mcp-server/src/legal_mcp/services/claude_session.py

@@ -60,6 +60,31 @@ _SESSION_MARKER_EXACT = frozenset({"AI_AGENT", "CLAUDE_EFFORT"})
 MAX_RETRIES = 3
 RETRY_BACKOFF_BASE = 5  # seconds; sleep = base * attempt_number
 
+# Phrases the CLI emits as the "result" of an exit-0 run that actually hit a
+# usage/rate limit (a refusal NOTICE, not a real answer). Matched only against a
+# result that is NOT structured output (doesn't start with [ or {), so a genuine
+# JSON extraction containing these words as content is never mis-flagged.
+_LIMIT_NOTICE_MARKERS = (
+    "usage limit",
+    "rate limit",
+    "limit reached",
+    "limit will reset",
+    "try again later",
+)
+
+
+def _looks_like_limit_notice(data: dict) -> bool:
+    """True if an exit-0 ``result`` is really a usage/rate-limit refusal."""
+    result = data.get("result")
+    if not isinstance(result, str):
+        return False
+    stripped = result.lstrip()
+    # Structured output (JSON array/object) is a real answer, never a notice.
+    if stripped.startswith(("[", "{")):
+        return False
+    low = result.lower()
+    return any(m in low for m in _LIMIT_NOTICE_MARKERS)
+
 
 def _clean_subprocess_env() -> dict[str, str]:
     """Copy the current env minus Claude Code session markers.
@@ -187,11 +212,27 @@ async def query(
                 try:
                     data = json.loads(stdout)
                     if isinstance(data, dict) and "result" in data:
-                        return data["result"]
-                    return stdout
+                        # A usage/rate-limit hit can exit 0 with a refusal NOTICE
+                        # as the "result" (is_error / error subtype). Returning it
+                        # as success makes callers treat a throttled run as a real
+                        # empty answer — e.g. the halacha extractor then checkpoints
+                        # the chunk as done-with-0-halachot and a resume skips it
+                        # forever (#138/#144 silent under-extraction). Treat it as a
+                        # transient failure → retry, and raise if it persists so the
+                        # chunk stays un-checkpointed for a real resume.
+                        if data.get("is_error") or _looks_like_limit_notice(data):
+                            last_err = (
+                                f"error result (subtype={data.get('subtype')}): "
+                                f"{str(data.get('result',''))[:200]}"
+                            )
+                        else:
+                            return data["result"]
+                    else:
+                        return stdout
                 except json.JSONDecodeError:
                     return stdout
-            last_err = "empty response"
+            else:
+                last_err = "empty response"
 
         # Transient failure — retry with linear backoff unless this was the last try.
         if attempt < MAX_RETRIES:

mcp-server/tests/test_claude_session_limit_notice.py

@@ -0,0 +1,40 @@
+"""Regression test for #144 — a usage/rate-limit refusal returned with exit 0
+must NOT be treated as a real (empty) answer.
+
+``_looks_like_limit_notice`` recognises a prose refusal in the CLI's exit-0
+``result`` so ``query`` raises instead of returning it — keeping the halacha
+extractor from checkpointing the chunk as done-with-0-halachot (silent
+under-extraction). A genuine JSON result (array/object) is never mis-flagged.
+"""
+
+from __future__ import annotations
+
+from legal_mcp.services import claude_session as cs
+
+
+def test_detects_usage_limit_prose():
+    assert cs._looks_like_limit_notice(
+        {"result": "Claude usage limit reached. Try again later."})
+    assert cs._looks_like_limit_notice(
+        {"result": "You've hit the rate limit for this model."})
+
+
+def test_json_array_result_is_not_a_notice():
+    # A real halacha extraction — must pass through as a valid answer.
+    assert not cs._looks_like_limit_notice(
+        {"result": '[{"rule_statement": "the usage limit of a permit ..."}]'})
+
+
+def test_json_object_result_is_not_a_notice():
+    assert not cs._looks_like_limit_notice(
+        {"result": '{"summary": "rate limit doctrine"}'})
+
+
+def test_plain_answer_without_markers_is_not_a_notice():
+    assert not cs._looks_like_limit_notice({"result": "אין הלכות בקטע זה."})
+
+
+def test_missing_or_nonstring_result():
+    assert not cs._looks_like_limit_notice({})
+    assert not cs._looks_like_limit_notice({"result": None})
+    assert not cs._looks_like_limit_notice({"result": 42})

