fix(precedents): Anthropic SDK fallback, format() crash, UI refresh

Three fixes to the precedent library after the first end-to-end test on
403-17 surfaced runtime issues:

1. Anthropic SDK fallback in claude_session. The legal-ai Docker container
   does not ship the `claude` CLI, so every halacha and metadata extraction
   was failing with "Claude CLI not found." Module now tries the CLI first
   (zero-cost local path) and falls back to the Anthropic SDK with
   ANTHROPIC_API_KEY when the binary is absent. Default model is
   claude-sonnet-4-6, overridable via CLAUDE_SDK_MODEL env. The system
   message gets cache_control: ephemeral so multi-chunk runs reuse the
   cached instruction prefix at ~10% read cost. Adds `anthropic` to
   pyproject deps.

2. precedent_metadata_extractor crashed with KeyError because the JSON
   example inside the prompt template contained literal { } characters
   that str.format() interpreted as placeholders. Switched to f-string
   concatenation; the prompt template no longer needs format() at all.

3. Library list query stays stale after upload because the upload
   mutation's onSuccess fires when the POST returns task_id, not when
   SSE reports completion. Added a second invalidate inside the SSE
   watcher in PrecedentUploadSheet so the new row appears with up-to-date
   chunk and halachot counts the moment processing finishes.

Halacha and metadata extractors now route the long static prompt through
the new `system=` parameter so the SDK path actually caches it; the CLI
path concatenates and behaves as before.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

mcp-server/pyproject.toml

@@ -8,6 +8,7 @@ dependencies = [
     "asyncpg>=0.29.0",
     "pgvector>=0.3.0",
     "voyageai>=0.3.0",
+    "anthropic>=0.45.0",
     "python-dotenv>=1.0.0",
     "pydantic>=2.0.0",
     "pymupdf>=1.25.0",

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

@@ -1,21 +1,26 @@
-"""Claude Code session bridge — runs prompts via `claude -p` instead of API.
+"""Claude Code session bridge — runs prompts via `claude -p` or Anthropic SDK.
 
-All LLM calls in the project go through this module. We shell out to the
+History: originally shelled out to `claude -p` exclusively (zero direct API
-local Claude Code CLI which uses the developer's claude.ai session — zero
+cost via the developer's claude.ai session). That works locally but fails
-direct Anthropic API cost.
+in the legal-ai Docker container, which does not ship the CLI. To keep the
+same call sites working in production, the module now tries the CLI first
+and falls back to the Anthropic SDK using ``ANTHROPIC_API_KEY`` when the
+CLI binary is absent.
 
-History: this module was originally synchronous (``subprocess.run``) with
+Both paths share a single shape: ``query()`` returns text, ``query_json()``
-a 120-second timeout. That broke for large legal documents:
+parses that text as JSON. Callers don't need to know which path executed.
 
-  1. Sync subprocess stalled the asyncio event loop in the MCP server
+Async history: originally synchronous (``subprocess.run``) with a 120 s
-     while a single LLM call was in flight.
+timeout. That broke for large legal documents — sync subprocess stalled
-  2. 120 seconds was far too short. A 25K-character Hebrew appeal on cold
+the asyncio loop, and 120 s was far too short for cold-cache Hebrew prompts
-     prompt cache routinely takes 130-180 seconds; we proved this in case
+(case 8174-24 hit three timeouts in a row). Fixed by going async with a
-     8174-24 (three timeouts in a row).
+30-minute ceiling.
 
-The fix: switch to async subprocess (non-blocking) and raise the default
+Caching contract (SDK path): pass long, repeated instruction text via
-ceiling to 30 minutes — long enough that no realistic document hits it,
+``system=...``. The SDK path adds ``cache_control: ephemeral`` so back-to-back
-but bounded so a runaway never zombifies forever.
+chunk calls reuse the cached prefix at ~10% of read cost. The CLI path doesn't
+expose API-level caching; with ``system`` set, we just prepend it to the
+prompt — same observable behavior, no caching benefit.
 """
 
 from __future__ import annotations
@@ -23,54 +28,82 @@ from __future__ import annotations
 import asyncio
 import json
 import logging
+import os
+import shutil
 
 from legal_mcp.config import parse_llm_json
 
 logger = logging.getLogger(__name__)
 
-# Default ceiling for any single ``claude -p`` invocation, in seconds.
+# Default ceiling for any single LLM call, in seconds.
 # 30 min covers any single-document call we make in practice (chunking
 # handles the rest); the bound exists only to prevent runaway zombies.
 DEFAULT_TIMEOUT = 1800
 LONG_TIMEOUT = 3600  # opus block writing on full case context
 
+# Anthropic SDK fallback config — used when `claude` CLI is not on PATH.
+# Default to Sonnet 4.6: strong balance of Hebrew legal-text quality and
+# cost for the per-chunk extraction workload. Override via env if needed.
+DEFAULT_SDK_MODEL = os.environ.get("CLAUDE_SDK_MODEL", "claude-sonnet-4-6")
+DEFAULT_SDK_MAX_TOKENS = int(os.environ.get("CLAUDE_SDK_MAX_TOKENS", "8192"))
 
-async def query(prompt: str, timeout: int = DEFAULT_TIMEOUT, max_turns: int = 1) -> str:
+_anthropic_client = None
-    """Send a prompt to Claude Code headless and return the text response.
 
