fi
-      done < <(jq -r ".domains[$i].layers[]" harness.config.json)
+      done < <(jp ".domains[$i].layers[]" harness.config.json)
       if [ "$TOUCHED_COUNT" -ge 2 ]; then
         echo "Domain '$DOMAIN' has changes spanning $TOUCHED_COUNT layers:$TOUCHED_NAMES" >> "$MULTI_OUT"
         MULTI_HIT=1
@@ -187,8 +244,8 @@ HEADLESS_SOURCE=""
 if [ "${AHK_HEADLESS_RECOVER:-}" = "1" ]; then
   HEADLESS_RECOVER=1
   HEADLESS_SOURCE="AHK_HEADLESS_RECOVER"
-elif [ -f harness.config.json ] && command -v jq >/dev/null 2>&1; then
-  CFG_VAL=$(jq -r '.recovery.headless // false' harness.config.json 2>/dev/null)
+elif [ -f harness.config.json ] && have_jp; then
+  CFG_VAL=$(jp '.recovery.headless // false' harness.config.json 2>/dev/null)
   if [ "$CFG_VAL" = "true" ]; then
     HEADLESS_RECOVER=1
     HEADLESS_SOURCE="harness.config.json:.recovery.headless"
@@ -196,12 +253,64 @@ elif [ -f harness.config.json ] && command -v jq >/dev/null 2>&1; then
 fi
 if [ "$HEADLESS_RECOVER" = "1" ] && command -v claude >/dev/null 2>&1; then
   FAILED_LIST=$(tr '\n' ' ' < "$TMPDIR_HOOK/failed.list")
-  echo "[ahk] headless recovery enabled ($HEADLESS_SOURCE) — spawning recovery turn for: $FAILED_LIST" >&2
-  claude -p \
-    "The pre-completion checklist failed: $FAILED_LIST. Read the failure output in $TMPDIR_HOOK and apply the smallest fix. Do not disable any check." \
-    --max-turns 5 \
-    >"$TMPDIR_HOOK/recover.out" 2>&1 &
-  # Don't wait — let the next session pick up the partially-applied fix.
+
+  # Concurrency guard. Two Stop events in different sessions (e.g. user
+  # working in two terminals, or an unattended CI rerun firing while a
+  # previous recovery is still active) used to race and edit the same
+  # files. The lock is a directory created atomically with `mkdir`; the
+  # PID file inside lets us detect stale locks left by a crashed parent.
+  mkdir -p .harness
+  LOCK_DIR=".harness/recovery.lock"
+  LOCK_STALE_MAX_SECS=${AHK_RECOVERY_LOCK_STALE_SECS:-1800}
+
+  if mkdir "$LOCK_DIR" 2>/dev/null; then
+    # We won the race — spawn the recovery turn. Snapshot the failure
+    # context into the lock dir BEFORE the parent's EXIT trap deletes
+    # TMPDIR_HOOK; otherwise the subshell's redirect to recover.out
+    # races the parent's cleanup and the subshell dies before claude
+    # can run. Everything the recovery needs (failed.list, per-check
+    # output, recover.out) now lives inside LOCK_DIR — self-contained.
+    cp -r "$TMPDIR_HOOK/." "$LOCK_DIR/snapshot/" 2>/dev/null || true
+    (
+      # Trap removes the lock on subshell EXIT (success, failure, or signal).
+      trap 'rm -rf "$LOCK_DIR"' EXIT
+      claude -p \
+        "The pre-completion checklist failed: $FAILED_LIST. Read the failure output in $LOCK_DIR/snapshot and apply the smallest fix. Do not disable any check." \
+        --max-turns 5 \
+        >"$LOCK_DIR/recover.out" 2>&1
+    ) &
+    SUB_PID=$!
+    # Parent writes metadata SYNCHRONOUSLY before printing the "spawned"
+    # message so a second Stop firing immediately after never sees an
+    # empty pid file. Subsecond races between mkdir and these writes are
+    # closed by the bounded read-loop in the lock-held branch below.
+    echo "$SUB_PID" > "$LOCK_DIR/pid"
+    date +%s > "$LOCK_DIR/started_at"
+    echo "$HEADLESS_SOURCE" > "$LOCK_DIR/source"
+    echo "[ahk] headless recovery spawned (source=$HEADLESS_SOURCE, wrapper-pid=$SUB_PID, lock=$LOCK_DIR)" >&2
+  else
+    # Lock already held. Read who holds it and decide: live → skip,
+    # stale → reclaim. We never block the user's Stop on the lock —
+    # worst case we skip a recovery turn that the next Stop can retry.
+    # Bounded wait for the pid file to materialize — closes the race
+    # window between the parent's `mkdir` and its `echo $SUB_PID > pid`.
+    for _ in 1 2 3 4 5 6 7 8 9 10; do
+      [ -s "$LOCK_DIR/pid" ] && break
+      sleep 0.05
+    done
+    EXISTING_PID=$(cat "$LOCK_DIR/pid" 2>/dev/null || true)
+    STARTED_AT=$(cat "$LOCK_DIR/started_at" 2>/dev/null || echo 0)
+    NOW=$(date +%s)
+    AGE=$((NOW - STARTED_AT))
+    if [ -n "$EXISTING_PID" ] && kill -0 "$EXISTING_PID" 2>/dev/null; then
+      echo "[ahk] headless recovery skipped — another session already running (pid=$EXISTING_PID, age=${AGE}s, lock=$LOCK_DIR)" >&2
+    elif [ "$AGE" -gt "$LOCK_STALE_MAX_SECS" ]; then
+      echo "[ahk] headless recovery: removing stale lock (pid=$EXISTING_PID, age=${AGE}s > ${LOCK_STALE_MAX_SECS}s); next stop will retry. lock=$LOCK_DIR" >&2
+      rm -rf "$LOCK_DIR"
+    else
+      echo "[ahk] headless recovery skipped — lock present with dead pid=$EXISTING_PID (age=${AGE}s, will reclaim after ${LOCK_STALE_MAX_SECS}s). lock=$LOCK_DIR" >&2
+    fi
+  fi
 fi
 
 exit 2