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

@@ -13,6 +13,8 @@ claude_session המקומי בלבד (כמו שאר המחלצים) — לא Ant
 from __future__ import annotations
 
 import logging
+import os
+import re
 from uuid import UUID
 
 from legal_mcp.services import claude_session, db
@@ -23,6 +25,18 @@ logger = logging.getLogger(__name__)
 # not a pinned model id — the session model is whatever is configured).
 MODEL_TAG = "claude_local"
 
+# ── mavat auto-enrichment (Phase C trigger 2) ──────────────────────────────────
+# When an extracted candidate is missing its validity (gazette_date / yalkut), we
+# fill the gaps from the official source (mavat) via the host bridge. Conservative
+# by design: only modern numeric plan numbers resolve on mavat search, each fetch
+# drives a real browser (~30-60s, serial), so we gate by format + cap per call and
+# fail soft. Set PLAN_ENRICH_FROM_MAVAT=0 to disable.
+_ENRICH_ENABLED = os.environ.get("PLAN_ENRICH_FROM_MAVAT", "1").strip() not in ("0", "false", "")
+_ENRICH_MAX_PER_CALL = int(os.environ.get("PLAN_ENRICH_MAX_PER_CALL", "8"))
+# mavat search resolves the modern "NN-NNNNNNN" identifiers; legacy forms
+# (מי/820, 5166/ב, תמ"א 38) don't, so don't waste a browser launch on them.
+_MAVAT_NUM_RE = re.compile(r"^\d{2,4}-\d{6,8}$")
+
 
 EXTRACT_PLANS_PROMPT = """אתה מחלץ מידע עובדתי על תכניות בניין-עיר (תב"ע) עבור מרשם-תכניות של ועדת ערר.
 
@@ -112,16 +126,69 @@ async def extract_plans_from_text(text: str) -> list[dict]:
     return candidates
 
 
+def _needs_enrichment(c: dict) -> bool:
+    """A candidate is worth enriching iff its validity is incomplete AND its
+    number is a mavat-resolvable modern identifier."""
+    if not (_ENRICH_ENABLED and _MAVAT_NUM_RE.match((c.get("plan_number") or "").strip())):
+        return False
+    return not (c.get("gazette_date") and c.get("yalkut_number"))
+
+
+async def _enrich_from_mavat(c: dict) -> tuple[dict, bool]:
+    """Fill a candidate's MISSING fields from mavat (never override case-grounded
+    values). Returns (candidate, enriched?). Fails soft — a bridge-down / not-found
+    / blocked fetch leaves the candidate untouched (logged, never swallowed)."""
+    from legal_mcp.services import plans_fetch
+
+    num = c["plan_number"].strip()
+    try:
+        fetched = await plans_fetch.fetch_plan(num)
+    except plans_fetch.PlanFetchUnavailable as e:
+        logger.info("plan-enrich: bridge unavailable for %s — %s", num, e)
+        return c, False
+    except plans_fetch.PlanFetchError as e:
+        logger.info("plan-enrich: mavat had no usable result for %s — %s", num, e)
+        return c, False
+    except Exception as e:  # noqa: BLE001 — never let enrichment break extraction
+        logger.warning("plan-enrich: unexpected error for %s — %s", num, e)
+        return c, False
+
+    enriched = dict(c)
+    filled: list[str] = []
+    for f in ("gazette_date", "yalkut_number", "display_name", "plan_type", "purpose"):
+        if not enriched.get(f) and fetched.get(f):
+            enriched[f] = fetched[f]
+            filled.append(f)
+    if filled:
+        logger.info("plan-enrich: %s filled %s from mavat (%s)",
+                    num, ",".join(filled), fetched.get("source_url", ""))
+        return enriched, True
+    return c, False
+
+
 async def upsert_candidates(
     candidates: list[dict],
     *,
     source_case_number: str = "",
     source_document_id: UUID | None = None,
     model_used: str = MODEL_TAG,
+    enrich: bool = True,
 ) -> list[dict]:
-    """Upsert extracted candidates into the registry as pending_review (G10)."""
+    """Upsert extracted candidates into the registry as pending_review (G10).
+
+    When ``enrich`` (default) and a candidate's validity is incomplete, its
+    missing fields are pulled from mavat first (capped per call). The row still
+    enters pending_review — enrichment changes the candidate, not the chair gate.
+    """
     out: list[dict] = []
+    enriched_count = 0
     for c in candidates:
+        used = model_used
+        if enrich and enriched_count < _ENRICH_MAX_PER_CALL and _needs_enrichment(c):
+            c, did = await _enrich_from_mavat(c)
+            if did:
+                enriched_count += 1
+                used = f"{model_used}+mavat"
         try:
             plan = await db.upsert_plan(
                 plan_number=c["plan_number"],
@@ -133,12 +200,18 @@ async def upsert_candidates(
                 review_status="pending_review",
                 source_case_number=source_case_number,
                 source_document_id=source_document_id,
-                model_used=model_used,
+                model_used=used,
             )
             out.append(plan)
         except ValueError as e:
             # Don't swallow — surface the bad candidate so it isn't silently dropped.
             logger.warning("upsert_candidates: skipped %r — %s", c.get("plan_number"), e)
+    if enrich and enriched_count >= _ENRICH_MAX_PER_CALL:
+        logger.warning(
+            "plan-enrich: hit the per-call cap (%d) — remaining candidates kept "
+            "as-extracted (no silent truncation; raise PLAN_ENRICH_MAX_PER_CALL).",
+            _ENRICH_MAX_PER_CALL,
+        )
     return out