-    Passes the prompt via stdin (not argv) to avoid the OS ARG_MAX limit —
-    prompts can be 500K+ chars when analyzing a full style corpus.
 
-    Args:
+def _has_cli() -> bool:
-        prompt: The prompt to send.
+    return shutil.which("claude") is not None
-        timeout: Max seconds before the subprocess is killed.
-        max_turns: Max conversation turns (1 = single response).
 
-    Returns:
-        The text response from Claude.
 
-    Raises:
+def _get_anthropic_client():
-        RuntimeError: If claude CLI is not available, fails, or times out.
+    """Lazy-init the AsyncAnthropic client. Raises with a clear message if
+    the package or API key is missing — better than letting the SDK 401 in
+    the middle of a multi-chunk extraction.
     """
+    global _anthropic_client
+    if _anthropic_client is not None:
+        return _anthropic_client
+    try:
+        import anthropic
+    except ImportError as e:
+        raise RuntimeError(
+            "The 'anthropic' package is required when the Claude CLI is "
+            "unavailable. Add it to mcp-server/pyproject.toml."
+        ) from e
+    if not os.environ.get("ANTHROPIC_API_KEY"):
+        raise RuntimeError(
+            "ANTHROPIC_API_KEY is not set; cannot fall back to Anthropic SDK."
+        )
+    _anthropic_client = anthropic.AsyncAnthropic()
+    return _anthropic_client
+
+
+async def _query_cli(
+    prompt: str, system: str | None, timeout: int, max_turns: int,
+) -> str:
+    """Run the prompt via the local `claude` CLI subprocess.
+
+    Uses the developer's claude.ai session — zero direct API cost. With
+    ``system`` set, we just prepend it to the prompt; the CLI doesn't
+    expose API-level caching anyway.
+    """
+    full_prompt = f"{system}\n\n{prompt}" if system else prompt
+
     cmd = [
         "claude", "-p",
         "--output-format", "json",
         "--max-turns", str(max_turns),
     ]
 
-    try:
+    proc = await asyncio.create_subprocess_exec(
-        proc = await asyncio.create_subprocess_exec(
+        *cmd,
-            *cmd,
+        stdin=asyncio.subprocess.PIPE,
-            stdin=asyncio.subprocess.PIPE,
+        stdout=asyncio.subprocess.PIPE,
-            stdout=asyncio.subprocess.PIPE,
+        stderr=asyncio.subprocess.PIPE,
-            stderr=asyncio.subprocess.PIPE,
+    )
-        )
-    except FileNotFoundError:
-        raise RuntimeError("Claude CLI not found. Install Claude Code or add 'claude' to PATH.")
 
     try:
         stdout_b, stderr_b = await asyncio.wait_for(
-            proc.communicate(input=prompt.encode("utf-8")),
+            proc.communicate(input=full_prompt.encode("utf-8")),
             timeout=timeout,
         )
     except asyncio.TimeoutError:
@@ -100,10 +133,104 @@ async def query(prompt: str, timeout: int = DEFAULT_TIMEOUT, max_turns: int = 1)
         return stdout
 
