`rerank.maybe_rerank` calls `base_search(limit=…, **base_kwargs)` on both
the rerank-on and rerank-off paths. Commit 242f668 moved the closure into
hybrid_search.py and renamed its parameter to `limit_inner`, so every call
to `/api/precedent-library/search` raised TypeError 500 regardless of the
VOYAGE_RERANK_ENABLED flag. Sibling `search_documents_hybrid` was unaffected
because it uses `lambda **kw:` which absorbs the kwarg.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Observed 2026-05-03: a `precedent_process_pending(halacha)` run that
chained two precedents (1110/20 → 317/10) succeeded for the first
(9 halachot, 129 chunks) and produced status=`no_halachot` for the
second despite it being a 47KB Supreme Court ruling with rich legal
analysis. A manual single-precedent re-run on 317/10 immediately
extracted 53 halachot. Diagnosis: every chunk's claude_session call
in the back-to-back run silently failed (likely Anthropic rate-limit
storm after the 1110/20 token burn), and the empty list was reported
as "Claude looked and found nothing" — same code path as a real
0-halacha ruling. The user couldn't tell the difference.
Three changes:
1. Surface chunk-level failures (halacha_extractor.py)
`_extract_chunk` now returns `(halachot, succeeded)` so the caller
can count how many chunks crashed. `extract()` uses this to
distinguish:
- `no_halachot` — chunks ran cleanly, Claude found nothing
- `extraction_failed` — ≥50% of chunks crashed AND zero halachot
came back (rate limit, subprocess crash, etc.)
When `extraction_failed`, DB status is left as 'processing' so the
request stays in the queue for the caller to retry — instead of
the old behaviour where it got marked 'completed' and silently
dropped from the queue.
2. Inter-precedent cooldown (precedent_library.py)
`process_pending_extractions` now sleeps 30s between precedents.
Anthropic rate-limits per-org, and back-to-back large rulings
(~4M tokens for 1110/20, immediately followed by another 2-3M)
was the empirical trigger. 30s gives the per-minute counter time
to drain.
3. Auto-retry on extraction_failed (precedent_library.py)
When a precedent comes back as `extraction_failed`, retry once
after a 60s cooldown before giving up. Rate-limit storms are
transient — the manual re-run of 317/10 minutes later succeeded
with 53 halachot and zero chunk failures, confirming a single
retry is sufficient. Only retries `extraction_failed`; never
`no_halachot` (Claude looked and there genuinely is no holding).
The DB status now ends up as 'failed' only after retries are
exhausted, matching the UI's terminal-failure chip.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
When Paperclip wakes the CEO and the model issues an mcp__legal-ai__*
call within ~10s of session init, Claude Code sometimes returns
"No such tool available" because the legal-ai MCP server hasn't
finished bringing up its tool catalog yet. Observed twice today on
CMPA precedent-extraction wakeups (sessions 9989fbaf and a9c61801);
the agent fell back to bash + .venv/bin/python and finished the work,
but the race needed fixing on the server side.
Three changes that close the window:
1. Lazy schema init (services/db.py + server.py)
`init_schema()` was awaited inside the FastMCP lifespan, blocking
the `initialize`/`tools/list` handshake until ~10 CREATE TABLE IF
NOT EXISTS statements ran. Under contention (two CEOs waking at
once for different companies) this stretched. Now the lifespan
returns immediately and `get_pool()` runs the schema migrations
exactly once on first DB access, guarded by an asyncio.Lock.
tools/list is answered in milliseconds regardless of DB state.
2. Lazy heavy imports
- services/embeddings.py: voyageai (~450ms) loaded only inside
_get_client()
- services/extractor.py: google.cloud.vision (~550ms) loaded only
inside _get_vision_client() and _ocr_with_google_vision()
These two were being imported at module top from
legal_mcp.tools.documents -> services.processor -> services.{
extractor,embeddings}, so the FastMCP server couldn't even start
responding until both finished. Cold start dropped from 2.7s to
1.17s end-to-end (init + tools/list response).
3. Agent-side warmup + retry guidance (.claude/agents/legal-ceo.md)
Even with a fast server, the model can still race on the very
first call. The precedent-extraction section now tells the CEO
to call workflow_status as a warmup probe and to retry after a
short sleep if it sees "No such tool available", before falling
back to the python bypass.
Also expanded the precedent-tool whitelists on the sub-agents that
delegate halacha/library work (commits 4a9a6b7 + 7ee90dc added the
tools to the MCP server but only the CEO got them in its allowed
list). Added to: legal-researcher (full extraction set), legal-analyst
(library_get/list + halacha review), legal-writer (library lookups +
halacha_review), legal-qa (library_get + halacha_review), and the two
that the CEO was already missing (halacha_review, halachot_pending).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The legacy chunker did not track which PDF page each chunk came from.
Stored chunks had page_number=NULL, which blocked the multimodal
hybrid retriever's text+image boost — it joins (chunk, image) on
(document_id, page_number) and the join could never fire.
This change:
- extractor.extract_text now returns (text, page_count, page_offsets);
page_offsets[i] is the start char offset of page (i+1) in the joined
text. None for non-PDFs.
- chunker.chunk_document accepts an optional page_offsets and tags
each chunk with the page that contains its first character (uses
the existing chunker logic; pages assigned post-hoc by content
search to keep the diff minimal).
- processor.process_document and precedent_library.ingest_precedent
forward page_offsets through the chunker. New uploads now carry
accurate page_number on every chunk.
- Other extract_text callers (tools/documents, tools/workflow,
web/app.py) updated to unpack the third element (ignored).
