feat: review chat — ask the run why it concluded a finding

Read-only Q&A on the review screen, per finding and per run, answered from
the job's own artifacts (evidence, cluster, extraction, verification, Brain
merge, sheet index, cover reconciliation, job.log). It never mutates findings,
decisions, or the report.

Turns are logged job-locally (review/chat_log.jsonl, transcript at
/jobs/{id}/review-chat/log) and to a cross-job feedback store
(REVIEW_FEEDBACK_DIR), which now also receives review decisions with their
category/severity corrections.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0115gGtrSxXE9DKvS9XPFSoT

backend/.env.example

@@ -51,6 +51,21 @@ AGENT_REVIEW_AUDIT_SAMPLE=5
 # Allow future cross-job review-feedback aggregation to include source_text/images/comments
 REVIEW_AGGREGATE_INCLUDE_TEXT=false
 
+# Review chat: read-only Q&A about findings and coverage on the review screen.
+# It explains what the run did from the job's artifacts; it never changes a
+# finding, a decision, or the report.
+ENABLE_REVIEW_CHAT=true
+# Model for chat answers (blank inherits TEXT_MODEL)
+REVIEW_CHAT_MODEL=
+REVIEW_CHAT_MAX_TOKENS=4096
+# Prior turns replayed into a thread's prompt
+REVIEW_CHAT_HISTORY_TURNS=6
+# Max job.log lines searched into the chat's context bundle
+REVIEW_CHAT_LOG_LINES=40
+REVIEW_CHAT_MAX_QUESTION_CHARS=2000
+# Cross-job store for review decisions + chat turns (blank = backend/outputs/_feedback)
+REVIEW_FEEDBACK_DIR=
+
 # Pipeline tuning
 PDF_DPI=100
 MAX_PAGES=60

backend/config.py

@@ -129,6 +129,26 @@ AGENT_REVIEW_AUDIT_SAMPLE = int(os.getenv("AGENT_REVIEW_AUDIT_SAMPLE", "5"))
 # NOTE: currently unwired - reserved for future cross-job aggregation tooling.
 REVIEW_AGGREGATE_INCLUDE_TEXT = os.getenv("REVIEW_AGGREGATE_INCLUDE_TEXT", "false").strip().lower() in ("1", "true", "yes")
 
+# -- Review chat (ask-the-run Q&A on the review screen) --------------
+# A read-only explainer: it answers "why did the run decide X?" from the job's
+# own artifacts and never mutates findings, decisions, or code. Every turn is
+# appended to <out_dir>/review/chat_log.jsonl AND to the cross-job feedback
+# store (REVIEW_FEEDBACK_DIR) so answers are available to future prompt priors.
+# HISTORY_TURNS caps how much of a thread is replayed into the prompt;
+# LOG_LINES caps how many job.log lines are searched into the context bundle.
+ENABLE_REVIEW_CHAT = _flag("ENABLE_REVIEW_CHAT", "true")
+REVIEW_CHAT_MODEL = os.getenv("REVIEW_CHAT_MODEL", "") or TEXT_MODEL
+REVIEW_CHAT_MAX_TOKENS = int(os.getenv("REVIEW_CHAT_MAX_TOKENS", "4096"))
+REVIEW_CHAT_HISTORY_TURNS = int(os.getenv("REVIEW_CHAT_HISTORY_TURNS", "6"))
+REVIEW_CHAT_LOG_LINES = int(os.getenv("REVIEW_CHAT_LOG_LINES", "40"))
+REVIEW_CHAT_MAX_QUESTION_CHARS = int(os.getenv("REVIEW_CHAT_MAX_QUESTION_CHARS", "2000"))
+# Cross-job feedback store: where review decisions and chat turns accumulate so
+# a future run can be primed with "what humans corrected last time". Job-local
+# artifacts stay the source of truth; this is the append-only roll-up.
+# (OUTPUT_DIR is defined further down; keep this in sync with it.)
+REVIEW_FEEDBACK_DIR = os.getenv("REVIEW_FEEDBACK_DIR", "") or os.path.join(
+    _BASE_DIR, "outputs", "_feedback")
+
 # -- Hybrid (local text LLM) ----------------------------------------
 # Optional OpenAI-compatible local endpoint (e.g. a vLLM box) for the text-only
 # QAQC stages. Vision stages ALWAYS use OpenRouter. The user picks hybrid per

