@routinecrew/routinecode 2.31.45 → 2.31.47

package/hooks/lib/known_solution.py

@@ -0,0 +1,148 @@
+# -*- coding: utf-8 -*-
+"""[KNOWN-SOLUTION] 주입 공유 모듈 — error-fix 지식(events.jsonl)의 결정적 소비 경로.
+
+설계: .routinecode/plans/design-knowledge-consumption-loop.md (design score 96)
+사용처: user-prompt-submit.sh(L1, additionalContext) / post-bash-knowledge.sh(L2, systemMessage)
+
+원칙: 보수 이중 문턱(12자+ 공통 부분열 OR 토큰 자카드>=0.5), 1건만, 저장된 지식 필드만 인용
+(사용자 프롬프트 원문 비포함), 세션x시그니처 1회(dedup 마커), 실패 시 조용히 None(fail-open).
+"""
+import hashlib
+import json
+import os
+import re
+import time
+
+KN_DIR = os.path.expanduser('~/.routinecode/knowledge')
+EVENTS = os.path.join(KN_DIR, 'events.jsonl')
+MAX_ROWS = 5000
+
+ERROR_HINT = re.compile(
+    r'(Error|에러|실패|Exception|Cannot |undefined|TypeError|Traceback|ERR_|error:)', re.I)
+
+
+def _normalize(s):
+    s = str(s or '').lower()
+    s = re.sub(r'/[\w./\-]{8,}', '*', s)        # 절대경로 → *
+    s = re.sub(r'\d+', '#', s)                   # 숫자 → #
+    s = re.sub(r'\s+', ' ', s).strip()
+    return s
+
+
+def _tokens(s):
+    return set(re.findall(r'[a-z가-힣_]{3,}', s))
+
+
+def _jaccard(a, b):
+    if not a or not b:
+        return 0.0
+    return len(a & b) / len(a | b)
+
+
+def _common_substr_len(a, b):
+    """최장 공통 부분열 길이 (O(n*m) — 시그니처 400자 캡이라 충분히 빠름)."""
+    a, b = a[:400], b[:400]
+    if not a or not b:
+        return 0
+    prev = [0] * (len(b) + 1)
+    best = 0
+    for i in range(1, len(a) + 1):
+        cur = [0] * (len(b) + 1)
+        for j in range(1, len(b) + 1):
+            if a[i - 1] == b[j - 1]:
+                cur[j] = prev[j - 1] + 1
+                if cur[j] > best:
+                    best = cur[j]
+        prev = cur
+    return best
+
+
+def _tenant_ok(entry_tid):
+    """pre-read-inject _tenant_ok와 동일 의미(무태그=통과). 현 events.jsonl 전행 무태그 — 미래 대비."""
+    legacy = entry_tid in (None, 0, '0', '')
+    if legacy:
+        return True
+    try:
+        cfg = json.load(open(os.path.expanduser('~/.routinecode/config.json')))
+        my = cfg.get('tenantId')
+    except Exception:
+        my = None
+    if my in (None, 0, '0', ''):
+        return False
+    return str(entry_tid) == str(my)
+
+
+def find_known_solution(error_text):
+    """에러 텍스트 → 최고 매칭 error-fix 1건 또는 None."""
+    try:
+        if not error_text or not ERROR_HINT.search(error_text):
+            return None
+        if not os.path.exists(EVENTS):
+            return None
+        q_norm = _normalize(error_text)
+        q_tok = _tokens(q_norm)
+        best, best_score = None, 0.0
+        n = 0
+        with open(EVENTS, errors='ignore') as f:
+            for line in f:
+                n += 1
+                if n > MAX_ROWS:
+                    break
+                try:
+                    e = json.loads(line)
+                except Exception:
+                    continue
+                sig = e.get('errorSignature')
+                if not sig or not e.get('rootCause') or not _tenant_ok(e.get('tenantId')):
+                    continue
+                s_norm = _normalize(sig)
+                jac = _jaccard(q_tok, _tokens(s_norm))
+                sub = _common_substr_len(q_norm, s_norm)
+                if sub >= 12 or jac >= 0.5:
+                    score = jac + min(sub, 40) / 40.0
+                    if score > best_score:
+                        best, best_score = e, score
+        return best
+    except Exception:
+        return None
+
+
+def build_injection(e):
+    root = str(e.get('rootCause', ''))[:120]
+    fix = str(e.get('fixApproach', ''))[:120]
+    txt = '[KNOWN-SOLUTION] 과거 동일 에러 해결 이력 — 원인: %s / 해결: %s' % (root, fix)
+    sid = e.get('serverId')
+    if isinstance(sid, int) or (isinstance(sid, str) and sid.isdigit()):
+        # resolvedFromId 는 number 타입 (mcp-intelligence index.ts:164)
+        txt += ' (검증 시 record_verification resolvedFromId=%d)' % int(sid)
+    return txt
+
+
+def _sig8(e):
+    return hashlib.sha1(_normalize(e.get('errorSignature', '')).encode()).hexdigest()[:8]
+
+
+def consume_once(session_id, error_text):
+    """조회+dedup — 세션x시그니처당 1회만 주입문 반환, 그 외 None."""
+    e = find_known_solution(error_text)
+    if not e:
+        return None
+    marker = os.path.join(KN_DIR, '.ks-%s-%s' % (str(session_id)[:12], _sig8(e)))
+    if os.path.exists(marker):
+        return None
+    try:
+        os.makedirs(KN_DIR, exist_ok=True)
+        open(marker, 'w').write(str(int(time.time())))
+        # GC: 24h 초과 .ks-* 자가 정리 (별도 cron 불요 — 설계 v2-2)
+        cutoff = time.time() - 86400
+        for fn in os.listdir(KN_DIR):
+            if fn.startswith('.ks-'):
+                fp = os.path.join(KN_DIR, fn)
+                try:
+                    if os.path.getmtime(fp) < cutoff:
+                        os.remove(fp)
+                except Exception:
+                    pass
+    except Exception:
+        pass
+    return build_injection(e)