- scripts/backfill_chunk_pages.py: per-case retrofit. Re-extracts each
PDF (re-OCRs via Google Vision if needed, ~$0.0015/page), computes
page_offsets, and updates page_number on every chunk by content
search. Idempotent; --force re-runs on already-tagged docs.
Forward-only would leave the 419 image embeddings backfilled on
cases 8174-24 + 8137-24 unable to boost their corresponding text
chunks. The retrofit script closes that gap (cost ~$0.60).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Cosine scores in voyage-3 (~0.4-0.5) and voyage-multimodal-3
(~0.2-0.25) live on different scales. The previous weighted-sum
merge let text always dominate — verified empirically: 0 image-only
hits across 7 queries on case 8174-24, image side contributed nothing.
RRF combines by *rank* in each list rather than raw score, robust
to scale differences. Per-item score:
rrf_score = text_weight / (k + text_rank)
+ image_weight / (k + image_rank)
A row that appears in both lists (joined on (id_field, page_number))
gets both terms — surfaced as match_type='text+image'.
After fix on 8174-24 (146 image rows): 2 image-only hits land in
top-5 across all 7 test queries, surfacing actual table/diagram/
signature pages (p12, p13 of שומת המשיבה for 'טבלת השוואת ערכי שומה',
p25 of שומת השגה for 'תרשים גוש וחלקה', etc).
On 8137-24 (273 image rows): 'חישוב היוון של דמי החכירה' goes from
0 baseline results → 5 hybrid results (3 text + 2 image), opening
recall on scanned content the OCR layer misses.
Default MULTIMODAL_TEXT_WEIGHT 0.65 → 0.5 (vanilla RRF) since the
prior 0.65 was tuned for raw cosine scales that no longer apply.
New env knob MULTIMODAL_RRF_K (default 60, standard literature).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Stage C: per-page image embeddings via voyage-multimodal-3 + hybrid
text+image search. Off by default; enable with MULTIMODAL_ENABLED=true.
- Schema V9: document_image_embeddings + precedent_image_embeddings
(vector(1024), page_number, image_thumbnail_path)
- extractor.render_pages_for_multimodal renders PDF pages at
MULTIMODAL_DPI (144) for embedding + JPEG thumbnails at
MULTIMODAL_THUMB_DPI (96) for UI preview, in one pass
- embeddings.embed_images calls voyage-multimodal-3 in 50-page batches
- services/hybrid_search.py orchestrator: rerank applied to text side
first (rerank-2 is text-only); image side cosine; weighted merge
with text_weight 0.65 (env-tunable); image-only pages surface as
match_type='image' so dense scanned content still appears
- processor.process_document and precedent_library.ingest_precedent
gated by flag — non-fatal on multimodal failure
- scripts/multimodal_backfill.py — idempotent per-case CLI to embed
existing documents without re-extracting text
Validated locally on a 5-page response brief: render 0.31s, embed 8.32s,
hybrid merge surfaces image rows correctly. Production rollout starts
with flag=false (no behavior change), then per-case A/B.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
PATCH /api/halachot/{id} was returning 500 because the row included
``embedding`` as a numpy.ndarray of np.float32, which FastAPI's
jsonable_encoder cannot serialize (vars() and dict() both fail on it).
The bug had been latent — it triggered for the first time today after
the auto-approve batch left only low-confidence halachot for the chair
to review manually, and her first PATCH hit the unserializable response.
Replace ``RETURNING *`` with an explicit column list (everything except
``embedding``). Callers that need the embedding can re-fetch via
``get_halacha`` — but no current caller does.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Halachot extracted by halacha_extractor with confidence >= 0.80 are now
inserted with review_status='approved' instead of 'pending_review' —
they appear in search_precedent_library immediately. Halachot below the
threshold still require manual chair approval.
Threshold tunable via env (HALACHA_AUTO_APPROVE_THRESHOLD), defaults to
0.80. Rationale: 89% of historical extractions (356/400) score 0.80+,
spot-checks confirmed quality, and the manual review backlog was the
single biggest reason rerank-2 was returning passages-only on
ההבחנה-style queries.
After this change + the one-time backfill UPDATE, search now returns
9/10 halachot for "ההבחנה בין השבחה לפיצויים" instead of 0 — and the
top-3 are exact-match rules, not adjacent passages.
Reviewer field records "auto-approved (confidence ≥ X.XX)" with the
threshold value at insert time, for traceability.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
set_case_law_extraction_status and set_case_law_halacha_status now NULL
the corresponding *_requested_at timestamp when status transitions to
"completed" or "failed". Without this, completed rows kept lingering in
the local-MCP work queue (which scans by `WHERE *_requested_at IS NOT NULL`)
and the UI's isPrecedentActive check, leaving them undeletable until a
manual SQL cleanup.
The pre-existing process_pending_extractions path already called
clear_extraction_request, but other paths (re-extraction, status set
during upload) didn't — so the cleanup belongs at the status setter.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The chair pointed out three UX gaps after uploading a new precedent:
1. The status said "מחלץ הלכות" but nothing was actually running — the
field only meant "halacha_extraction_status != completed", which
includes the post-upload "pending" state where the local MCP worker
hasn't been told to drain anything yet. Misleading.
2. The page didn't refresh on its own. The chair had to F5 to see new
counts after extraction completed.
3. Clicking the trash icon mid-extraction would cascade-delete the row
while the extractor was still using it (FK errors, partial writes).
Fixes:
- ingest_precedent now auto-queues both metadata and halacha extraction
on upload by stamping the request timestamps. The chair (or me) drains
the queue with one `precedent_process_pending` call from chat —
no need to click any button before that.