backend/main.py

@@ -23,6 +23,7 @@ import backend.jobs
 from backend import config, llm
 from backend.jobs import PIPELINE_MODES, create_job, get_job, _set
 from backend.pipeline.pdf_processor import render_page_jpeg
+from backend.review import chat as review_chat
 from backend.review.feedback import decision_to_label, write_label
 from backend.review.finalizer import finalize_review
 from backend.review.store import ReviewStore
@@ -252,6 +253,67 @@ def finalize_review_endpoint(job_id: str):
     return {"status": "finalizing"}
 
 
+# Statuses in which the review chat may be used. The chat is read-only, so it
+# stays available after finalization - a reviewer often asks "why did it say
+# that?" about a report they have already sent.
+_CHAT_STATES = ("needs_review", "reviewing", "finalizing", "done", "finalization_error")
+
+
+def _chat_out_dir(job_id: str) -> str:
+    """Resolve a job's output dir for a chat request, or raise an HTTP error."""
+    job = get_job(job_id)
+    if not job:
+        raise HTTPException(status_code=404, detail="Job not found")
+    if job.get("status") not in _CHAT_STATES:
+        raise HTTPException(status_code=409, detail={
+            "detail": f"review chat is not available for a job in status {job.get('status')}",
+        })
+    return job.get("out_dir") or os.path.join(config.OUTPUT_DIR, job_id)
+
+
+@app.post("/jobs/{job_id}/review-chat")
+def review_chat_ask(job_id: str, payload: dict):
+    """Ask one question about a finding, or about the run as a whole.
+
+    Read-only: this answers from the job's artifacts and appends to the chat
+    log. It never changes a finding, a decision, or the report.
+    """
+    out_dir = _chat_out_dir(job_id)
+    store = ReviewStore(out_dir, create=False)
+    try:
+        turn = review_chat.ask(
+            job_id=job_id,
+            out_dir=out_dir,
+            question=payload.get("question"),
+            review_item_id=payload.get("review_item_id") or None,
+            queue=store.read_queue(),
+            decisions=store.read_decisions(),
+        )
+    except review_chat.ChatError as e:
+        raise HTTPException(status_code=422, detail=str(e))
+    except Exception as e:
+        # A failed model call is an upstream problem, not a bad request; the
+        # review screen shows it inline and the reviewer can retry.
+        raise HTTPException(status_code=502, detail=f"review chat failed: {e}")
+    return {"turn": turn}
+
+
+@app.get("/jobs/{job_id}/review-chat")
+def review_chat_history(job_id: str, review_item_id: Optional[str] = None):
+    """Logged chat turns, oldest first. Without review_item_id, all threads."""
+    out_dir = _chat_out_dir(job_id)
+    turns = review_chat.read_log(out_dir, review_item_id=review_item_id)
+    return {"turns": turns, "enabled": config.ENABLE_REVIEW_CHAT}
+
+
+@app.get("/jobs/{job_id}/review-chat/log")
+def review_chat_log(job_id: str):
+    """The chat log as a readable transcript: issue, questions, findings."""
+    out_dir = _chat_out_dir(job_id)
+    markdown = review_chat.render_log_markdown(review_chat.read_log(out_dir))
+    return Response(content=markdown, media_type="text/markdown; charset=utf-8")
+
+
 @app.get("/jobs/{job_id}/sheet-image/{page}")
 def sheet_image(job_id: str, page: int):
     """Render one page of a completed job's source PDF as JPEG (sheet viewer)."""