package/hooks/lib/receipt.py

@@ -201,6 +201,8 @@ def render(r):
         parts = []
         if t.get("measured_s"):
             parts.append("실측 %s [병렬 서브에이전트]" % _fmt_dur(t["measured_s"]))
+        if t.get("paired_s"):
+            parts.append("실측쌍 %s [탐색 cold→warm 동일모듈]" % _fmt_dur(t["paired_s"]))
         if t.get("anchored_s"):
             parts.append("앵커 %s [지식·루틴 — 실측앵커 하한]" % _fmt_dur(t["anchored_s"]))
         if parts:

package/hooks/lib/time_savings.py

@@ -23,6 +23,11 @@ _RC = os.path.expanduser('~/.routinecode')
 RATES_PATH = os.path.join(_RC, 'rdpe', 'time-rates.json')
 LEDGER = os.path.join(_RC, 'knowledge', 'time-savings.jsonl')
 TELEMETRY = os.path.join(_RC, 'knowledge', 'agent-telemetry.jsonl')
+# 탐색 시간 cold/warm 장부 — 모듈 단위 실측쌍(같은 모듈의 cold 중앙값 vs warm 실측).
+# "코드 전체 스캔 시간" 절감의 정직한 계산: 양변 전부 실측, 상수 없음.
+EXPLORE_LEDGER = os.path.join(_RC, 'knowledge', 'explore-ledger.jsonl')
+IDLE_GAP_S = 600          # 이 이상 간격은 사람 부재로 보고 탐색 시간에서 제외
+PAIRED_SESSION_CAP_S = 900  # 실측쌍 적립의 세션 상한(보수)
 
