legal-ai/scripts/final_learning_pipeline.py

#!/usr/bin/env python3
"""One-shot LOCAL pipeline for the 'run-learning' button (voice learning).

The container can't run the LLM steps (claude/DeepSeek/Gemini keys are local), so
the /api/cases/{case}/final/run-learning endpoint wakes the Hermes curator, which
runs THIS single deterministic command. Collapsing the flow into one script (rather
than asking the agent to assemble several tool calls) makes the autonomous path
reliable.

Steps:
  [1] ingest_final_version(case, file_path)  → Opus distils draft↔final into
      draft_final_pairs.analysis (status→analyzed). INV-LRN5 separates style↔substance.
  [2] enroll the final into style_corpus (idempotent) so lessons have a corpus_id.
  [3] style_lesson_panel --apply  → DeepSeek+Gemini vote per style lesson; 2/2-keep →
      decision_lesson (source=panel:deepseek+gemini); split → chair (INV-G10).

The fold into SKILL.md / legal-decision-lessons.md stays a manual chair gate.
Local-only. Idempotent — safe to re-run.

Durable (X16 / INV-DUR1): the 3 steps run through scripts/_pipeline_runtime.py
(shared with final_halacha) with a SQLite checkpoint per case
(data/checkpoints/learning.sqlite). A crash/OOM in the long style panel [3]
RESUMES from [3] instead of re-paying the Opus distillation [1]. Default =
auto-resume; ``--fresh`` forces a clean run. Needs the host extra
``pip install -e ".[durable]"``; without it the steps run linearly (as before).

    cd ~/legal-ai/mcp-server
    .venv/bin/python ../scripts/final_learning_pipeline.py --case 8126-03-25
"""
from __future__ import annotations

import argparse
import asyncio
import json
import sys
from argparse import Namespace
from pathlib import Path

# scripts/ is not a package — make style_lesson_panel + the runtime importable.
sys.path.insert(0, str(Path(__file__).resolve().parent))

import _pipeline_runtime  # noqa: E402 — durable runtime (X16); scripts/ on sys.path
from legal_mcp import config  # noqa: E402
from legal_mcp.services import db  # noqa: E402
from legal_mcp.tools.documents import document_upload_training  # noqa: E402
from legal_mcp.tools.workflow import ingest_final_version  # noqa: E402


def _resolve_final_path(case_number: str) -> str | None:
    """The canonical final saved by /final/upload, with a graceful fallback."""
    export_dir = config.find_case_dir(case_number) / "exports"
    canonical = export_dir / f"סופי-{case_number}.docx"
    if canonical.exists():
        return str(canonical)
    cands = sorted(export_dir.glob("סופי-*.docx"))
    return str(cands[0]) if cands else None


async def _has_style_corpus(decision_number: str) -> bool:
    pool = await db.get_pool()
    async with pool.acquire() as conn:
        row = await conn.fetchrow(
            "SELECT 1 FROM style_corpus WHERE decision_number = $1 LIMIT 1",
            decision_number,
        )
    return bool(row)


async def _latest_pair_status(case_id) -> str | None:
    pool = await db.get_pool()
    async with pool.acquire() as conn:
        return await conn.fetchval(
            "SELECT status FROM draft_final_pairs WHERE case_id = $1 "
            "ORDER BY created_at DESC LIMIT 1",
            case_id,
        )


async def main(args: argparse.Namespace) -> int:
    case_number = args.case
    case = await db.get_case_by_number(case_number)
    if not case:
        print(f"✗ תיק {case_number} לא נמצא")
        return 1

    final_path = _resolve_final_path(case_number)
    if not final_path:
        print(f"✗ לא נמצא קובץ סופי ל-{case_number} (העלה דרך 'העלאת החלטה סופית של היו\"ר')")
        return 1
    print(f"final: {final_path}\n")

    # The 3 steps as durable nodes (X16 / INV-DUR1) — shared runtime with
    # final_halacha (scripts/_pipeline_runtime.py). A crash/OOM in the long style
    # panel [3] resumes from [3] instead of re-paying the Opus distillation [1].

