@qijenchen/design-system 0.1.0-beta.8 → 0.1.0-beta.80

package/ds-canonical/hooks/retired/tests/test_pre_write_subsumption_check.sh

@@ -0,0 +1,63 @@
+#!/bin/bash
+# Tests for pre_write_subsumption_check.sh
+#
+# Scope A:new file detection / Scope B:M-row addition
+
+set -u
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+HOOK="$SCRIPT_DIR/../pre_write_subsumption_check.sh"
+[ -x "$HOOK" ] || { echo "FATAL: hook not executable"; exit 1; }
+
+PASS=0; FAIL=0; FAILED=""
+
+run_hook() {
+  STDOUT=$(mktemp); STDERR=$(mktemp)
+  set +e; printf '%s' "$1" | bash "$HOOK" >"$STDOUT" 2>"$STDERR"; EXIT=$?; set -e
+  STDOUT_TEXT=$(cat "$STDOUT"); rm -f "$STDOUT" "$STDERR"
+}
+
+expect_fire_with() {
+  local name="$1"; local needle="$2"
+  if [ "$EXIT" = "0" ] && echo "$STDOUT_TEXT" | grep -q "$needle"; then
+    echo "  PASS  $name"; PASS=$((PASS+1))
+  else
+    echo "  FAIL  $name (no '$needle' in output;exit=$EXIT)"
+    FAIL=$((FAIL+1)); FAILED="${FAILED}\n  - $name"
+  fi
+}
+
+expect_silent() {
+  local name="$1"
+  if [ "$EXIT" = "0" ] && [ -z "$STDOUT_TEXT" ]; then
+    echo "  PASS  $name"; PASS=$((PASS+1))
+  else
+    echo "  FAIL  $name (expected silent, got exit=$EXIT, stdout='$STDOUT_TEXT')"
+    FAIL=$((FAIL+1)); FAILED="${FAILED}\n  - $name"
+  fi
+}
+
+# Scope B: M-row addition → fire 3-Q context
+echo "Test 1: M-row addition fires 3-Q"
+P=$(jq -n --arg fp "/abs/CLAUDE.md" --arg os "old" --arg ns "| **M19** | new pattern |" '{tool_name:"Edit",tool_input:{file_path:$fp,old_string:$os,new_string:$ns}}')
+run_hook "$P"; expect_fire_with "Test 1 M-row fires" "Meta-Pattern"
+
+# Scope B: non-M-row CLAUDE.md edit → silent
+echo "Test 2: non-M-row CLAUDE.md edit silent"
+P=$(jq -n --arg fp "/abs/CLAUDE.md" --arg os "old" --arg ns "regular text without M pattern" '{tool_name:"Edit",tool_input:{file_path:$fp,old_string:$os,new_string:$ns}}')
+run_hook "$P"; expect_silent "Test 2 regular CLAUDE.md edit silent"
+
+# Scope B: Write CLAUDE.md with M-row(file might not exist)→ fire
+echo "Test 3: Write CLAUDE.md with M-row in content"
+P=$(jq -n --arg fp "/abs/CLAUDE.md" --arg c "| **M20** | test |" '{tool_name:"Write",tool_input:{file_path:$fp,content:$c}}')
+run_hook "$P"; expect_fire_with "Test 3 Write M-row fires" "Meta-Pattern"
+
+# Scope B: M-row pattern in non-CLAUDE.md → silent
+echo "Test 4: M-row in random spec → silent (Scope B 限 CLAUDE.md)"
+P=$(jq -n --arg fp "/abs/packages/design-system/src/components/Foo/foo.spec.md" --arg os "o" --arg ns "| **M99** | not real |" '{tool_name:"Edit",tool_input:{file_path:$fp,old_string:$os,new_string:$ns}}')
+run_hook "$P"; expect_silent "Test 4 spec.md not Scope B"
+
+echo ""
+echo "════ Results: $PASS PASS, $FAIL FAIL ════"
+[ "$FAIL" -gt 0 ] && { printf "Failed:%b\n" "$FAILED"; exit 1; }
+exit 0

package/ds-canonical/hooks/session_start_governance_check.sh

@@ -0,0 +1,310 @@
+#!/bin/bash
+set -uo pipefail
+# SessionStart hook: check if governance hygiene actions are overdue; if so,
+# inject reminder into session context so Claude proactively addresses them.