backend/review/chat.py

@@ -0,0 +1,368 @@
+"""Review-screen chat: ask the run why it concluded something.
+
+Read-only by construction. The chat reads job artifacts, calls one LLM, and
+appends a log record; it never mutates findings, review decisions, or the
+report, and the prompt forbids it from emitting code or config changes.
+
+Every turn is logged twice, on purpose:
+
+- ``<out_dir>/review/chat_log.jsonl`` - job-local, the auditable record of what
+  was asked about which finding and what came back.
+- ``REVIEW_FEEDBACK_DIR/chat_turns.jsonl`` - cross-job, append-only, so the
+  corrections a reviewer makes in conversation ("that is not a floor drain, it
+  is a power floor box") accumulate somewhere a future run can be primed from.
+  Nothing reads this yet; writing it is what makes that possible later.
+"""
+
+import json
+import os
+import uuid
+from datetime import datetime, timezone
+from typing import Any, Dict, List, Optional
+
+from backend import config
+from backend.llm import call_json
+from backend.pipeline._serialize import dumps
+from backend.review.chat_context import build_context
+from backend.review.chat_prompts import (
+    REVIEW_CHAT_SYSTEM_PROMPT,
+    REVIEW_CHAT_USER_PROMPT,
+)
+from backend.review.feedback import append_shared_feedback
+
+_ANSWERABLE = {"yes", "partial", "no"}
+_ASSESSMENTS = {"looks_supported", "looks_unsupported", "cannot_tell", "not_applicable"}
+_CONFIDENCE = {"high", "medium", "low"}
+_MAX_FINDINGS = 12
+_MAX_EVIDENCE = 12
+
+
+class ChatError(Exception):
+    """Raised for a caller-fixable problem (bad question, chat disabled)."""
+
+
+def _now() -> str:
+    return datetime.now(timezone.utc).isoformat()
+
+
+def _one_of(value: Any, allowed: set, default: str) -> str:
+    text = str(value or "").strip().lower()
+    return text if text in allowed else default
+
+
+def _clean_question(raw: Any) -> str:
+    question = str(raw or "").strip()
+    if not question:
+        raise ChatError("question is required")
+    if len(question) > config.REVIEW_CHAT_MAX_QUESTION_CHARS:
+        raise ChatError(
+            f"question is too long (max {config.REVIEW_CHAT_MAX_QUESTION_CHARS} characters)")
+    return question
+
+
+def _log_path(out_dir: str) -> str:
+    return os.path.join(out_dir, "review", "chat_log.jsonl")
+
+
+def read_log(out_dir: str, review_item_id: Optional[str] = None,
+             scope_only: bool = False) -> List[Dict[str, Any]]:
+    """Chat turns for this job, oldest first.
+
+    ``review_item_id`` filters to one finding's thread; with ``scope_only`` and
+    no id, returns only the run-scope turns. Corrupt lines are skipped rather
+    than failing the read - a truncated log must not hide the rest.
+    """
+    path = _log_path(out_dir)
+    if not os.path.isfile(path):
+        return []
+    turns: List[Dict[str, Any]] = []
+    try:
+        with open(path, encoding="utf-8") as f:
+            for line in f:
+                line = line.strip()
+                if not line:
+                    continue
+                try:
+                    turn = json.loads(line)
+                except json.JSONDecodeError:
+                    continue
+                if not isinstance(turn, dict):
+                    continue
+                if review_item_id is not None:
+                    if turn.get("review_item_id") != review_item_id:
+                        continue
+                elif scope_only and turn.get("review_item_id") is not None:
+                    continue
+                turns.append(turn)
+    except OSError:
+        return []
+    return turns
+
+
+def _append_log(out_dir: str, turn: Dict[str, Any]) -> None:
+    """Append one turn as a JSON line; never raises on I/O failure."""
+    try:
+        os.makedirs(os.path.join(out_dir, "review"), exist_ok=True)
+        with open(_log_path(out_dir), "a", encoding="utf-8") as f:
+            f.write(json.dumps(turn) + "\n")
+    except OSError as e:
+        print(f"[ReviewChat] chat log write failed: {e}")
+
+
+def _issue_snapshot(item: Optional[Dict]) -> Optional[Dict[str, Any]]:
+    """The issue as it stood when asked about - the log's 'issue in question'.
+
+    Copied rather than referenced by id so the log stays readable after
+    finalization renumbers or suppresses the finding.
+    """
+    if not item:
+        return None
+    payload = item.get("payload") or {}
+    if item.get("kind") == "clean_cluster":
+        return {
+            "review_item_id": item.get("review_item_id"),
+            "kind": item.get("kind"),
+            "cluster_key": payload.get("key"),
+            "location": payload.get("location"),
+            "disciplines": payload.get("disciplines"),
+        }
+    return {
+        "review_item_id": item.get("review_item_id"),
+        "kind": item.get("kind"),
+        "issue_id": payload.get("issue_id"),
+        "source_stage": payload.get("source_stage"),
+        "category": payload.get("category"),
+        "severity": payload.get("severity"),