 DEFAULT_RATES = {
     "knowledge_error_reuse_s": 60,     # C1 폴백 (보수 하한)
@@ -74,20 +79,39 @@ def collect(transcript_path):
     gate_blocks = 0
     session_start = None
     session_end = None
+    # 훅 주입 소비(C1/C2) — attachment(hook_success) stdout 마커. 주입문 sha1로 dedup
+    # (tool_use 경로와 단일 집합 공유 → 이중계수 차단. 설계 v2-6)
+    injected_c1 = set()
+    injected_c2 = set()
 
     for o in _iter_transcript(transcript_path):
         ts = o.get('timestamp')
         if ts:
             session_start = session_start or ts
             session_end = ts
+        att = o.get('attachment') or {}
+        if att.get('type') in ('hook_success', 'hook_blocking_error'):
+            out = str(att.get('stdout', ''))[:2000]
+            if '[KNOWN-SOLUTION]' in out:
+                import hashlib as _h
+                injected_c1.add(_h.sha1(out.split('[KNOWN-SOLUTION]')[1][:200].encode()).hexdigest())
+            if '[DOMAIN KNOWLEDGE]' in out:
+                import hashlib as _h
+                injected_c2.add(_h.sha1(out.split('[DOMAIN KNOWLEDGE]')[1][:200].encode()).hexdigest())
         raw = None
         for c in (o.get('message', {}).get('content') or []):
             if not isinstance(c, dict):
                 continue
             if c.get('type') == 'tool_use':
                 tool_uses.append((c.get('name', ''), c.get('id', '')))
-                if c.get('name') in ('Edit', 'Write', 'MultiEdit'):
+                if c.get('name') in ('Edit', 'Write', 'MultiEdit', 'NotebookEdit'):
                     had_edit = True
+                elif c.get('name') == 'Bash':
+                    # 모델이 Edit 대신 Bash로 파일을 고치는 경우(sed -i/리다이렉션/tee)도 구현 신호
+                    _cmd = str((c.get('input') or {}).get('command', ''))[:500]
+                    if ('sed -i' in _cmd or ' tee ' in _cmd or '>>' in _cmd
+                            or ('>' in _cmd and ('.ts' in _cmd or '.js' in _cmd or '.py' in _cmd or '.json' in _cmd))):
+                        had_edit = True
             elif c.get('type') == 'tool_result':
                 txt = c.get('content')
                 results[c.get('tool_use_id', '')] = json.dumps(txt, ensure_ascii=False)[:400] if txt else ''
@@ -117,6 +141,10 @@ def collect(transcript_path):
         elif short in C4_MAP:
             c4_s += rates[C4_MAP[short]]
 
+    # 훅 주입 소비 합산 (동일 게이트·상한 적용 — 설계 v2-6)
+    c1 += len(injected_c1)
+    c2 += len(injected_c2)
+
     # 지식 소비(C1/C2)는 이후 실제 구현(Edit/Write)이 있어야 적립 — 조회만으론 0
     if not had_edit:
         c1 = c2 = 0
@@ -131,16 +159,120 @@ def collect(transcript_path):
     # C3 병렬 — agent-telemetry에서 이 세션 시간창 내 배치 실측
     measured_s = _parallel_savings(session_start, session_end)
 
+    # C-explore 실측쌍 — 같은 모듈 cold 중앙값 vs 이번 warm 탐색 (양변 실측, 상수 없음)
+    paired_s = record_explore(transcript_path)
+
     return {
         'at': datetime.now(timezone.utc).isoformat(),
         'transcript': os.path.basename(transcript_path),
         'measured_s': round(measured_s),          # 전액 실측 (병렬)
+        'paired_s': round(paired_s),              # 실측쌍 (cold/warm 동일모듈 탐색)
         'anchored_s': round(anchored_s),          # 실측앵커 상수 기반
         'potential_gate_blocks': gate_blocks,     # C5: 적립 0, 잠재 건수만
         'detail': {'c1_error_reuse': c1, 'c2_domain_use': c2, 'c4_mcp_routine_s': c4_s},
     }
 