+#
+# Why: M10 — silent tech debt violates. If /knowledge-prune last ran 3+ months
+# ago, CLAUDE.md is over 800 lines, or user-corrections.jsonl has pending
+# entries not yet codified → user / Claude should know at session start, not
+# discover later when things break.
+#
+# Two-tier thresholds(2026-04-25 G1):
+#   • Soft — inject reminder,non-blocking
+#   • Hard(2x soft)— inject BLOCKER context requiring Claude's first action
+#     to be /knowledge-prune(SessionStart hooks cannot truly block session,
+#     but hard-tier context is prefixed with 🚨 BLOCKER / REQUIRED_FIRST_ACTION
+#     so Claude reads it as must-address-first instruction)
+#
+# Checks + thresholds:
+#   1. CLAUDE.md line count     — soft 400 / hard 800(CLAUDE.md L34 SSOT: transition 400 / hard cap 800)
+#   2. Days since last prune    — soft 90   / hard 180
+#   3. user-corrections pending — soft 20   / hard 40
+#   4. Benchmarks freshness     — auto-fetch at 30 days(no hard tier)
+
+# Per-hook fire logging(enables /knowledge-prune D2 dead-hook detection)
+source "$(dirname "$0")/_log-fire.sh" 2>/dev/null && log_hook_fire
+
+# 2026-05-30:移除 -e(對齊 L2 set -uo)。本 hook = 非阻塞 governance reminder,必永遠 exit 0;
+# set -e 會讓任何未 guard 的 command(eg. L38 grep no-match pipeline 在 node-missing degraded env)
+# 殺掉整個 session-start hook。fail-open > fail-closed。adversarial-verify 2026-05-30 抓出此真 bug。
+set -uo pipefail
+
+PROJECT_DIR="${CLAUDE_PROJECT_DIR:-$(pwd)}"
+cd "$PROJECT_DIR" || exit 0
+
+# 2026-05-26 folded auto_sync_memory.sh in here(retire that hook for D8a budget):
+# Per CLAUDE.md governance hook count budget,SessionStart-only utility helper inline
+# under main governance hook 比 separate hook 更合 anti-bloat。
+# Memory sync = harness ↔ repo mirror auto-fix-up,not blocking,always exit 0。
+if [ -f scripts/sync-memory.mjs ]; then
+  SYNC_OUT=$(node scripts/sync-memory.mjs 2>&1 || true)
+  COPIED=$(echo "$SYNC_OUT" | grep -oE 'copied: [0-9]+' | grep -oE '[0-9]+' | head -1 || true)
+  # COPIED > 0 → 加入 REMINDERS(下面 main flow inject)
+  if [ -n "$COPIED" ] && [ "$COPIED" -gt 0 ]; then
+    MEMSYNC_NOTE="\n- 🔄 auto sync-memory(SessionStart): harness → repo mirrored ${COPIED} memory file(s)。"
+  fi
+fi
+
+REMINDERS="${MEMSYNC_NOTE:-}"
+BLOCKERS=""
+
+# Check 1: CLAUDE.md size(soft 800 / hard 1000)
+if [ -f CLAUDE.md ]; then
+  LINES=$(wc -l < CLAUDE.md | tr -d ' ')
+  # 2026-04-26 tightened thresholds(對應 M19 + user 質問「auto self-improve」要更主動):
+  # SSOT alignment per CLAUDE.md L35: target ≤ 200 / transition ≤ 400 / hard cap 800
+  # Stratified soft warnings: 500 approaching / 600 strong / 800 hard-cap blocker
+  if [ "$LINES" -gt 800 ]; then
+    BLOCKERS="${BLOCKERS}\n- CLAUDE.md is ${LINES} lines(hard cap 800 breached). /knowledge-prune REQUIRED FIRST ACTION this session."
+  elif [ "$LINES" -gt 600 ]; then
+    REMINDERS="${REMINDERS}\n- CLAUDE.md is ${LINES} lines(50% over 400 transition cap). /knowledge-prune strongly recommended this session."
+  elif [ "$LINES" -gt 500 ]; then
+    REMINDERS="${REMINDERS}\n- CLAUDE.md is ${LINES} lines(approaching 600 strong-warn). Consider /knowledge-prune."