- StatusPill is now five-state with proper labels:
"נכשל" (extraction_status=failed) — red
"מעבד טקסט" — shimmer (extraction_status=processing)
"בתור" — neutral (chunks queued, not yet running)
"מחלץ הלכות" — shimmer (halacha_extraction_status=processing)
"ממתין לחילוץ" — neutral (queued for local MCP worker)
"לא חולץ" — neutral (pending without queue stamp — shouldn't happen)
"X/Y מאושרות" — gold (done, with halachot count)
The shimmer is a CSS-only sliding-stripe animation defined in globals.
- usePrecedents has a conditional refetchInterval — polls every 5s while
any row is mid-extraction or queued, then stops once everything settles
to completed/failed. New helper isPrecedentActive() centralises the
"is this row mid-something" check so the UI and the destructive-action
guard agree.
- Trash button is disabled (opacity 30%, tooltip explains) while the row
is active. Pencil/edit stays enabled — editing metadata fields during
extraction is safe (last write wins, low-stakes race).
Schema: list_external_case_law now exposes the two *_requested_at
timestamps so the UI can distinguish "queued" from "never asked".
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The chair wanted a one-click "extract metadata" button on the edit sheet.
The constraint stays the same — claude_session needs the local CLI which
the container doesn't have, so the button can't run the extractor itself.
Compromise: button stamps a queue marker; the local MCP server drains the
queue on demand.
DB (V8): two nullable timestamps on case_law,
metadata_extraction_requested_at and halacha_extraction_requested_at,
with partial indexes for cheap "find pending" scans.
API:
POST /api/precedent-library/{id}/request-metadata → stamp the row
POST /api/precedent-library/{id}/request-halachot → same for halacha
GET /api/precedent-library/queue/pending?kind=... → read-only view
UI: Sparkles button in the edit sheet header. Click → toast tells the
chair what to run from Claude Code. The button never triggers the
extractor directly from the container.
MCP tool: precedent_process_pending(kind, limit) — runs from Claude Code
with the local CLI, picks up everything stamped, calls the extractor for
each, clears the timestamp on success. Failures keep the timestamp so the
next invocation retries them.
Architectural rule (claude_session local-only) is preserved end-to-end
and called out in the new endpoint comment + tool docstring.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two follow-ups after running the metadata extractor on 403-17:
1. Library table: shadcn TableCell defaults to whitespace-nowrap and
the table wrapper has overflow-x-auto, so the long citation forced
a horizontal scrollbar inside the row. Override on the citation
cell only — whitespace-normal + break-words + min/max-w to keep the
column readable. Same for the case-name cell. Row aligns to top so
wrapping doesn't push neighbours up.
2. Extractor now also fills source_type (court_ruling /
appeals_committee). The previous round added decision_date_iso,
precedent_level, and court but left source_type empty. Same
closed-enum + merge-only-if-empty policy.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The first end-to-end run on 403-17 surfaced three fields the auto-fill
left blank because the chair didn't set them in the upload form: date,
precedent_level, and court. All three are right there in the ruling's
header text — there's no reason to require manual entry.
Prompt now asks for:
- decision_date_iso (YYYY-MM-DD parsed from "ניתנה היום, … 5 בספטמבר 2022"
style signatures)
- precedent_level (closed enum: עליון/מנהלי/ועדת_ערר_ארצית/ועדת_ערר_מחוזית)
- court (the full court name from the title block)
Validation is unchanged: precedent_level only accepts the four enum
values; decision_date_iso is parsed into a Python date object before
being handed to update_case_law (asyncpg doesn't coerce strings to
DATE columns); court is stored verbatim.
Merge policy is unchanged — only fills empty fields. Anything the
chair typed in the upload form survives.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Architectural correction: every claude_session caller in this project
runs through the local MCP server (~/.claude.json points at
/home/chaim/legal-ai/mcp-server/.venv/bin/python). The Coolify container
has no `claude` CLI and no claude.ai session, so any LLM call originating
from web/ FastAPI fails with "Claude CLI not found" — which is exactly
what we hit on 403-17.
The earlier Anthropic SDK fallback would have made it work, but at
direct API cost. The chair's preference is to stay on the claude.ai
session for everything. So:
- claude_session.py: removed the SDK fallback, restored CLI-only.
The error message now points the next person at the architectural
rule in the module docstring instead of papering over it.
- precedent_library.py:ingest_precedent (called from FastAPI on upload)
now does only the non-LLM half: extract → chunk → embed → store.
Sets halacha_extraction_status='pending' for the chair to act on.
- reextract_halachot / reextract_metadata kept, but lazy-import their
extractors so the FastAPI path can't accidentally pull them in. They
are reachable only via the MCP tools precedent_extract_halachot /
precedent_extract_metadata, which run locally with CLI.
- Removed POST /api/precedent-library/{id}/extract-halachot and
/extract-metadata — they were dead ends from the container.
- Dropped the `anthropic` Python dep that the SDK fallback required.
- UI: removed the "refresh halachot" and "sparkles metadata" buttons
that called those endpoints. Edit sheet now points the chair at the
MCP tool names instead.
Halacha and metadata extraction for an uploaded precedent now happen
when the chair (via Claude Code) runs:
mcp__legal-ai__precedent_extract_metadata <case_law_id>
mcp__legal-ai__precedent_extract_halachot <case_law_id>
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
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>
Three improvements to the precedent library based on usage feedback:
1. Auto-fill metadata at upload time. New service
precedent_metadata_extractor reads the ruling's full_text and
suggests case_name (short), summary, headnote, key_quote,
subject_tags, appeal_subtype. The merge policy fills only empty
fields, preserving everything the chair typed in the upload form.