 
+def _module_key(path):
+    """Read 파일경로 → 모듈 키 (마지막 2세그먼트 디렉토리). 프로젝트 이동에도 안정."""
+    try:
+        d = os.path.dirname(str(path))
+        segs = [s for s in d.split(os.sep) if s]
+        return '/'.join(segs[-2:]) if segs else ''
+    except Exception:
+        return ''
+
+
+def measure_explore(transcript_path):
+    """탐색 구간 실측: 세션 시작 → 첫 구현 행위까지, idle(>IDLE_GAP_S) 제외.
+    반환 {module, warm, explore_s} 또는 None(구현행위/Read 부재 시 — 장부 미기록)."""
+    first_ts = None
+    prev_ts = None
+    explore_s = 0.0
+    impl_seen = False
+    warm = False
+    reads = {}
+    for o in _iter_transcript(transcript_path):
+        ts = _parse_iso(o.get('timestamp') or '')
+        att = o.get('attachment') or {}
+        if not impl_seen and att.get('type') in ('hook_success', 'hook_system_message'):
+            blob = str(att.get('stdout', '')) + str(att.get('content', ''))
+            if '[DOMAIN KNOWLEDGE]' in blob or '[KNOWN-SOLUTION]' in blob:
+                warm = True
+        for c in (o.get('message', {}).get('content') or []):
+            if not isinstance(c, dict) or c.get('type') != 'tool_use':
+                continue
+            name = c.get('name', '')
+            if name in ('Edit', 'Write', 'MultiEdit', 'NotebookEdit'):
+                impl_seen = True
+            elif name == 'Bash':
+                _cmd = str((c.get('input') or {}).get('command', ''))[:500]
+                if ('sed -i' in _cmd or ' tee ' in _cmd or '>>' in _cmd
+                        or ('>' in _cmd and ('.ts' in _cmd or '.js' in _cmd or '.py' in _cmd or '.json' in _cmd))):
+                    impl_seen = True
+            elif name == 'Read' and not impl_seen:
+                mk = _module_key((c.get('input') or {}).get('file_path', ''))
+                if mk:
+                    reads[mk] = reads.get(mk, 0) + 1
+            elif not impl_seen and ('get_domain_knowledge' in name or 'get_similar_errors' in name
+                                    or 'get_patterns' in name or 'get_known_quirks' in name):
+                warm = True
+        if ts:
+            if first_ts is None:
+                first_ts = ts
+            if not impl_seen and prev_ts is not None:
+                gap = ts - prev_ts
+                if 0 < gap <= IDLE_GAP_S:
+                    explore_s += gap
+            prev_ts = ts
+    if not impl_seen or not reads or explore_s <= 0:
+        return None
+    module = max(reads.items(), key=lambda kv: kv[1])[0]
+    return {'module': module, 'warm': warm, 'explore_s': round(explore_s, 1)}
+
+
+def _explore_cold_median(module):
+    """같은 모듈의 cold 탐색 실측 중앙값 (장부 이력에서)."""
+    vals = []
+    if not os.path.exists(EXPLORE_LEDGER):
+        return None
+    for line in open(EXPLORE_LEDGER, errors='ignore'):
+        try:
+            r = json.loads(line)
+        except Exception:
+            continue
+        if r.get('module') == module and not r.get('warm') and r.get('explore_s', 0) > 0:
+            vals.append(float(r['explore_s']))
+    if not vals:
+        return None
+    vals.sort()
+    return vals[len(vals) // 2]
+
+
+def record_explore(transcript_path):
+    """탐색 장부 기록 + warm이면 cold 중앙값과의 실측쌍 절감 반환(초)."""
+    m = measure_explore(transcript_path)
+    if not m:
+        return 0.0
+    paired = 0.0
+    if m['warm']:
+        cold = _explore_cold_median(m['module'])
+        if cold is not None and cold > m['explore_s']:
+            paired = min(cold - m['explore_s'], PAIRED_SESSION_CAP_S)
+    try:
+        os.makedirs(os.path.dirname(EXPLORE_LEDGER), exist_ok=True)
+        with open(EXPLORE_LEDGER, 'a') as f:
+            f.write(json.dumps({
+                'at': datetime.now(timezone.utc).isoformat(),
+                'transcript': os.path.basename(transcript_path),
+                'module': m['module'], 'warm': m['warm'],
+                'explore_s': m['explore_s'], 'paired_saved_s': round(paired),
+            }, ensure_ascii=False) + '\n')
+    except Exception:
+        pass
+    return paired
+
+
 def _parse_iso(s):
     try:
         return datetime.fromisoformat(s.replace('Z', '+00:00')).timestamp()
@@ -180,7 +312,8 @@ def _parallel_savings(session_start, session_end):
 