scripts/SCRIPTS.md

@@ -33,6 +33,7 @@
 | `drain_metadata_queue.py` | python | **ריקון תור חילוץ-המטא של הפסיקה** — `process_pending_extractions(kind='metadata')` ב-batches עד ריק. רץ על **Gemini Flash** (structured JSON, `gemini_session`) — מהיר ואמין, במקום ה-claude CLI ה-agentic שפגע ב-`error_max_turns`. no-op מהיר כשריק. הרצה ידנית: `mcp-server/.venv/bin/python scripts/drain_metadata_queue.py [batch]`. | דרך `legal-metadata-drain.config.cjs` (pm2 cron) |
 | `legal-metadata-drain.config.cjs` | pm2/js | **תזמון כל 15 דק' של `drain_metadata_queue.py`** (cron `*/15 * * * *`, `METADATA_DRAIN_CRON` לעקיפה) — מונע סתימה של תור חילוץ-המטא ב-/precedents. דורש `GEMINI_API_KEY` ב-`~/.env`. התקנה: `pm2 start scripts/legal-metadata-drain.config.cjs && pm2 save`. | pm2 cron (host-side) |
 | `reconcile_metadata_status.py` | python | **נרמול `metadata_extraction_status` תקוע (G1)** — שורות עם ברירת-המחדל `'pending'` שאינן בצנרת-Gemini נערמות כ-backlog-רפאים שהדריינר (סורק `*_requested_at IS NOT NULL`) לעולם לא מנקה ומנפח את מונה "ממתין" ב-/operations. מיישב כל שורה למצב-אמת במקור: `internal_committee`→`completed` (מטא דטרמיניסטי, מחוץ ל-Gemini), `external_upload` מלא→`completed`, `external_upload` עם טקסט וחסר שם/תקציר→חותם `requested_at` (הדריינר יטפל), `cited_only` (אין טקסט)→`skipped`. **מכסה את שני התורים (#140):** אותו `cited_only→skipped` מוחל גם על `halacha_extraction_status` (תור-תאום, G2). אידמפוטנטי. תיקון-המקור הנלווה ב-`db.create_internal_committee_decision` + מסנן `EXTRACTION_ELIGIBLE_PREDICATE` ב-`list_pending_extraction_requests`. הרצה: `mcp-server/.venv/bin/python scripts/reconcile_metadata_status.py`. | חד-פעמי / re-runnable כהגנת-drift |
+| `reconcile_under_extracted_halacha.py` | python | **#144 — שחזור פסיקה תת-מחולצת** שהושלמה אך עם 0 הלכות למרות ≥3 מקטעי-נימוק (legal_analysis/ruling/conclusion) — חתימת ה-checkpoint-הריק שנוצרה לפני תיקון limit-notice ב-claude_session. מאפס checkpoints + `request_halacha_extraction` (נתיב קנוני, G2) → הדריינר מחלץ מחדש. שמרני (≥3 מקטעים → לא מטפל ב-remand לגיטימי חסר-הלכה; אפס אובדן כי 0 הלכות ממילא). מחריג cited_only. אידמפוטנטי, dry-run כברירת-מחדל / `--apply`. הרצה: `HOME=/home/chaim mcp-server/.venv/bin/python scripts/reconcile_under_extracted_halacha.py --apply`. | חד-פעמי / re-runnable |
 | `backfill_plans_registry.py` | python | **ייבוא מרשם-התכניות (V38) מקורפוס-ההחלטות** — סורק `data/cases/*/drafts/decision.md` + `data/training/cmp/*.md`, מאתר פסקאות-תוקף ("פורסמה למתן תוקף"), מחלץ רשומת-תכנית מובנית (`plans_extractor`, claude CLI מקומי) ועושה `upsert_plan(review_status='pending_review')` עם provenance. ה-SSOT לזהות+תוקף של תכנית, פעם-אחת במקום גזירה-מחדש מהשומות בכל תיק (G2). idempotent על plan_number מנורמל (G1/G3). `--dry-run` (ברירת-מחדל, כלום לא נכתב) / `--apply` / `--glob` (תת-קבוצה). אחרי הרצה: אישור-יו"ר ב-`plan_review`/תור-האישור (G10). הרץ: `mcp-server/.venv/bin/python scripts/backfill_plans_registry.py`. | ידני (חד-פעמי + לפי-צורך כשנוספות החלטות) |
 | `backfill_precedent_citations.py` | python | **#145** — backfill ל-`citation_formatted` (מראה-מקום) ברשומות `case_law` ריקות, באמצעות `db.format_precedent_citation` הדטרמיניסטי (X1 §3 / INV-ID2 — שדה-תצוגה נגזר, לא מעוצב ע"י LLM ש-הפיל אותו, #145). שני מעברים לכל שורה: (1) **ללא-LLM** — הרכבה מהשדות השמורים (ממלא שורות-ועדה עם parties+docket+date); (2) **LLM** — אם (1) נמנע ויש full_text, מריץ את מחלץ-המטא (extract_and_apply) שמחלץ רכיבים (parties, citation_prefix) ואז מרכיב — זה ממלא את 171 פסקי-בתי-המשפט מהכותרת. שורות בלי רובריקה (אין צדדים) נשארות ריקות ומדווחות, לא מנוחשות (INV-AH). idempotent — רק שדה ריק (G3). `--apply` / `--limit N` / `--no-llm`. הרץ: `HOME=/home/chaim mcp-server/.venv/bin/python scripts/backfill_precedent_citations.py`. | ידני (חד-פעמי + לפי-צורך) |
 | `auto-sync-cases.sh` | bash | סנכרון תיקי ערר ל-Gitea — רץ כל דקה | `* * * * *` (cron) |