+  fi
+fi
+
+# Check 2: Last /knowledge-prune commit(soft 90d / hard 180d)
+LAST_PRUNE=$(git log --format='%ct' --grep='knowledge-prune\|prune:\|governance.*prune' -1 2>/dev/null || echo "")
+if [ -n "$LAST_PRUNE" ]; then
+  NOW=$(date +%s)
+  DAYS=$(( (NOW - LAST_PRUNE) / 86400 ))
+  if [ "$DAYS" -gt 180 ]; then
+    BLOCKERS="${BLOCKERS}\n- Last /knowledge-prune was ${DAYS} days ago(HARD THRESHOLD 180 breached). /knowledge-prune REQUIRED FIRST ACTION this session."
+  elif [ "$DAYS" -gt 90 ]; then
+    REMINDERS="${REMINDERS}\n- Last /knowledge-prune commit was ${DAYS} days ago (target: quarterly ≤ 90 days)."
+  fi
+fi
+
+# Check 3: user-corrections pending(soft 20 / hard 40)
+CORRECTIONS_LOG="$PROJECT_DIR/.claude/logs/user-corrections.jsonl"
+if [ -f "$CORRECTIONS_LOG" ]; then
+  CORRECTION_COUNT=$(wc -l < "$CORRECTIONS_LOG" | tr -d ' ')
+  if [ "$CORRECTION_COUNT" -gt 40 ]; then
+    BLOCKERS="${BLOCKERS}\n- ${CORRECTION_COUNT} user-corrections pending(HARD THRESHOLD 40 breached). /codify-corrections REQUIRED FIRST ACTION this session."
+  elif [ "$CORRECTION_COUNT" -gt 20 ]; then
+    REMINDERS="${REMINDERS}\n- ${CORRECTION_COUNT} user-correction signals pending codification (.claude/logs/user-corrections.jsonl). Review + codify pending ones."
+  fi
+fi
+
+# Check 4: benchmarks freshness — AUTO-RUN fetcher if > 30 days or never fetched
+# (對齊 M14 AUTO integrate pipeline — external signal refresh 不等 user 提醒)
+BENCH_DIR="$PROJECT_DIR/.claude/benchmarks"
+if [ -d "$BENCH_DIR" ]; then
+  LAST_FETCH_FILE="$BENCH_DIR/last-fetch.txt"
+  SHOULD_AUTO_FETCH=0
+
+  if [ ! -f "$LAST_FETCH_FILE" ]; then
+    SHOULD_AUTO_FETCH=1
+    FETCH_REASON="never fetched"
+  else
+    LAST_TS=$(stat -f '%m' "$LAST_FETCH_FILE" 2>/dev/null || echo "0")
+    NOW=$(date +%s)
+    DAYS=$(( (NOW - LAST_TS) / 86400 ))
+    if [ "$DAYS" -gt 30 ]; then
+      SHOULD_AUTO_FETCH=1
+      FETCH_REASON="${DAYS} days stale"
+    fi
+  fi
+
+  # Background fetch(不 block session 起動,寫結果供下次 session 用)
+  # 只有 fetcher 存在才跑,容忍網路錯誤(fetch.sh 內建 fail-silent)
+  if [ "$SHOULD_AUTO_FETCH" = "1" ] && [ -x "$BENCH_DIR/fetch.sh" ]; then
+    (bash "$BENCH_DIR/fetch.sh" > "$BENCH_DIR/.last-auto-fetch.log" 2>&1 &) 2>/dev/null
+    REMINDERS="${REMINDERS}\n- Benchmarks ${FETCH_REASON} → auto-fetching in background(結果寫 .last-auto-fetch.log,下次 session 生效)"
+  elif [ "$SHOULD_AUTO_FETCH" = "1" ]; then
+    REMINDERS="${REMINDERS}\n- Benchmarks ${FETCH_REASON} → manually run \`bash .claude/benchmarks/fetch.sh\`"
+  fi
+fi
+
+# Check 6: Fix commits without /scan-similar-bugs invoke(M10 mechanical 落地)
+# Detect 24h 內 fix( commit 但 skill-invokes log 沒對應 scan-similar-bugs invoke
+RECENT_FIX_COMMITS=$(git log --since='24 hours ago' --grep='^fix(\|^bugfix:\|^fix:' --oneline 2>/dev/null | head -5 || true)