 
 def append_ledger(snap):
-    if snap['measured_s'] <= 0 and snap['anchored_s'] <= 0 and snap['potential_gate_blocks'] <= 0:
+    if (snap['measured_s'] <= 0 and snap['anchored_s'] <= 0
+            and snap.get('paired_s', 0) <= 0 and snap['potential_gate_blocks'] <= 0):
         return False
     os.makedirs(os.path.dirname(LEDGER), exist_ok=True)
     with open(LEDGER, 'a') as f:
@@ -190,7 +323,7 @@ def append_ledger(snap):
 
 def totals():
     """영수증(receipt.py)용 누계."""
-    t = {'measured_s': 0, 'anchored_s': 0, 'potential_gate_blocks': 0, 'sessions': 0}
+    t = {'measured_s': 0, 'paired_s': 0, 'anchored_s': 0, 'potential_gate_blocks': 0, 'sessions': 0}
     if not os.path.exists(LEDGER):
         return t
     for line in open(LEDGER, errors='ignore'):
@@ -199,6 +332,7 @@ def totals():
         except Exception:
             continue
         t['measured_s'] += r.get('measured_s', 0)
+        t['paired_s'] += r.get('paired_s', 0)
         t['anchored_s'] += r.get('anchored_s', 0)
         t['potential_gate_blocks'] += r.get('potential_gate_blocks', 0)
         t['sessions'] += 1

package/hooks/post-bash-knowledge.sh

@@ -0,0 +1,45 @@
+#!/bin/bash
+# KS(L2): Bash 도구 출력에 에러가 보이면 과거 해결 이력([KNOWN-SOLUTION])을 systemMessage로 주입.
+# 설계: .routinecode/plans/design-knowledge-consumption-loop.md (score 96)
+# - 조회/매칭/dedup 전부 lib/known_solution.py 공유 (L1 user-prompt-submit와 동일 모듈)
+# - PostToolUse에서 systemMessage는 유효 (post-build-check.sh와 동일 포맷)
+# - fail-open: 어떤 실패도 {"continue":true} 무주입
+
+INPUT=$(cat)
+HOOKS_DIR="$(cd "$(dirname "$0")" && pwd)"
+
+ROUTINECODE_HOOK_INPUT="$INPUT" _HOOK_LIB_DIR="$HOOKS_DIR/lib" python3 - <<'PYEOF' 2>/dev/null || echo '{"continue":true}'
+import json, os, re, sys
+
+def emit(obj):
+    print(json.dumps(obj, ensure_ascii=False))
+    sys.exit(0)
+
+try:
+    inp = json.loads(os.environ.get('ROUTINECODE_HOOK_INPUT', '') or '{}')
+except Exception:
+    emit({"continue": True})
+
+session_id = inp.get('session_id') or 'unknown'
+resp = inp.get('tool_response')
+text = json.dumps(resp, ensure_ascii=False) if not isinstance(resp, str) else resp
+if not text:
+    emit({"continue": True})
+
+# 에러 라인만 추출 (마지막 3줄) — 전체 출력 전달 금지(노이즈·성능)
+err_lines = [l for l in text.split('\\n') if re.search(r'(Error|error:|에러|Traceback|Cannot |ERR_|Exception)', l)]
+if not err_lines:
+    emit({"continue": True})
+probe = ' '.join(err_lines[-3:])[:400]
+
+try:
+    sys.path.insert(0, os.environ.get('_HOOK_LIB_DIR', ''))
+    import known_solution
+    msg = known_solution.consume_once(session_id, probe)
+except Exception:
+    msg = None
+
+if msg:
+    emit({"continue": True, "systemMessage": msg})
+emit({"continue": True})
+PYEOF