+        "confidence": payload.get("confidence"),
+        "location": payload.get("location"),
+        "disciplines": payload.get("disciplines"),
+        "sheets": payload.get("sheets"),
+        "description": payload.get("description"),
+        "blocking": item.get("blocking"),
+    }
+
+
+def _history_block(turns: List[Dict[str, Any]]) -> str:
+    if not turns:
+        return ""
+    recent = turns[-config.REVIEW_CHAT_HISTORY_TURNS:]
+    lines = ["Earlier turns in this thread (oldest first):"]
+    for turn in recent:
+        lines.append(f"Reviewer: {turn.get('question') or ''}")
+        lines.append(f"You: {turn.get('answer') or ''}")
+    lines.append("")
+    return "\n".join(lines)
+
+
+def _normalize_answer(parsed: Optional[Dict]) -> Optional[Dict[str, Any]]:
+    """Coerce the model's JSON into the log/API shape, or None if unusable."""
+    if not isinstance(parsed, dict):
+        return None
+    answer = str(parsed.get("answer") or "").strip()
+    if not answer:
+        return None
+    findings = [
+        str(item).strip()
+        for item in (parsed.get("findings") or [])
+        if isinstance(item, (str, int, float)) and str(item).strip()
+    ][:_MAX_FINDINGS]
+    evidence = []
+    for item in (parsed.get("evidence_cited") or [])[:_MAX_EVIDENCE]:
+        if not isinstance(item, dict):
+            continue
+        evidence.append({
+            "artifact": str(item.get("artifact") or "").strip() or None,
+            "sheet": item.get("sheet"),
+            "quote": str(item.get("quote") or "").strip() or None,
+            "why_it_matters": str(item.get("why_it_matters") or "").strip() or None,
+        })
+    correction = parsed.get("suggested_category_correction")
+    correction = str(correction).strip() if correction else ""
+    missing = parsed.get("missing_information")
+    return {
+        "answer": answer,
+        "findings": findings,
+        "evidence_cited": evidence,
+        "answerable": _one_of(parsed.get("answerable"), _ANSWERABLE, "partial"),
+        "missing_information": str(missing).strip() if missing else None,
+        "assessment_of_finding": _one_of(parsed.get("assessment_of_finding"),
+                                         _ASSESSMENTS, "cannot_tell"),
+        # The feedback signal: a reviewer correcting a misidentification in
+        # conversation ("that is a power floor box") lands here as structured
+        # data instead of dying in free text.
+        "suggested_category_correction": correction or None,
+        "confidence": _one_of(parsed.get("confidence"), _CONFIDENCE, "low"),
+    }
+
+
+def _feedback_record(turn: Dict[str, Any]) -> Dict[str, Any]:
+    """Cross-job roll-up of one turn: metadata + the correction signal.
+
+    Mirrors the privacy stance of the decision labels - no images, no raw sheet
+    dumps. The question and answer ARE carried, because a chat turn without its
+    question is not usable as feedback; keep this store job-internal.
+    """
+    issue = turn.get("issue") or {}
+    return {
+        "kind": "review_chat_turn",
+        "turn_id": turn.get("turn_id"),
+        "job_id": turn.get("job_id"),
+        "created_at": turn.get("created_at"),
+        "review_item_id": turn.get("review_item_id"),
+        "scope": turn.get("scope"),
+        "issue_id": issue.get("issue_id"),
+        "source_stage": issue.get("source_stage"),
+        "category": issue.get("category"),
+        "severity": issue.get("severity"),
+        "confidence": issue.get("confidence"),
+        "sheets": issue.get("sheets"),
+        "question": turn.get("question"),
+        "answer": turn.get("answer"),
+        "findings": turn.get("findings"),
+        "assessment_of_finding": turn.get("assessment_of_finding"),
+        "suggested_category_correction": turn.get("suggested_category_correction"),
+        "answerable": turn.get("answerable"),
+        "model": turn.get("model"),
+    }
+
+
+def ask(job_id: str, out_dir: str, question: str,
+        review_item_id: Optional[str] = None,
+        queue: Optional[List[Dict]] = None,
+        decisions: Optional[Dict[str, Dict]] = None) -> Dict[str, Any]:
+    """Answer one reviewer question and log the turn.
+
+    Returns the logged turn. Raises ChatError for a bad question or a disabled
+    chat, and RuntimeError when the model call fails outright (the caller maps
+    both to HTTP status codes).
+    """
+    if not config.ENABLE_REVIEW_CHAT:
+        raise ChatError("review chat is disabled (ENABLE_REVIEW_CHAT=false)")
+    question = _clean_question(question)
+    queue = queue or []
+
+    item = next((candidate for candidate in queue
+                 if candidate.get("review_item_id") == review_item_id), None)
+    if review_item_id and item is None:
+        raise ChatError(f"unknown review_item_id {review_item_id!r}")
+
+    context = build_context(out_dir, review_item_id, queue, decisions, question)
+    history = read_log(out_dir, review_item_id=review_item_id) if review_item_id \
+        else read_log(out_dir, scope_only=True)