scripts/halacha_drain_supervisor.py

@@ -527,6 +527,16 @@ def tick():
     if not staging_ok:
         notes.append("⚠️ תיקים הושלמו אך checkpoints לא התקדמו — לבדוק staging.")
 
+    # Churn guard (#144): a precedent regressing completed→pending (Δdone<0) or a
+    # net loss of halachot beyond consolidation noise (Δhal<-2) means a re-extract
+    # is DEGRADING the corpus (the 3→1 incident). Surface it loudly — never silent.
+    churn_ok = not (d_done < 0 or d_hal < -2)
+    if not churn_ok:
+        notes.append(
+            f"⚠️ churn: Δהלכות={d_hal} Δתיקים={d_done} — חילוץ-מחדש מדרדר "
+            "(completed→pending / אובדן-הלכות); לבדוק לפני המשך."
+        )
+
     if mode in ("ratelimited", "weekly_exhausted") and cd_dt:
         next_wake = max(300, min(int((cd_dt - now).total_seconds()) + 180, 3600))
     elif mode in ("done", "idle_off_window"):
@@ -562,7 +572,8 @@ def tick():
         "halachot_total": hal, "checkpointed_chunks": ck,
         "delta_halachot": d_hal, "delta_checkpoints": d_ck, "delta_done": d_done,
         "pm2": status, "progress_age_sec": int(progress_age),
-        "staging_ok": staging_ok, "rate_limited": in_cooldown, "cooldown_until": cooldown_until,
+        "staging_ok": staging_ok, "churn_ok": churn_ok,
+        "rate_limited": in_cooldown, "cooldown_until": cooldown_until,
     }, ensure_ascii=False))
 