Wired into the ingest pipeline; also exposed as a re-run endpoint
POST /api/precedent-library/{id}/extract-metadata for existing
records.
2. Edit sheet in the UI. Pencil icon on each library row opens a
pre-populated form covering every field. A Sparkles button on the
sheet runs the metadata extractor on demand and refreshes the
form. The case_number is read-only because halachot are FK'd to
it; renaming requires delete + re-upload.
3. Halacha extractor branches on is_binding. Sources marked binding
(Supreme/Administrative) keep the strict halacha prompt. Non-binding
sources (other appeals committees, district courts on planning
matters) get a different prompt that extracts applications,
interpretive principles, and persuasive conclusions — labeled with
new rule_types 'application' and 'persuasive'. The fallback also
widens chunk selection: if the chunker labeled nothing as
legal_analysis/ruling/conclusion, we now run on all chunks rather
than returning zero halachot for a usable ruling.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds a third corpus of legal authority distinct from style_corpus
(Daphna's prior decisions for voice) and case_precedents (chair-attached
quotes per case). The new corpus holds chair-uploaded court rulings and
other appeals committee decisions, with binding rules (הלכות) extracted
automatically and queued for chair approval.
Pipeline (web/app.py + services/precedent_library.py):
file → extract → chunk → Voyage embed → halacha_extractor → store +
publish progress over the existing Redis SSE channel.
Schema V7 (services/db.py): extends case_law with source_kind +
extraction status fields under a CHECK constraint pinning practice_area
to the three appeals committee domains (rishuy_uvniya, betterment_levy,
compensation_197). New precedent_chunks (vector(1024)) and halachot
tables (vector(1024) over rule_statement, IVFFlat indexes, gin on
practice_areas/subject_tags). Halachot start as pending_review; only
approved/published rows are visible to search_precedent_library.
Agents: legal-writer, legal-researcher, legal-analyst, legal-ceo,
legal-qa get search_precedent_library. legal-writer prompt explains
the three-corpus distinction and CREAC use; legal-qa now verifies that
every cited halacha resolves to an approved row in the corpus.
UI: /precedents page with four tabs — library / semantic search /
pending review (J/K nav, A/R/E shortcuts, badge count) / stats.
Reuses the existing upload-sheet progress + SSE pattern.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The case repo is the user's backup, so anything in the dir must end up
on Gitea. Two layers:
1. Periodic sweep (every 30s) — git_sync.sweep_loop runs as a FastAPI
background task. It scans every case dir, runs git status --porcelain
on each, and commit_and_push's any dirty changes with an auto-built
Hebrew message ("אוטו: טיוטות (2) · מסמכים"). Catches files written
outside the API path: agent research artefacts, manual edits, etc.
2. Explicit commits at known write paths — DOCX export, interim draft,
apply_user_edit, revise_draft, mark-final, analysis DOCX export.
These give immediate feedback with descriptive messages instead of
waiting up to 30s for the sweep.
safe.directory injection added to _git_env so sweep + explicit commits
work even when the running uid differs from the case-dir owner (host
runs vs. uniform-root container).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The auto-creation in case_create had two failure modes that combined to
make repos silently missing: a stale GITEA_TOKEN returning 401, and the
outer try/except in case_create that swallowed every exception with a
bare pass. Result: cases like 8174-24 ended up with a local git repo and
Paperclip project but no Gitea repo, with no signal anywhere.
_setup_gitea_remote now returns {ok, url, error} and never raises; the
result is attached to the case JSON and the FastAPI endpoint logs a
warning when ok=false. The UI gets a "צור ריפו ב-Gitea" button on the
case header that appears only when the repo or remote is missing.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The case_delete tool in tools/cases.py and the DELETE /api/cases endpoint
in web/app.py both invoke await db.delete_case(case_id), but no such
function existed in services/db.py — every call returned 500 with an
AttributeError. Discovered while wiping case 8174-24 for a clean rerun.
Implementation is straightforward because the FK graph already does the
work: 7 dependent tables CASCADE on cases.id (documents, document_chunks,
claims, appraiser_facts, decisions, qa_results, case_precedents) and 2
SET NULL (audit_log, chair_feedback). A single DELETE FROM cases is
enough — no manual ordering needed.
Documented in the docstring that this only touches the legal-ai DB —
Paperclip projects/issues and Gitea repos for the case are separate
systems and must be cleaned up by the caller.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The claude_session bridge had two structural defects that made any
non-trivial document extraction unreliable:
1. subprocess.run() blocks the asyncio event loop in the MCP server
for the full duration of every LLM call (60-180s typical).
2. The 120-second timeout was below the cold-cache cost of any
document over ~12K Hebrew characters. Three back-to-back timeouts
on case 8174-24 dropped 43 appellant claims on the floor.
Phase 1 of the remediation plan — keeps claude_session as the engine
(no Anthropic API switch) and restructures around it:
claude_session.py
• query / query_json are now async — asyncio.create_subprocess_exec
instead of subprocess.run, so MCP server can serve other coroutines
while a call is in flight.
• DEFAULT_TIMEOUT 120 → 1800 (30 min). High enough that no realistic
document hits it; bounded so a runaway never zombifies forever.
• LONG_TIMEOUT 300 → 3600 for opus block writing on full case context.
• TimeoutError now actually kills the subprocess (asyncio.wait_for
cancellation alone leaves the child running).
claims_extractor.py
• _split_by_sections: chunks at numbered sections / Hebrew letter
headings / "פרק" markers / markdown ##, falls back to paragraph
breaks, then to hard splits. Targets 12K chars per chunk — small
enough that each chunk reliably finishes inside the timeout.