+
+    scope_line = (
+        f"Scope: this question is about review item {review_item_id}."
+        if item else
+        "Scope: this question is about the run as a whole, not one finding."
+    )
+    user_text = (REVIEW_CHAT_USER_PROMPT
+                 .replace("{scope_line}", scope_line)
+                 .replace("{question}", question)
+                 .replace("{history_block}", _history_block(history))
+                 .replace("{context}", dumps(context)))
+
+    parsed = call_json(
+        system_prompt=REVIEW_CHAT_SYSTEM_PROMPT,
+        user_text=user_text,
+        max_tokens=config.REVIEW_CHAT_MAX_TOKENS,
+        model=config.REVIEW_CHAT_MODEL,
+        usage_stage="review.chat",
+    )
+    answer = _normalize_answer(parsed)
+    if answer is None:
+        raise RuntimeError("the model did not return a usable answer")
+
+    turn = {
+        "turn_id": uuid.uuid4().hex[:12],
+        "job_id": job_id,
+        "created_at": _now(),
+        "review_item_id": review_item_id,
+        "scope": context.get("scope"),
+        "issue": _issue_snapshot(item),
+        "question": question,
+        "reviewer_decision_at_time": context.get("reviewer_decision_so_far"),
+        "artifacts_consulted": sorted(
+            name for name, present
+            in (context.get("artifacts_available") or {}).items() if present
+        ),
+        "model": config.REVIEW_CHAT_MODEL,
+        **answer,
+    }
+    _append_log(out_dir, turn)
+    append_shared_feedback(_feedback_record(turn))
+    return turn
+
+
+def render_log_markdown(turns: List[Dict[str, Any]]) -> str:
+    """Human-readable transcript: the issue, the questions, the findings.
+
+    Grouped by review item so one finding's whole thread reads together, with
+    run-scope questions last under their own heading.
+    """
+    by_item: Dict[str, List[Dict[str, Any]]] = {}
+    for turn in turns:
+        by_item.setdefault(turn.get("review_item_id") or "", []).append(turn)
+
+    lines = ["# Review chat log", ""]
+    if not turns:
+        lines.append("_No questions have been asked about this run._")
+        return "\n".join(lines) + "\n"
+    lines.append(f"{len(turns)} turn(s) across {len(by_item)} thread(s).")
+    lines.append("")
+
+    for item_id in sorted(by_item, key=lambda key: (key == "", key)):
+        item_turns = by_item[item_id]
+        issue = next((turn.get("issue") for turn in item_turns if turn.get("issue")), None)
+        if not item_id:
+            lines += ["## Run-scope questions", "",
+                      "_Not about a single finding._", ""]
+        elif issue:
+            lines.append(f"## {issue.get('issue_id') or item_id}")
+            lines.append("")
+            meta = [
+                ("Category", issue.get("category")),
+                ("Severity", issue.get("severity")),
+                ("Run confidence", issue.get("confidence")),
+                ("Location", issue.get("location")),
+                ("Sheets", ", ".join(str(s) for s in issue.get("sheets") or []) or None),
+                ("Stage", issue.get("source_stage")),
+            ]
+            for label, value in meta:
+                if value:
+                    lines.append(f"- **{label}:** {value}")
+            if issue.get("description"):
+                lines += ["", f"> {issue['description']}"]
+            lines.append("")
+        else:
+            lines += [f"## {item_id}", ""]
+
+        for turn in item_turns:
+            lines.append(f"### Q ({turn.get('created_at') or ''})")
+            lines += ["", turn.get("question") or "", "", "**Answer**", "",
+                      turn.get("answer") or "", ""]
+            if turn.get("findings"):
+                lines.append("**Findings**")
+                lines.append("")
+                lines += [f"- {finding}" for finding in turn["findings"]]
+                lines.append("")
+            if turn.get("evidence_cited"):
+                lines += ["**Evidence cited**", ""]
+                for item in turn["evidence_cited"]:
+                    where = item.get("artifact") or "?"
+                    sheet = f" ({item['sheet']})" if item.get("sheet") else ""
+                    quote = item.get("quote") or ""
+                    lines.append(f"- `{where}`{sheet}: \"{quote}\"")
+                    if item.get("why_it_matters"):
+                        lines.append(f"  - {item['why_it_matters']}")
+                lines.append("")
+            tail = [
+                ("Answerable", turn.get("answerable")),
+                ("Assessment", turn.get("assessment_of_finding")),
+                ("Confidence", turn.get("confidence")),
+                ("Missing", turn.get("missing_information")),
+                ("Suggested correction", turn.get("suggested_category_correction")),
+                ("Model", turn.get("model")),
+            ]
+            lines.append(" | ".join(f"{label}: {value}" for label, value in tail if value))
+            lines.append("")
+    return "\n".join(lines) + "\n"