    async def step_ingest(results: dict) -> dict:
        # [1] distillation (Opus) — skip if already analyzed (idempotent; --force to redo).
        status = await _latest_pair_status(case["id"])
        if status == "analyzed" and not args.force:
            print("[1/3] ingest_final_version — דולג (הזוג כבר analyzed; --force לחידוש)")
            return {"ingest": "skipped:analyzed"}
        print("[1/3] ingest_final_version — דיסטילציית טיוטה↔סופי…", flush=True)
        raw = await ingest_final_version(case_number, file_path=final_path)
        try:
            env = json.loads(raw)
        except Exception:
            print(f"  (ingest returned: {raw[:200]})")
            return {"ingest": "unparsed"}
        if env.get("status") == "error":  # fatal — halt (resume retries)
            raise RuntimeError(f"ingest_final_version failed: {env.get('message')}")
        d = env.get("data", {})
        ds = d.get("diff_stats", {})
        print(f"  ✓ change {ds.get('change_percent')}% · lessons {d.get('lessons_count')} "
              f"· new_expr {d.get('new_expressions')}")
        return {"ingest": "done"}

    async def step_enroll(results: dict) -> dict:
        # [2] enroll into style_corpus (idempotent) — lessons need a corpus_id.
        print("[2/3] רישום לקורפוס-הסגנון (idempotent)…", flush=True)
        if await _has_style_corpus(case_number):
            print("  ✓ כבר רשום בקורפוס-הסגנון")
            return {"enroll": "exists"}
        r = await document_upload_training(
            final_path,
            decision_number=case_number,
            title=f"החלטה סופית — {case.get('proceeding_type', '')} {case_number}".strip(),
            practice_area=case.get("practice_area") or "appeals_committee",
            appeal_subtype=case.get("appeal_subtype") or "",
        )
        try:
            print(f"  ✓ corpus_id {json.loads(r).get('data', {}).get('corpus_id')}")
        except Exception:
            print(f"  (training upload returned: {r[:160]})")
        return {"enroll": "done"}

    async def step_panel(results: dict) -> dict:
        # [3] two-judge style panel (DeepSeek + Gemini) — the long step durability protects.
        apply = not args.dry_run
        print(f"[3/3] פאנל-סגנון דו-סוכני (DeepSeek+Gemini) {'--apply' if apply else '(dry-run)'}…",
              flush=True)
        import style_lesson_panel as slp
        rc = await slp.main(Namespace(
            case=case_number, pair_id=None, apply=apply, limit=0, concurrency=4,
        ))
        return {"panel_rc": rc or 0}

    steps = [
        _pipeline_runtime.Step("ingest_final_version", step_ingest),
        _pipeline_runtime.Step("enroll_style_corpus", step_enroll),
        _pipeline_runtime.Step("style_panel", step_panel),
    ]
    checkpoint_db = config.DATA_DIR / "checkpoints" / "learning.sqlite"
    thread_id = f"learning:{case_number}" + (":dryrun" if args.dry_run else "")
    try:
        results = await _pipeline_runtime.run_pipeline(
            steps,
            thread_id=thread_id,
            checkpoint_db=checkpoint_db,
            fresh=bool(args.fresh) or args.dry_run,
        )
    except Exception as e:  # fatal step (e.g. ingest error) — clean non-zero exit
        print(f"\n✗ pipeline-למידה נכשל: {e}")
        return 1
    print("\n✓ pipeline-למידה הושלם" + (" (dry-run)" if args.dry_run else ""))
    return int(results.get("panel_rc", 0) or 0)


if __name__ == "__main__":
    ap = argparse.ArgumentParser(description=__doc__,
                                 formatter_class=argparse.RawDescriptionHelpFormatter)
    ap.add_argument("--case", required=True, help="case_number, e.g. 8126-03-25")
    ap.add_argument("--dry-run", dest="dry_run", action="store_true",
                    help="run the chain but the style panel in dry-run (no decision_lesson writes)")
    ap.add_argument("--force", action="store_true",
                    help="re-run ingest_final_version even if the pair is already analyzed")
    ap.add_argument("--fresh", action="store_true",
                    help="ignore any incomplete checkpoint and run from step [1] "
                         "(default: auto-resume an interrupted run; X16/INV-DUR1)")
    raise SystemExit(asyncio.run(main(ap.parse_args())))