• _extract_chunk: per-chunk retry (1 attempt by default) with
structured logging on failure. Failed chunks no longer crash the
overall extraction; they're skipped with a partial-result warning.
• extract_claims_with_ai now runs chunks in parallel via
asyncio.gather bounded by a semaphore (CHUNK_CONCURRENCY=3).
For a 25K-char appeal: was sequential 150-300s, now ~70-90s.
Updated all 9 callers (claims, appraiser facts, block writer, qa
validator, brainstorm, learning loop, style analyzer × 3) to await
the now-async API.
The one-shot scripts/extract_claims_8174.py used to recover 43
appellant claims on case 8174-24 has been moved to .archive/ — phase 1
makes it obsolete. SCRIPTS.md updated.
Phase 2 (background-task wrapper around LLM-bound MCP tools, persistent
llm_tasks table, SSE progress) is the structural follow-up — separate PR.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Hebrew was rendering LTR or in Times New Roman fallback in some Word
contexts. Root cause: incomplete RTL marking and missing font hints
on the run level.
Three layers of RTL are required (per skills/docx/SKILL.md):
1. Section: <w:bidi/> in sectPr (now inherited from template)
2. Paragraph: <w:bidi/> directly in pPr (paragraph direction)
3. Run: <w:rtl/> in rPr — tells Word to use cs (complex-script) font
Without an explicit font on the run, Hebrew renders in the ascii slot
(Times New Roman). Force David on all four slots (ascii / hAnsi / cs /
eastAsia) so every shaping path picks the correct font.
Changes:
- TEMPLATE_PATH now points to skills/docx/decision_template.docx
(carries David, RTL, margins, styles); replaces hard-coded constants.
- _mark_run_rtl: writes rFonts on all four slots, then appends <w:rtl/>.
- _mark_paragraph_rtl: places <w:bidi/> directly in pPr (not nested in
rPr — that was the bug), and adds <w:rtl/> to the paragraph-mark rPr.
- _set_paragraph_jc: forces explicit jc, overriding style-inherited.
Tests:
- test_mark_paragraph_rtl_adds_bidi_directly_in_pPr — guards against
the regression where bidi was nested inside rPr.
- test_mark_run_rtl_forces_david_on_all_font_slots — ensures all four
font slots are set, not just cs.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Cases failed to push silently after the Gitea token in Infisical was
rotated: the embedded credential in each case repo's origin URL was
the old token, the rotation never propagated, and capture_output=True
hid the auth failure as a logger.warning. Three cases (1033-25,
1130-25, 1194-25) accumulated unpushed commits over weeks before
this was noticed.
Fixes the root cause in two places: web/gitea_client.py for uploads
through the FastAPI endpoint, and mcp-server/services/git_sync.py
for case_update / document_upload through MCP tools (which previously
committed but never pushed at all).
The new commit_and_push helper:
- re-injects the current GITEA_ACCESS_TOKEN into the existing origin
URL on every call, so pushes survive token rotation
- logs push failures at WARNING with the actual stderr (the previous
code suppressed errors entirely)
- continues to push even when the commit was a no-op, in case earlier
commits are still unpushed
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds a comprehensive archive flow for closed cases — separate /archive
screen in the UI, archive/restore actions on the case detail page, and
automatic two-way sync with Paperclip.
Backend (web/app.py + mcp-server/services/db.py):
- New SCHEMA_V6 migration: cases.archived_at TIMESTAMPTZ + partial index
- list_cases gains include_archived/archived_only flags; default excludes
archived rows so the main /api/cases list hides closed cases
- archive_case / restore_case helpers in db.py
- POST /api/cases/{n}/archive sets archived_at and calls
pc_archive_project (sets Paperclip projects.archived_at via direct DB)
- POST /api/cases/{n}/restore clears archived_at and calls
pc_restore_project (clears Paperclip archived_at)
- archive_project / restore_project in paperclip_client.py — name-based
match consistent with create_project's lookup
Frontend (web-ui):
- cases.ts: scope param ("active"|"archived"|"all") on useCases;
useArchiveCase / useRestoreCase mutations
- /archive page (new): table of archived cases with restore button +
search, sort, empty state matching the editorial aesthetic of /
- case-archive-action.tsx: button on case detail header. Active case →
confirm dialog → archive. Archived case → restore (no confirm).
Toast announces both legal-ai and Paperclip outcomes (synced, not
found in pc, error)
- case-header shows "בארכיון" badge when archived_at is set
- Nav: ארכיון link added to AppShell after בית
Tested end-to-end against the live DB:
- 1130-25 archive → list_cases(include_archived=False) excludes it,
list_cases(archived_only=True) includes it, restore reverses
- pc archive/restore on 1194-25 verified via direct DB lookup
- TypeScript compiles clean
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The "על כן" pattern for block-yod-bet was too greedy and matched mid-discussion
transitional sentences (e.g. "על כן, במקום בו..."), which caused forward-scan
to skip block-yod-alef ("סוף דבר") via the pointer advance.
Tightened to require an operative subject (אנו / הערר / הוועדה / ועדת הערר)
so terminal "על כן, אנו מחליטים" still matches but mid-block transitions don't.
Added structural_fallback for cover blocks (alef/bet/gimel/dalet) — these are
template metadata not present in user-edited DOCX bodies. Inject zero-content
anchors so apply_user_edit can still target them later. The frontend toast
distinguishes real content gaps from fallback anchors.