 

scripts/reconcile_under_extracted_halacha.py

@@ -0,0 +1,93 @@
+"""Recover halacha precedents that completed extraction but are under-extracted (#144).
+
+Before the claude_session limit-notice fix, a chunk that hit a usage/rate limit
+could exit 0 with a refusal NOTICE as its "result" → the extractor read it as an
+empty answer and checkpointed the chunk as done-with-0-halachot. A ``completed``
+precedent could thus carry substantial reasoning yet ZERO holdings, and a resume
+would skip the empty-checkpointed chunks forever.
+
+This finds the strongly-degraded survivors — ``status='completed'``,
+``source_kind <> 'cited_only'``, with >= MIN_REASONING_CHUNKS reasoning chunks
+(legal_analysis / ruling / conclusion) yet 0 halachot — and re-queues them for a
+clean re-extraction: clear their chunk checkpoints, set status 'pending', stamp
+``halacha_extraction_requested_at`` so the nightly drain re-runs them.
+
+CONSERVATIVE by design: requires >= 3 reasoning chunks so a genuine remand (a
+1-2 chunk ruling that just sends the case back, legitimately holding-free) is
+NOT churned into a pointless re-extraction loop. Nothing is lost (these have 0
+halachot already). Chair-approved/published rows are untouched (there are none
+on a 0-halachot precedent, but the drain's reset preserves them regardless).
+
+Idempotent / re-runnable. Dry-run by default; pass ``--apply`` to write.
+Host-only (reads POSTGRES_URL from ~/.env). Run:
+    HOME=/home/chaim mcp-server/.venv/bin/python scripts/reconcile_under_extracted_halacha.py [--apply]
+"""
+
+from __future__ import annotations
+
+import asyncio
+import os
+import sys
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", "mcp-server", "src"))
+
+from legal_mcp.services import db
+
+MIN_REASONING_CHUNKS = 3
+_REASONING = ("legal_analysis", "ruling", "conclusion")
+
+_CANDIDATES_SQL = """
+SELECT cl.id, cl.case_number,
+       (SELECT count(*) FROM precedent_chunks pc
+          WHERE pc.case_law_id = cl.id
+            AND pc.section_type = ANY($1::text[])) AS reasoning_chunks
+FROM case_law cl
+WHERE cl.halacha_extraction_status = 'completed'
+  AND cl.source_kind <> 'cited_only'
+  AND (SELECT count(*) FROM halachot h WHERE h.case_law_id = cl.id) = 0
+  AND (SELECT count(*) FROM precedent_chunks pc
+         WHERE pc.case_law_id = cl.id
+           AND pc.section_type = ANY($1::text[])) >= $2
+ORDER BY reasoning_chunks DESC
+"""
+
+
+async def main(apply: bool) -> int:
+    pool = await db.get_pool()
+    rows = await pool.fetch(_CANDIDATES_SQL, list(_REASONING), MIN_REASONING_CHUNKS)
+
+    if not rows:
+        print("no under-extracted completed precedents found — nothing to do")
+        return 0
+
+    print(f"under-extracted candidates (completed, 0 halachot, >= "
+          f"{MIN_REASONING_CHUNKS} reasoning chunks): {len(rows)}")
+    for r in rows:
+        print(f"  {r['case_number']:<22} reasoning_chunks={r['reasoning_chunks']}  id={r['id']}")
+
+    if not apply:
+        print("\n(dry-run — pass --apply to re-queue these for clean re-extraction)")
+        return 0
+
+    requeued = 0
+    for r in rows:
+        cid = r["id"]
+        async with pool.acquire() as conn:
+            async with conn.transaction():
+                # Clear checkpoints so the resume re-extracts every chunk (these
+                # carry 0 halachot, so there is nothing to preserve/lose).
+                await conn.execute(
+                    "UPDATE precedent_chunks SET halacha_extracted_at = NULL "
+                    "WHERE case_law_id = $1", cid,
+                )
+        # Status→pending + stamp requested_at via the canonical request path so
+        # the drain picks it up (G2: same enqueue path as everything else).
+        await db.request_halacha_extraction(cid)
+        requeued += 1
+
+    print(f"\n✓ re-queued {requeued} precedent(s) — the halacha drain will re-extract them")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(asyncio.run(main("--apply" in sys.argv)))