 
-async def query_json(prompt: str, timeout: int = DEFAULT_TIMEOUT) -> dict | list | None:
+async def _query_sdk(prompt: str, system: str | None, timeout: int) -> str:
+    """Run the prompt via the Anthropic SDK with 5-min ephemeral caching
+    on the system message. Streams the response to dodge HTTP read timeouts
+    on long Hebrew JSON outputs.
+    """
+    import anthropic
+
+    client = _get_anthropic_client()
+
+    kwargs: dict = {
+        "model": DEFAULT_SDK_MODEL,
+        "max_tokens": DEFAULT_SDK_MAX_TOKENS,
+        "messages": [{"role": "user", "content": prompt}],
+    }
+    if system:
+        # cache_control: ephemeral → 5-min TTL. The same system text repeats
+        # across every chunk in an extraction run, so we get
+        # cache_read_input_tokens on every call after the first.
+        kwargs["system"] = [{
+            "type": "text",
+            "text": system,
+            "cache_control": {"type": "ephemeral"},
+        }]
+
+    try:
+        async with client.messages.stream(**kwargs) as stream:
+            message = await asyncio.wait_for(
+                stream.get_final_message(),
+                timeout=timeout,
+            )
+    except asyncio.TimeoutError:
+        raise RuntimeError(f"Anthropic SDK call timed out after {timeout}s")
+    except anthropic.APIError as e:
+        raise RuntimeError(f"Anthropic SDK call failed: {e}") from e
+
+    text_parts: list[str] = []
+    for block in message.content:
+        if getattr(block, "type", None) == "text":
+            text_parts.append(block.text)
+    out = "".join(text_parts).strip()
+    if not out:
+        raise RuntimeError("Anthropic SDK returned no text content")
+
+    usage = getattr(message, "usage", None)
+    if usage is not None:
+        logger.debug(
+            "claude_session SDK usage: input=%s cache_read=%s cache_write=%s output=%s",
+            getattr(usage, "input_tokens", None),
+            getattr(usage, "cache_read_input_tokens", None),
+            getattr(usage, "cache_creation_input_tokens", None),
+            getattr(usage, "output_tokens", None),
+        )
+    return out
+
+
+async def query(
+    prompt: str,
+    timeout: int = DEFAULT_TIMEOUT,
+    max_turns: int = 1,
+    *,
+    system: str | None = None,
+) -> str:
+    """Send a prompt to Claude and return the text response.
+
+    Tries the Claude CLI first (zero API cost, uses claude.ai session).
+    Falls back to the Anthropic SDK with ANTHROPIC_API_KEY when the CLI is
+    absent — this is the production-Docker path.
+
+    Pass repeating instruction text via ``system=`` so the SDK path can
+    cache it (5-min ephemeral). Pass only the per-chunk content via
+    ``prompt``. The CLI path concatenates them; the SDK path keeps them
+    separate so caching works.
+
+    Args:
+        prompt: The user-facing prompt text.
+        timeout: Max seconds before the call is aborted.
+        max_turns: CLI-only — max conversation turns (1 = single response).
+        system: Optional system message. With the SDK path, gets cached
+            with 5-min ephemeral TTL when set.
+
+    Raises:
+        RuntimeError: if both paths fail or time out. The message includes
+            which path raised so the caller can distinguish CLI from SDK.
+    """
+    if _has_cli():
+        return await _query_cli(prompt, system, timeout, max_turns)
+    return await _query_sdk(prompt, system, timeout)
+
+
+async def query_json(
+    prompt: str,
+    timeout: int = DEFAULT_TIMEOUT,
+    *,
+    system: str | None = None,
+) -> dict | list | None:
     """Send a prompt and parse the response as JSON.
 
     Uses parse_llm_json for robust parsing (handles markdown wrapping, truncation).
     """
-    raw = await query(prompt, timeout=timeout)
+    raw = await query(prompt, timeout=timeout, system=system)
     return parse_llm_json(raw)

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

@@ -279,8 +279,9 @@ async def _extract_chunk(
         else HALACHA_EXTRACTION_PROMPT_PERSUASIVE
     )
     chunk_label = f" (חלק {chunk_index + 1}/{chunk_total})" if chunk_total > 1 else ""
-    prompt = (
+    # Pass the static instruction prompt as `system` so the SDK path can cache
-        f"{base_prompt}\n\n"
+    # it (5-min ephemeral). Only the per-chunk content varies via `prompt`.
+    user_msg = (
         f"## הקלט\n"
         f"סוג קטע: {section_type}\n"
         f"{context}{chunk_label}\n\n"
@@ -289,7 +290,7 @@ async def _extract_chunk(
     last_err: Exception | None = None
     for attempt in range(CHUNK_RETRY_ATTEMPTS + 1):
         try:
-            result = await claude_session.query_json(prompt)
+            result = await claude_session.query_json(user_msg, system=base_prompt)
         except Exception as e:
             last_err = e
             logger.warning(

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

@@ -28,6 +28,10 @@ _HEAD_CHARS = 12_000
 _TAIL_CHARS = 6_000
 
 
+# Note: this template is concatenated with f-strings at call-time rather
+# than using .format(), because the JSON example below contains '{' / '}'
+# which str.format would interpret as placeholders and crash with
+# KeyError on the field names.
 METADATA_EXTRACTION_PROMPT = """אתה מסייע משפטי בכיר. קרא את פסק הדין/ההחלטה הבא וחלץ ממנו מטא-דאטה לקטלוג הקורפוס.
 