Also expanded heading patterns based on training corpus inspection:
- block-vav: על המקרקעין חלות / במצב התכנוני / התכניות החלות
- block-zayin: טענות העוררת
- block-chet: עיקר תגובת המשיב
- block-tet: הדיון בוועדת הערר
For case 1130-25, this raises detection from 6/12 to 11/12 blocks — only
block-yod-bet remains missing (Daphna's edit ends at "סוף דבר" + numbered
ruling, no terminal "ההחלטה" or "על כן אנו מחליטים" paragraph).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Until now changing a document's doc_type required a manual SQL update.
Adds an inline editor on the document badge so the chair can retag
without leaving the case page, and threads an appraiser_side tag
(committee / appellant / deciding) through the appraisal pipeline so
betterment-levy cases — which usually have 2-3 appraisers — render
conflicts with the deciding appraiser's view marked as governing.
Backend
- New appraiser_facts.appraiser_side column (V5.1) populated from
documents.metadata.appraiser_side at extraction time.
- extract_appraiser_facts now returns status='sides_missing' with the
list of untagged appraisals instead of running with empty side
labels — chair must tag every appraisal first via the UI.
- Conflict detection orders entries committee → appellant → deciding so
the deciding appraiser appears last; block-tet's prompt instructs the
writer to phrase the deciding appraiser's view as the governing
factual finding ("ואולם, השמאי המכריע קבע...").
- New PATCH /api/cases/{n}/documents/{doc_id} (Pydantic model with
whitelist validation) and matching document_update MCP tool. Both
merge appraiser_side into metadata JSONB instead of touching the
schema.
UI
- New shared doc-types module exports the canonical 11 doc_type
options plus the 3 appraiser-side options; both upload-sheet and
the document badge now read from it instead of duplicating Hebrew
labels.
- New DocumentTypeEditor renders a Popover off the doc-type Badge
with two Selects. The save button stays disabled while doc_type is
appraisal but no side has been picked, mirroring the backend
enforcement so the user finds out before triggering extraction.
- usePatchDocument React-Query mutation invalidates the case detail
on success so the badge updates without a manual refresh.
Lets the chair generate a partial decision DOCX before the discussion-and-
ruling block is decided. Same template, skill and DOCX styling as the final
decision (David, RTL, bookmarks) — only the block selection and order differ:
רקע (ו) → תכניות+היתרים (ט) → טענות (ז) → הליכים (ח). The opening (ה),
ruling (י), summary (יא), and signatures (יב) are omitted.
- New appraiser_facts table + CRUD + conflict detection in db.py (V5 schema).
Conflict = same plan/permit identifier reported differently by 2+ appraisers.
- New appraiser_facts_extractor service: per-appraisal Claude extraction of
plans + permits with raw quotes and page numbers.
- block-tet prompt extended with a permits sub-section sourced from the
extracted facts, plus an explicit instruction to flag inter-appraiser
conflicts in neutral wording without resolving them (deferred to block-yod).
- block-chet prompt extended with a post-hearing materials context sourced
from documents.metadata.is_post_hearing.
- docx_exporter.export_decision now accepts mode='interim' which reorders
the blocks per the chair's mental model and writes
טיוטת-ביניים-v{N}.docx (versioned independently of regular drafts).
- 3 new MCP tools: extract_appraiser_facts, write_interim_draft,
export_interim_draft. write_interim_draft auto-runs extraction if the
appraiser_facts table is empty for the case.
Fixes critical bug in 1033-25: user-uploaded עריכה-*.docx files were
orphaned on disk while exports kept rebuilding from stale DB blocks.
New architecture:
- User-uploaded DOCX becomes the source of truth (cases.active_draft_path)
- System edits via XML surgery with real Word <w:ins>/<w:del> revisions
- User can Accept/Reject each change from within Word
Components:
- docx_reviser.py: XML surgery for Track Changes (15 tests)
- docx_retrofit.py: retroactive bookmark injection with Hebrew marker
detection + heading heuristic (9 tests)
- docx_exporter.py: emits bookmarks around each of the 12 blocks
- 3 new MCP tools: apply_user_edit, list_bookmarks, revise_draft
- 4 new/updated endpoints: upload (auto-registers active draft),
/exports/revise, /exports/bookmarks, /exports/{filename}/retrofit,
/active-draft
- DB migration: cases.active_draft_path column
- UI: correct banner using real v-numbers, "מקור האמת" badge,
detailed upload toast with bookmarks_added/missing_blocks
- agents: legal-exporter (3 export modes), legal-ceo (stage G for
revision handling), legal-writer (revision mode)
Multi-tenancy:
- Works for both CMP (1xxx cases) and CMPA (8xxx/9xxx cases)
- New revise-draft skill added to both companies
- deploy-track-changes.sh syncs skills CMP ↔ CMPA
- retrofit_case.py: one-off retrofit of existing files
Tests: 34 passing (15 reviser + 9 retrofit + 4 exporter bookmarks + 6 e2e)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- legal-analyst: opus 4.6 → opus 4.7
- legal-proofreader: opus 4.6 → opus 4.7
- legal-writer: sonnet 4.6 → opus 4.7 (complex block writing benefits from stronger model)
- block_writer MODEL_MAP: updated opus ID to 4.7
Opus 4.7 brings: high-res images (2576px), better file-based memory,
improved DOCX generation, and task budgets for agentic loops.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Add strip_nevo_preamble() to extractor.py — auto-removes Nevo database
headers (bibliography, legislation, mini-ratio) during training upload
- Add appeal_subtype column to style_patterns table — patterns are now
stored per subtype instead of globally mixed
- Update clear_style_patterns() to support subtype-scoped deletion
- Pass appeal_subtype through analyze_corpus → store → upsert pipeline
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Support ingestion of betterment levy (היטל השבחה) decisions into a
separate training corpus (CMPA). Key changes:
- Add .doc file extraction via LibreOffice conversion in extractor
- Add practice_area/appeal_subtype columns to style_corpus table
- Route training files to cmp/ or cmpa/ subdirs based on appeal subtype
- Fix derive_subtype to handle ARAR-YY-NNNN format (was matching year digit)
