Merge pull request 'feat(drafting): טיוטת-ביניים דטרמיניסטית — Opus 4.8 + effort per-בלוק (#204)' (#356) from worktree-agent-a0d2502e8308bc27c into main
All checks were successful
Build & Deploy / build-and-deploy (push) Successful in 1m27s
G12 Leak-Guard / leak-guard (push) Successful in 4s
Lint — undefined names / undefined-names (push) Successful in 10s

This commit was merged in pull request #356.
This commit is contained in:
2026-06-30 12:03:42 +00:00
7 changed files with 142 additions and 52 deletions

View File

@@ -35,29 +35,53 @@ logger = logging.getLogger(__name__)
# ── Block configuration ───────────────────────────────────────────
# Output token limits per Anthropic docs:
# Opus 4.7: up to 128K output tokens (new tokenizer — ~35% more tokens)
# Sonnet 4.6: up to 64K output tokens
# Streaming required when max_tokens > 21,333
# Generation is structurally deterministic (#204 / WS5): every AI block is
# pinned to ONE model (Opus 4.8) with a per-block reasoning `effort`. The
# generation path is claude_session.query → `claude -p` (subscription, near-zero
# cost; local-only — see reference_claude_generation_path / claude_session
# docstring). Opus 4.7/4.8 REMOVED temperature/top_p/top_k (sending them → HTTP
# 400); the only knob is `effort` (low/medium/high/xhigh/max; default high).
#
# DEPRECATED FIELDS — DO NOT REINTRODUCE: earlier revisions carried per-block
# `temp` (0/0.1/0.4) and `model` ("sonnet"/"opus"/"script"). Both were DEAD
# METADATA: write_block called claude_session.query(prompt, …) WITHOUT model/
# temp/effort, so generation ran on the CLI's default model with no temperature
# knob at all (the table only stored a number in decision_blocks.temperature and
# fed MODEL_MAP→timeout). They are removed here so the table no longer misleads.
# The conceptual origin (gen_type → thinking-budget) now lives in `effort`; see
# docs/block-schema.md §3.
#
# `gen_type` is retained (documentation/UI metadata + audit `generation_type`).
# `model` below is the DISPATCH key only: "script" = template-fill (no LLM),
# "ai" = generated via the pinned Opus model. It is NOT a model alias anymore.
#
# Output token note (Anthropic): Opus 4.8 supports large outputs; streaming is
# handled by the CLI. `max_tokens` is advisory context for callers, not sent.
GENERATION_MODEL = "claude-opus-4-8" # single pinned model for every AI block (#204)
BLOCK_CONFIG = {
"block-alef": {"index": 1, "title": "כותרת מוסדית", "gen_type": "template-fill", "temp": 0, "model": "script"},
"block-bet": {"index": 2, "title": "הרכב הוועדה", "gen_type": "template-fill", "temp": 0, "model": "script"},
"block-gimel":{"index": 3, "title": "צדדים", "gen_type": "template-fill", "temp": 0, "model": "script"},
"block-dalet":{"index": 4, "title": "החלטה", "gen_type": "template-fill", "temp": 0, "model": "script"},
"block-he": {"index": 5, "title": "פתיחה", "gen_type": "paraphrase", "temp": 0.2, "model": "sonnet", "max_tokens": 4096},
"block-vav": {"index": 6, "title": "רקע עובדתי", "gen_type": "reproduction", "temp": 0, "model": "sonnet", "max_tokens": 16384},
"block-zayin":{"index": 7, "title": "טענות הצדדים", "gen_type": "paraphrase", "temp": 0.1, "model": "sonnet", "max_tokens": 16384},
"block-chet": {"index": 8, "title": "הליכים", "gen_type": "reproduction", "temp": 0, "model": "sonnet", "max_tokens": 8192},