 המטרה: למלא שדות בטופס העלאה שהמשתמש הזין באופן חלקי. **אל תמציא** — אם המידע לא מופיע בטקסט, השאר ריק (מחרוזת ריקה / מערך ריק).
@@ -51,13 +55,6 @@ METADATA_EXTRACTION_PROMPT = """אתה מסייע משפטי בכיר. קרא א
 4. **headnote** — לא מצטטים, מסכמים. סגנון נבו: ביטוי קצר אחד.
 5. **key_quote** — חייב להיות הדבקה מילולית מהקלט. אם אין ציטוט בולט — השאר ריק.
 6. **subject_tags** — 3-7 תגיות בעברית, snake_case (חניה, קווי_בניין, שיקול_דעת, פגם_פרוצדורלי, סמכות, מועדים, פגיעה_במקרקעין, ירידת_ערך, תכנית_רחביה, מימוש_במכר, וכד'). שייך לתחום של ועדת ערר תכנון ובניה.
-
-## הקלט
-{context}
-
---- תחילת הטקסט ---
-{text_window}
---- סוף הטקסט ---
 """
 
 
@@ -104,12 +101,18 @@ async def extract_metadata(case_law_id: UUID | str) -> dict:
         f"תאריך: {date_str}\n"
         f"תחום: {practice_area}"
     )
-    prompt = METADATA_EXTRACTION_PROMPT.format(
+    text_window = _build_text_window(full_text)
-        context=context, text_window=_build_text_window(full_text),
+    # Static instructions go via `system` so the SDK path can cache them
+    # across uploads. Per-precedent content goes in the user prompt.
+    user_msg = (
+        f"## הקלט\n{context}\n\n"
+        f"--- תחילת הטקסט ---\n{text_window}\n--- סוף הטקסט ---"
     )
 
     try:
-        result = await claude_session.query_json(prompt)
+        result = await claude_session.query_json(
+            user_msg, system=METADATA_EXTRACTION_PROMPT,
+        )
     except Exception as e:
         logger.warning("precedent_metadata_extractor: query failed: %s", e)
         return {}

web-ui/src/components/precedents/precedent-upload-sheet.tsx

@@ -3,6 +3,7 @@
 import { useEffect, useState } from "react";
 import { Upload, Loader2, CheckCircle2, AlertCircle } from "lucide-react";
 import { toast } from "sonner";
+import { useQueryClient } from "@tanstack/react-query";
 import {
   Sheet, SheetContent, SheetHeader, SheetTitle, SheetDescription,
 } from "@/components/ui/sheet";
@@ -14,7 +15,10 @@ import {
   Select, SelectContent, SelectItem, SelectTrigger, SelectValue,
 } from "@/components/ui/select";
 import { Progress } from "@/components/ui/progress";
-import { useUploadPrecedent, type PracticeArea, type SourceType } from "@/lib/api/precedent-library";
+import {
+  useUploadPrecedent, libraryKeys,
+  type PracticeArea, type SourceType,
+} from "@/lib/api/precedent-library";
 import { useProgress } from "@/lib/api/documents";
 import {
   PRACTICE_AREAS, PRECEDENT_LEVELS, SOURCE_TYPES,
@@ -44,6 +48,7 @@ export function PrecedentUploadSheet({ open, onOpenChange }: Props) {
   const [taskId, setTaskId] = useState<string | null>(null);
   const upload = useUploadPrecedent();
   const progress = useProgress(taskId);
+  const qc = useQueryClient();
 
   // Reset form when the sheet closes — fields, file input, and any in-flight
   // task subscription. We accept the cascade-render warning because resetting
@@ -60,17 +65,23 @@ export function PrecedentUploadSheet({ open, onOpenChange }: Props) {
     setHeadnote(""); setIsBinding(true); setTaskId(null);
   }, [open]);
 
-  // Auto-close on completion
+  // Auto-close on completion + refresh library list/stats so the new
+  // row appears with up-to-date counts (halachot, approved). The mutation's
+  // onSuccess fires when POST returns the task_id; we need a second
+  // invalidation when SSE reports terminal status, otherwise the table
+  // shows stale data.
   useEffect(() => {
     if (progress?.status === "completed") {
+      qc.invalidateQueries({ queryKey: libraryKeys.all });
       toast.success("הפסיקה הוכנסה לקורפוס. ההלכות ממתינות לאישור.");
       const t = window.setTimeout(() => onOpenChange(false), 1200);
       return () => window.clearTimeout(t);
     }
     if (progress?.status === "failed") {
+      qc.invalidateQueries({ queryKey: libraryKeys.all });
       toast.error(`כשל בעיבוד: ${progress.error || "שגיאה לא ידועה"}`);
     }
-  }, [progress, onOpenChange]);
+  }, [progress, onOpenChange, qc]);
 
   const onSubmit = async (e: React.FormEvent) => {
     e.preventDefault();