- Expose practice_area/appeal_subtype params in MCP upload_training tool
- Add appeal_subtype filter to analyze_style for per-type style analysis
- Update betterment levy methodology in lessons.py: checklist (from generic
to corpus-based), opening/closing strategies, and discussion rules
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
UUID and datetime objects from PostgreSQL RETURNING * were not
serializable. All other tool files already used default=str.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
CEO agent was reverting case status from "processing" to "new" when
updating metadata fields. Added ordered status list — case_update now
silently ignores status changes that would move backwards.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- auto-sync-cases.sh: fix broken directory scan (was looking for
status subdirs that don't exist), fix env var word-splitting bug,
add safe.directory handling and error logging
- cases.py: auto-create Gitea repo on case_create, fix
documents/original → documents/originals naming mismatch
- app.py: add GET /api/cases/{case_number}/git-status endpoint
- web-ui: add SyncIndicator component in case header showing
sync status (synced/pending/no remote) with last commit time
- pyproject.toml: add httpx dependency
- CLAUDE.md: update Paperclip wakeup API docs
- settings page: switch tag input from Select to free-text with datalist
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Install git in Docker image and wrap all subprocess git calls in
try/except so a missing or failing git binary never kills an upload
that already succeeded.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When a case is created, a Paperclip project is now automatically created in
the correct company based on the appeal_subtype tag. Tag-to-company mappings
are managed via a new Settings page that pulls companies from Paperclip DB.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The merge of ui-rewrite removed these parameters from db.search_similar()
and db.create_case() but left the callers passing them, causing TypeError
on any corpus search. Restores the parameters and adds schema migration.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Four functions were called by tools/precedents.py but never implemented
in services/db.py: create_case_precedent, list_case_precedents,
delete_case_precedent, search_precedent_library. This caused 500 errors
on the /api/cases/{n}/precedents endpoint.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Documents tab was limited to ~9 visible items due to fixed max-height
without overflow-hidden. Now uses 70vh with proper overflow. Added
click-to-preview (shows extracted text in dialog) and delete button
with confirmation dialog + backend DELETE endpoint.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The Next.js app was proxying /api/* to the old Flask/FastAPI server
at legal-ai.nautilus.marcusgroup.org. When that server went down,
the Next.js app's API calls failed with 503.
Now both services run in the same container:
- FastAPI (uvicorn) on :8000 — the API backend
- Next.js (node) on :3000 — proxies /api/* to localhost:8000
Changes:
- Dockerfile: multi-stage build with Python 3.12 + Node.js
- next.config.ts: default proxy target is now 127.0.0.1:8000
- start.sh: launches uvicorn in background + node in foreground
- pyproject.toml: add fastapi + uvicorn as explicit deps
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Major changes from ui-rewrite branch:
- Decision-writing methodology (decision-methodology.md) based on FJC, Garner, Posner
- 5 source books downloaded and processed (341K words)
- Methodology integrated into block-yod prompt
- All 8 Paperclip agents updated for methodology compliance
- DB schema V4: claim handling, standard of review, precedent hierarchy
- 15 pipeline gaps identified and fixed after test run on case 1130-25
- Negative checks layer added to CEO and QA agents
- HEARTBEAT: wakeup CEO on completion + blocked status
- Flexible claim handling (bundle/skip via chair_directions)
Conflicts resolved: all 5 files use ui-rewrite version (the latest).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
New columns for methodology-aware decision pipeline:
claims table:
- claim_handling (address/bundle/skip) — per-claim handling mode
- bundle_group — group name for bundled claims
- handling_reason — explanation for skip/bundle
cases table:
- standard_of_review — review standard (independent discretion / etc.)
- subject_categories — JSONB array of topics in the appeal
case_law table:
- precedent_level — hierarchy (supreme/administrative/national/district)
- is_binding — binding holding vs. obiter dictum
- creac_role — how it serves reasoning (rule/explanation/analogy)
decisions table:
- issue_order — JSONB array of ordered issues with type
- claim_handling — JSONB overrides from chair_directions
Migration tested and applied successfully on production DB.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
"בית ספר להחלטות" Phase 2 — the system now has formal analytical
methodology for building quasi-judicial decisions, separate from
Dafna's writing style (SKILL.md) and content checklists.
What was done:
- Downloaded 5 authoritative sources (~341K words): FJC Judicial
Writing Manual (1991+2020), Garner Legal Writing in Plain English,
Posner How Judges Think, Scalia/Garner Making Your Case
- Extracted principles from all sources into intermediate docs
- Synthesized into docs/decision-methodology.md (3,400 words,
12 sections, 10 guiding principles)
- Integrated methodology into block-yod prompt via {methodology_guidance}
- Restructured legal-writer agent workflow to follow analytical stages
- Made "answer all claims" flexible (bundle/skip via chair_directions)
- Added methodology compliance check (#7) to legal-qa agent
- Updated all knowledge files (CLAUDE.md, SKILL.md, lessons, corpus)
Three-layer architecture:
1. Methodology (decision-methodology.md) — universal, how to think
2. Content checklists (lessons.py) — specific per appeal subtype
3. Style (SKILL.md) — Dafna's personal writing patterns
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Backend changes cherry-picked from ui-rewrite branch to enable
feedback API endpoints for the Next.js staging UI.