+if [ -n "$RECENT_FIX_COMMITS" ]; then
+  SKILL_LOG="$PROJECT_DIR/.claude/logs/skill-invokes.jsonl"
+  RECENT_SCAN_INVOKE=0
+  if [ -f "$SKILL_LOG" ]; then
+    # Check 24h 內有 scan-similar-bugs invoke
+    NOW_EPOCH=$(date -u +%s)
+    DAY_AGO_EPOCH=$(( NOW_EPOCH - 86400 ))
+    while IFS= read -r line; do
+      [ -z "$line" ] && continue
+      if echo "$line" | grep -q "scan-similar-bugs"; then
+        TS=$(echo "$line" | jq -r '.ts // empty' 2>/dev/null)
+        if [ -n "$TS" ]; then
+          TS_EPOCH=$(date -j -f "%Y-%m-%dT%H:%M:%SZ" "$TS" +%s 2>/dev/null || echo 0)
+          if [ "$TS_EPOCH" -gt "$DAY_AGO_EPOCH" ]; then
+            RECENT_SCAN_INVOKE=1
+            break
+          fi
+        fi
+      fi
+    done < "$SKILL_LOG"
+  fi
+  if [ "$RECENT_SCAN_INVOKE" = "0" ]; then
+    FIX_LIST=$(echo "$RECENT_FIX_COMMITS" | sed 's/^/  - /' | head -3)
+    REMINDERS="${REMINDERS}\n- 24h 內 fix commit(s) 未 follow /scan-similar-bugs(M10 mechanical 落地):\n${FIX_LIST}\n  考慮 invoke /scan-similar-bugs 確認 DS-wide 同類 bug 全清。"
+  fi
+fi
+
+# Check 5: Fire-weighted test gap(G7)— hooks with fires > 100 but no test
+FIRES_LOG="$PROJECT_DIR/.claude/logs/hook-fires-per-hook.jsonl"
+TESTS_DIR="$PROJECT_DIR/.claude/hooks/tests"
+if [ -f "$FIRES_LOG" ] && [ -d "$TESTS_DIR" ]; then
+  # Top hot hooks by fire count(讀近 10000 lines 防 log 太長跑太久)
+  HOT_HOOKS=$(tail -10000 "$FIRES_LOG" 2>/dev/null \
+    | jq -r '.hook // empty' 2>/dev/null \
+    | sort | uniq -c | sort -rn \
+    | awk '$1 > 100 { sub(/\.sh$|\.py$/, "", $2); print $1, $2 }')
+  HOT_GAPS=""
+  while IFS= read -r line; do
+    [ -z "$line" ] && continue
+    COUNT=$(echo "$line" | awk '{print $1}')
+    NAME=$(echo "$line" | awk '{print $2}')
+    if [ ! -f "$TESTS_DIR/test_${NAME}.sh" ]; then
+      HOT_GAPS="${HOT_GAPS}\n  - ${NAME}(${COUNT} fires,無 test)"
+    fi
+  done <<< "$HOT_HOOKS"
+  if [ -n "$HOT_GAPS" ]; then
+    REMINDERS="${REMINDERS}\n- Fire-weighted test gap(hook > 100 fires 仍無 test):${HOT_GAPS}\n  推力:寫 \`.claude/hooks/tests/test_<name>.sh\`,參考 test_check_story_anatomy.sh。"
+  fi
+fi
+
+# Check 7: Hook count auto-trigger(soft 26 / hard 60 — Anthropic guideline ~15;真值見下方 -gt 判斷 + CLAUDE.md `# 治理 canonical`)
+# 2026-05-09 fix:tree-recursive count(含 lib/ helpers)。前身 -maxdepth 1 只 count root,
+# 漏 16 個 lib/ helpers → metric reports 19,reality 35 = system gaming own metric。
+# 2026-05-13 prune consolidation:`_*.sh` 約定 = internal helper(Unix convention)。
+# Lib helpers renamed `_<name>.sh` 並 fold under `post_edit_dispatcher.sh`(1 registered hook)。
+# Find rule 排除 `_*` 因 _ prefix 表「非 first-class hook,由 dispatcher orchestrate」—
+# 對齊 `_log-fire.sh` 既有 exclusion 原則。Settings.json 中 _* files 不可作 hook registration。
+# 排除:retired/ + tests/ + _* internal helpers.