"block-tet": {"index": 9, "title": "תכניות חלות", "gen_type": "guided-synthesis", "temp": 0.2, "model": "opus", "max_tokens": 16384},
"block-yod": {"index": 10, "title": "דיון והכרעה", "gen_type": "rhetorical-construction", "temp": 0.4, "model": "opus", "max_tokens": 16384},
"block-yod-alef": {"index": 11, "title": "סיכום", "gen_type": "paraphrase", "temp": 0.1, "model": "sonnet", "max_tokens": 8192},
"block-yod-bet": {"index": 12, "title": "חתימות", "gen_type": "template-fill", "temp": 0, "model": "script"},
"block-alef": {"index": 1, "title": "כותרת מוסדית", "gen_type": "template-fill", "model": "script"},
"block-bet": {"index": 2, "title": "הרכב הוועדה", "gen_type": "template-fill", "model": "script"},
"block-gimel":{"index": 3, "title": "צדדים", "gen_type": "template-fill", "model": "script"},
"block-dalet":{"index": 4, "title": "החלטה", "gen_type": "template-fill", "model": "script"},
"block-he": {"index": 5, "title": "פתיחה", "gen_type": "paraphrase", "model": "ai", "effort": "medium", "max_tokens": 4096},
"block-vav": {"index": 6, "title": "רקע עובדתי", "gen_type": "reproduction", "model": "ai", "effort": "medium", "max_tokens": 16384},
"block-zayin":{"index": 7, "title": "טענות הצדדים", "gen_type": "paraphrase", "model": "ai", "effort": "high", "max_tokens": 16384},
"block-chet": {"index": 8, "title": "הליכים", "gen_type": "reproduction", "model": "ai", "effort": "medium", "max_tokens": 8192},
"block-tet": {"index": 9, "title": "תכניות חלות", "gen_type": "guided-synthesis", "model": "ai", "effort": "high", "max_tokens": 16384},
"block-yod": {"index": 10, "title": "דיון והכרעה", "gen_type": "rhetorical-construction", "model": "ai", "effort": "xhigh", "max_tokens": 16384},
"block-yod-alef": {"index": 11, "title": "סיכום", "gen_type": "paraphrase", "model": "ai", "effort": "high", "max_tokens": 8192},
"block-yod-bet": {"index": 12, "title": "חתימות", "gen_type": "template-fill", "model": "script"},
}
MODEL_MAP = {
"sonnet": "claude-sonnet-4-20250514",
"opus": "claude-opus-4-7",
}
# Default effort when a block lacks an explicit one (defensive; every AI block
# above sets one). High is the safe default per the generation-path reference.
DEFAULT_EFFORT = "high"
# Blocks that take longer (deep reasoning) get the LONG timeout. The interim set
# [he, vav, tet, zayin, chet] and the discussion block all run on the same pinned
# Opus model, so timeout is driven by effort, not by a model split.
_LONG_EFFORTS = frozenset({"high", "xhigh", "max"})
# ── Template blocks (א-ד, יב) ────────────────────────────────────
@@ -120,11 +144,14 @@ TEMPLATE_WRITERS = {
BLOCK_PROMPTS = {
"block-he": """כתוב את בלוק הפתיחה (בלוק ה) של החלטת ועדת ערר.
## כללים:
- פתח ב"לפנינו ערר..." או "עניינה של החלטה זו..."
- הגדר "להלן" מרכזיים: הוועדה המקומית, התכנית/הבקשה, המגרש
- 1-2 סעיפים בלבד
- אין ניתוח, אין ערכי שיפוט, אין ציטוטים מצדדים
## מבנה קבוע (חובה — אותו מבנה בכל תיק):
- **המשפט הראשון פותח תמיד במילה "לפנינו"** — נוסח קבוע: "לפנינו ערר על החלטת
[הוועדה המקומית] מיום [תאריך] בעניין [נושא הבקשה/התכנית]." (אל תשתמש ב"עניינה של
החלטה זו" או בכל פתיח חלופי — הפתיח אחיד.)
- סעיף 1: הצגת הערר במשפט הקבוע לעיל + הגדרת ה"להלן" המרכזיים בסדר קבוע:
הוועדה המקומית → התכנית/הבקשה → המגרש/המקרקעין.
- סעיף 2 (רק אם נדרש להשלמת "להלן" נוספים): הגדרות-נוספות בלבד.
- **בדיוק 1-2 סעיפים.** אין ניתוח, אין ערכי שיפוט, אין ציטוטים מצדדים.