- chair_feedback DB table + API endpoints (GET/POST/PATCH)
- Content checklists by appeal subtype injected into block-yod prompt
- MCP tools for recording and listing chair feedback
- Corpus analysis documentation (24 decisions)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Addresses Dafna's observation that licensing decisions lack comprehensive
planning discussion. Systematic corpus analysis of all 24 training decisions
revealed the system learned writing style but not substantive content.
Changes:
- Corpus analysis of all 24 decisions (docs/corpus-analysis.md)
- 5 content checklists by appeal subtype injected into block-yod prompt
- chair_feedback DB table + API endpoints + MCP tools
- Feedback management page in Next.js UI (/feedback)
- Navigation updated with "הערות יו״ר" link
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
New self-contained table + MCP tools + FastAPI endpoints for letting
the chair attach external case-law quotes (quote + citation מראה מקום,
optional chair note, optional archived PDF) to either a specific
threshold_claim / issue or the case as a whole.
Data model
- case_precedents (SCHEMA_V5_SQL) — case_id, section_id NULL/
"threshold_N"/"issue_N", quote, citation (free-text), chair_note,
pdf_document_id FK to documents, denormalized practice_area for
cross-case library filtering.
- Deliberately NOT linked to the existing case_law table — that one
has UNIQUE(case_number) which would force parsing the free-text
citation into a structured key. A backfill pass into case_law is
a later follow-up once the UI stabilizes.
- db.py gains 4 helpers: create_case_precedent, list_case_precedents,
delete_case_precedent, search_precedent_library. The last uses
DISTINCT ON (citation) for the cross-case typeahead so each
precedent appears once even if reused across many cases.
MCP tools (legal_mcp/tools/precedents.py)
- precedent_attach, precedent_list, precedent_remove,
precedent_search_library — registered in server.py.
FastAPI (web/app.py)
- POST /api/cases/{n}/precedents — create, with PrecedentCreateRequest
- POST /api/cases/{n}/precedents/upload-pdf — one-shot PDF upload to
a dedicated documents/precedents/ subdirectory, creates a
documents row with doc_type="precedent_archive" and no text
extraction (archive only)
- GET /api/cases/{n}/precedents — list
- DELETE /api/precedents/{id} — uses path param since precedent_id
is a UUID (slash-safe, unlike case numbers)
- GET /api/precedents/search?q=...&practice_area=... — library
typeahead
Block-writer integration into _build_precedents_context is a deferred
follow-up — Phase 1 surfaces the feature in the compose UI only.
Plan: ~/.claude/plans/woolly-cooking-graham.md
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Wires a new case-deletion path across the three layers that needed it:
- db.delete_case(case_id) — single SQL DELETE; documents, chunks, and
qa_results cascade via existing schema FKs, audit_log nullifies.
- cases_tools.case_delete(case_number, remove_files=False) — MCP tool
wrapper. File tree on disk is kept by default (audit trail); pass
remove_files=True for a hard delete.
- DELETE /api/cases?case_number=... — FastAPI endpoint taking the case
number as a QUERY param rather than a path segment. Case numbers
like "1000/0426" can't be passed through a path parameter because
FastAPI routing decodes %2F before matching, so a query param is
the only shape that works for historical data.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Adds two orthogonal columns — practice_area (top-level legal domain:
appeals_committee / national_insurance / labor_law) and appeal_subtype
(building_permit / betterment_levy / compensation_197) — denormalized
into cases, documents, document_chunks, decisions, and style_corpus so
vector searches can filter without JOINs.
Why: the system handles two unrelated sub-domains under the same
appeals committee (1xxx building permits and 8xxx/9xxx betterment/197),
with different rules and writing style. Without a separation axis,
search_similar() and the block-writer's precedent lookup were free to
surface betterment-levy paragraphs while drafting a building-permit
decision — a real risk of cross-domain contamination. The same axis
also lets future domains (national insurance, labor law) coexist
without separate schemas.
Schema (V4 migration in db.py):
- ALTER ... ADD COLUMN IF NOT EXISTS on all five tables + composite
indexes (practice_area first).
- Idempotent backfill: case_number ~ '^1' → building_permit, '^8' →
betterment_levy, '^9' → compensation_197; propagated to documents,
chunks, and decisions via case_id; training-corpus rows (case_id NULL)
default to appeals_committee.
Code:
- New services/practice_area.py with derive_subtype, validate, and
is_override + enum constants.
- db.create_case / create_document / store_chunks / create_decision
inherit practice_area from the parent case (or take an explicit
override for the case_id=None training corpus).
- db.search_similar and search_similar_paragraphs accept practice_area
+ appeal_subtype filters using the denormalized columns.
- tools/search.py auto-resolves the filter from case_number when given.
- block_writer._build_precedents_context now passes the active case's
practice_area to search_similar_paragraphs — closes the contamination
hole for the discussion-block precedent fetch.
- tools/cases.case_create auto-derives subtype from case_number; an
explicit override that disagrees writes a case_subtype_override entry
to audit_log so we can spot bad classifications later.
- tools/documents.document_upload_training tags new training material
with practice_area + subtype end-to-end (corpus, document, chunks).
UI (web/static/index.html + web/app.py):
- New-case wizard gets a practice_area dropdown (others disabled until
national_insurance / labor_law arrive) and an appeal_subtype dropdown
with JS auto-fill from the case-number prefix; manual edits stick.
- Case header shows a blue badge with practice_area · subtype.
- CaseCreateRequest plumbs both fields through to cases_tools.case_create.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