+PRUNE_TRIGGERS=""
+HOOKS_DIR="$PROJECT_DIR/.claude/hooks"
+HOOK_COUNT=0
+if [ -d "$HOOKS_DIR" ]; then
+  HOOK_COUNT=$(find "$HOOKS_DIR" \( -name "*.sh" -o -name "*.py" \) \
+    -not -path "*/retired/*" -not -path "*/tests/*" -not -name "_*" \
+    2>/dev/null | wc -l | tr -d ' ')
+  HOOK_COUNT=${HOOK_COUNT:-0}
+fi
+# 2026-06-11 deep-audit R2(n=8)重估:訊息原以 3 consumer hooks(check_consumer_no_ds_catalog /
+# check_consumer_story_baseline / check_consumer_ds_primitive_misuse)作 60 cap rationale,該 3 hook
+# 已 2026-06-11 prune-merge 入 check_consumer_app_invariants.sh(retired/2026-06-11-prune-merge/),
+# 現值 52。重估 verdict:60 維持 — 52 + 8 headroom(~15%)符合「升 cap 只為已 justified 新 hook」
+# 歷史節奏;降 cap 屬治理 substantive(soft 26 已在 27+ 提供 advisory),留 /knowledge-prune 評估。
+if [ "$HOOK_COUNT" -gt 60 ]; then
+  BLOCKERS="${BLOCKERS}\n- Hook count ${HOOK_COUNT}(hard 60 — Anthropic guideline ~15;含 root + lib/,排 retired/tests/). 2026-05-27 升 50→55→60(當時 3 consumer hooks ship;該 3 hook 已 2026-06-11 prune-merge 入 check_consumer_app_invariants,現值基準 52,cap 60 經 2026-06-11 重估維持)。超 60 = 先跑 /knowledge-prune 評估 retire / consolidate,不直接 re-raise。"
+elif [ "$HOOK_COUNT" -gt 26 ]; then
+  # 2026-05-15 raised soft cap 25→26 per /knowledge-prune D2 audit:
+  # 26 wired hooks reflects M30 wrapper-schema-drift 新增 dedicated hook(justified evolution
+  # not bloat,per Task #19 closed analysis)。Re-raise to 27+ 需 audit re-justify。
+  PRUNE_TRIGGERS="${PRUNE_TRIGGERS}\n- Hook count ${HOOK_COUNT}(soft 26 trigger,2026-05-15 升 — Anthropic guideline ~15;含 root + lib/). /knowledge-prune 評估 retire / consolidate 候選."
+fi
+
+# Check 8: Memory entries auto-trigger(soft 18 / hard 20)
+# Path resolution(2026-05-08 fix):isolation-friendly precedence —
+#   1. PROJECT_DIR/.claude/memory(讓 test sandbox CLAUDE_PROJECT_DIR 隔離 work)
+#   2. harness user-local SSOT(實際 production 路徑)
+#   3. repo mirror fallback(cloud sandbox snapshot)
+HARNESS_MEMORY_DIR="$HOME/.claude/projects/-Users-chenqiren-Library-CloudStorage-GoogleDrive-qijenchen-gmail-com--------my-project/memory"
+REPO_MEMORY_DIR="$PROJECT_DIR/.claude/memory"
+if [ -d "$REPO_MEMORY_DIR" ]; then
+  MEMORY_DIR="$REPO_MEMORY_DIR"
+elif [ -d "$HARNESS_MEMORY_DIR" ]; then
+  MEMORY_DIR="$HARNESS_MEMORY_DIR"
+else
+  MEMORY_DIR=""
+fi
+MEM_COUNT=0
+if [ -n "$MEMORY_DIR" ] && [ -d "$MEMORY_DIR" ]; then
+  MEM_COUNT=$(find "$MEMORY_DIR" -maxdepth 1 -name "*.md" -not -name "MEMORY.md" -not -name "README.md" 2>/dev/null | wc -l | tr -d ' ')
+  MEM_COUNT=${MEM_COUNT:-0}
+fi
+if [ "$MEM_COUNT" -gt 20 ]; then
+  BLOCKERS="${BLOCKERS}\n- Memory entries ${MEM_COUNT}(hard 20 cap). /knowledge-prune REQUIRED FIRST ACTION."
+elif [ "$MEM_COUNT" -ge 18 ]; then
+  PRUNE_TRIGGERS="${PRUNE_TRIGGERS}\n- Memory entries ${MEM_COUNT}(soft 18 trigger,20 = hard cap). /knowledge-prune 建議排程."