- מספור: 1.
## פרטי התיק:
@@ -441,10 +468,19 @@ async def write_block(
f"Reduce documents or call extract_appraiser_facts first."
)
# Call Claude via Claude Code session (no API)
model_key = block_cfg["model"]
timeout = claude_session.LONG_TIMEOUT if model_key == "opus" else claude_session.DEFAULT_TIMEOUT
content = await claude_session.query(prompt, timeout=timeout, tools="") # prose gen — no tool_use → no error_max_turns
# Call Claude via Claude Code session (no API). #204: pin the model + per-block
# reasoning effort so generation is structurally deterministic — these were
# previously NOT forwarded (the source of inconsistency). model/effort flow
# through claude_session.query → `claude -p --model … --effort …`.
effort = block_cfg.get("effort", DEFAULT_EFFORT)
timeout = claude_session.LONG_TIMEOUT if effort in _LONG_EFFORTS else claude_session.DEFAULT_TIMEOUT
content = await claude_session.query(
prompt,
timeout=timeout,
model=GENERATION_MODEL,
effort=effort,
tools="", # prose gen — no tool_use → no error_max_turns
)
sources = await _collect_block_sources(case_id, block_id)
sources["case_law_ids"] = _precedent_case_law_ids
@@ -455,6 +491,7 @@ async def write_block(
def _build_result(block_id: str, content: str, block_cfg: dict) -> dict:
word_count = len(content.split())
is_ai = block_cfg["model"] == "ai"
return {
"block_id": block_id,
"block_index": block_cfg["index"],
@@ -462,8 +499,14 @@ def _build_result(block_id: str, content: str, block_cfg: dict) -> dict:
"content": content,
"word_count": word_count,
"generation_type": block_cfg["gen_type"],
"model_used": block_cfg["model"],
"temperature": block_cfg["temp"],
# AI blocks record the pinned model; template blocks record "script".
"model_used": GENERATION_MODEL if is_ai else block_cfg["model"],
# The real generation knob (#204). None for template/script blocks.
"effort": block_cfg.get("effort", DEFAULT_EFFORT) if is_ai else None,
# DEPRECATED: temperature is not a real knob on Opus 4.7/4.8 (sending it
# → HTTP 400). Kept only to satisfy decision_blocks.temperature
# NUMERIC(3,2); always 0. Read `effort` instead.
"temperature": 0,
}

View File

@@ -516,9 +516,18 @@ async def export_docx(case_number: str, output_path: str = "") -> str:
# ── Interim draft (pre-ruling) ────────────────────────────────────
# Blocks written for the interim draft, in display order.
# Blocks written for the interim draft, in a FIXED order.
# This is the same content the chair sees in the final decision (same template,
# same skill, same prompts) — minus opening, ruling, summary, signatures.
# same skill, same prompts, same single canonical write path — block_writer) —
# minus ruling, summary, signatures.
#
# Determinism (#204 / WS5): the interim draft is STRUCTURALLY deterministic.
# Each block is generated by block_writer.write_block, which now pins the model
# (Opus 4.8) and a per-block reasoning effort, and uses fixed structural prompts
# (e.g. block-he's opening is always "לפנינו ערר…"). The display order on export
# is fixed separately by docx_exporter._INTERIM_BLOCK_ORDER. There is no parallel
# generation path (G2): write_interim_draft is a thin orchestrator over the same
# write_and_store_block used by the full decision.
_INTERIM_BLOCKS = ["block-he", "block-vav", "block-tet", "block-zayin", "block-chet"]