+fi
+
+# Check 9: Branch sprawl(M28 — solo work = 1 chat = 1 branch / 0-1 active feature)
+# 開場掃 local + remote claude/* branch,> 1 active 或 local main divergent → warn。
+LOCAL_CLAUDE_BRANCHES=$(git -C "$PROJECT_DIR" branch 2>/dev/null | grep -c "claude/" || true)
+REMOTE_CLAUDE_BRANCHES=$(git -C "$PROJECT_DIR" branch -r 2>/dev/null | grep -cE "origin/claude/" || true)
+if [ "$LOCAL_CLAUDE_BRANCHES" -gt 1 ]; then
+  PRUNE_TRIGGERS="${PRUNE_TRIGGERS}\n- Local branch sprawl ${LOCAL_CLAUDE_BRANCHES} active claude/* branches(M28: 1 session = 1 branch). git branch -d <merged> 清掉,只留當前 active feature."
+fi
+if [ "$REMOTE_CLAUDE_BRANCHES" -gt 1 ]; then
+  PRUNE_TRIGGERS="${PRUNE_TRIGGERS}\n- Remote branch sprawl ${REMOTE_CLAUDE_BRANCHES} stale claude/* branches on origin(sandbox HTTP 403 攔 push --delete). User GitHub UI 手動刪 OR 'git push origin --delete <branch>'."
+fi
+# Local main divergent from origin/main(sandbox commits 殘留 / 未 merge)
+if git -C "$PROJECT_DIR" rev-parse origin/main >/dev/null 2>&1; then
+  AHEAD_BEHIND=$(git -C "$PROJECT_DIR" rev-list --left-right --count main...origin/main 2>/dev/null || echo "0	0")
+  AHEAD=$(echo "$AHEAD_BEHIND" | cut -f1)
+  BEHIND=$(echo "$AHEAD_BEHIND" | cut -f2)
+  if [ -n "$AHEAD" ] && [ -n "$BEHIND" ]; then
+    if [ "$AHEAD" -gt 0 ] || [ "$BEHIND" -gt 0 ]; then
+      PRUNE_TRIGGERS="${PRUNE_TRIGGERS}\n- Local main divergent (ahead ${AHEAD} / behind ${BEHIND} vs origin/main). 'git fetch && git checkout main && git reset --hard origin/main' 對齊。"
+    fi
+  fi
+fi
+
+# Check 10: SSOT auto-sync drift(2026-05-23 — sync-governance-counters drift report)
+SSOT_DRIFT=""
+if command -v node >/dev/null 2>&1 && [ -f scripts/sync-governance-counters.mjs ]; then
+  DRIFT_OUT=$(node scripts/sync-governance-counters.mjs --quiet 2>&1 || true)
+  if echo "$DRIFT_OUT" | grep -q "Hardcoded drift detected"; then
+    SSOT_DRIFT=$(echo "$DRIFT_OUT" | sed -n '/Hardcoded drift detected/,/^Log:/p' | grep '^  - ' | head -8)
+    if [ -n "$SSOT_DRIFT" ]; then
+      PRUNE_TRIGGERS="${PRUNE_TRIGGERS}\n- SSOT drift detected:\n${SSOT_DRIFT}\n  → 跑 node scripts/sync-governance-counters.mjs 看完整 + 對齊 SSOT 數字 / npm scope / plugin manifest。"
+    fi
+  fi
+fi
+
+# Check 11: Cross-repo env smoke(2026-05-30 — NON-BLOCKING:只進 PRUNE_TRIGGERS soft channel,
+# 永不進 BLOCKER 路徑、永不非零退出。set -uo pipefail(無 -e)→ 探針非零返回不殺 script;
+# 但 set -u 下 unset var 必 ${VAR:-} guard。每探針獨立、無 network、無 blocking subshell。
+ENV_SMOKE=""
+# (a) plugin-mode 完整性 — CLAUDE_PLUGIN_ROOT 在 ds-repo native mode 可能 UNSET → 先 guard 才用
+if [ -n "${CLAUDE_PLUGIN_ROOT:-}" ] && [ ! -d "${CLAUDE_PLUGIN_ROOT:-}/hooks" ]; then
+  ENV_SMOKE="${ENV_SMOKE}\n    - Plugin mode:\$CLAUDE_PLUGIN_ROOT 有設但 hooks/ 找不到 → plugin install 可能不完整(C-prime 主路徑 = committed-config + \`npm run sync-all\`〔npm-only〕;刻意用 plugin 才跑 /plugin marketplace update)。"
+fi
+# (b) node 在 PATH — audit scripts 依賴
+if ! command -v node >/dev/null 2>&1; then
+  ENV_SMOKE="${ENV_SMOKE}\n    - node 不在 PATH → audit scripts(dispatch-audit-dims / content-quality)在此環境跑不動。"
+fi
+# (c) codex transport(informational;缺 = fork repo 正常,Phase B 自動 fallback)
+if [ ! -x node_modules/.bin/codex ] && ! command -v codex >/dev/null 2>&1; then
+  ENV_SMOKE="${ENV_SMOKE}\n    - codex CLI 缺 → deep-audit Phase B dual-track 自動 fallback Phase-A-only(fork repo 屬正常;要雙軌跑 npm i -D @openai/codex)。"
+fi
+# (d) consumer-mode DS resolution — fork repo 引 npm DS 但未安裝
+if [ -f package.json ] && grep -q '"@qijenchen/design-system"' package.json 2>/dev/null && [ ! -d node_modules/@qijenchen/design-system ]; then
+  ENV_SMOKE="${ENV_SMOKE}\n    - Consumer repo 引用 @qijenchen/design-system 但 node_modules 沒裝(跑 npm install)。"
+fi
+if [ -n "$ENV_SMOKE" ]; then
+  PRUNE_TRIGGERS="${PRUNE_TRIGGERS}\n- 🩺 env-smoke(non-blocking,環境健檢):${ENV_SMOKE}"
+fi
+
+# Inject if HARD BLOCKERS(must)or auto-prune-triggers or quarterly-prune-overdue
+QUARTERLY_DUE=""
+if [ -f .claude/logs/.last-prune ]; then
+  PRUNE_DAYS=$(( ( $(date +%s) - $(stat -f %m .claude/logs/.last-prune 2>/dev/null || stat -c %Y .claude/logs/.last-prune 2>/dev/null || echo 0) ) / 86400 ))
+  [ "$PRUNE_DAYS" -ge 90 ] && QUARTERLY_DUE="\n- Last /knowledge-prune ${PRUNE_DAYS} days ago(quarterly target ≤ 90 days). Invoke /knowledge-prune this session if convenient."
+fi
+
+[ -z "$BLOCKERS" ] && [ -z "$QUARTERLY_DUE" ] && [ -z "$PRUNE_TRIGGERS" ] && exit 0
+
+if [ -n "$BLOCKERS" ]; then
+  MSG="🚨 BLOCKER — governance hard thresholds breached (SessionStart):${BLOCKERS}\n\n"
+  MSG="${MSG}⚠️ REQUIRED_FIRST_ACTION:先 invoke 上述 skill(/knowledge-prune 或 /codify-corrections)"
+  MSG="${MSG}把 governance 帶回健康區間,再處理 user 的實際請求。"
+  [ -n "$PRUNE_TRIGGERS" ] && MSG="${MSG}\n\n附 soft prune triggers:${PRUNE_TRIGGERS}"
+  [ -n "$QUARTERLY_DUE" ] && MSG="${MSG}\n${QUARTERLY_DUE}"
+elif [ -n "$PRUNE_TRIGGERS" ]; then
+  MSG="⚙️ Auto-prune triggers fired (SessionStart):${PRUNE_TRIGGERS}\nAI 必 AUTO-RUN /knowledge-prune(禁問 user 要不要跑;P0+P1 自動執行,P2 retire 候選列拍板清單 — per deep-audit-cross-codex SKILL C.0a,2026-06-11 user 糾正 codify).${QUARTERLY_DUE}"
+else
+  MSG="🧭 Governance hygiene reminder (SessionStart):${QUARTERLY_DUE}\nNot blocking — address inline when convenient."
+fi
+# fail-open:無 jq(eg. 殘缺 PATH / minimal cloud sandbox)→ 靜默 exit 0,不吐 malformed JSON(2026-05-30 硬化)
+command -v jq >/dev/null 2>&1 || exit 0
+ESCAPED=$(printf '%b' "$MSG" | jq -Rs .)
+printf '{"hookSpecificOutput":{"hookEventName":"SessionStart","additionalContext":%s}}\n' "$ESCAPED"
+exit 0