backend/review/chat_context.py

@@ -0,0 +1,315 @@
+"""Evidence bundles for the review chat.
+
+The chat is an explainer, not an investigator: it may only answer from what the
+run actually produced. This module assembles that material from the job's own
+artifacts and hands the model a bounded, slimmed view.
+
+Two shapes, matching the two kinds of question a reviewer asks:
+
+- item scope  ("why does it think the AC unit is on the ground?") - the finding,
+  its evidence, the cluster the finding came from, the sheets those assertions
+  were extracted from, any wave-5b verification verdict, the Brain's merge/drop
+  decision, and the reviewer's own saved decision.
+- run scope   ("why didn't it pick up the Civil set?") - the sheet index by
+  discipline, the deterministic cover-index reconciliation, per-stage counts,
+  what got suppressed and why, and matching job.log lines.
+
+Everything here is read-only and degrades to empty on a missing or corrupt
+artifact; a chat request must never be the thing that breaks a review screen.
+"""
+
+import json
+import os
+import re
+from typing import Any, Dict, List, Optional
+
+from backend import config
+
+# Assertion/evidence text is quoted back verbatim so the reviewer can check the
+# answer against the sheet, but a whole cluster of them would swamp the prompt.
+_MAX_CLUSTER_ASSERTIONS = 40
+_MAX_SHEET_ASSERTIONS = 25
+_MAX_SOURCE_TEXT_CHARS = 400
+_MAX_SHEETS_IN_ROSTER = 400
+_MAX_SUPPRESSED = 25
+_MAX_LOG_LINE_CHARS = 400
+
+
+def _read_json(path: str, default):
+    try:
+        with open(path, encoding="utf-8") as f:
+            return json.load(f)
+    except (OSError, json.JSONDecodeError):
+        return default
+
+
+def _truncate(value: Any, limit: int = _MAX_SOURCE_TEXT_CHARS) -> Any:
+    if not isinstance(value, str) or len(value) <= limit:
+        return value
+    return value[:limit] + "..."
+
+
+def _slim_assertion(assertion: Dict) -> Dict:
+    """Drop base64/bookkeeping; keep what explains where a value came from."""
+    out = {
+        "sheet_number": assertion.get("sheet_number"),
+        "discipline": assertion.get("discipline"),
+        "attribute": assertion.get("attribute"),
+        "value": assertion.get("value"),
+        "source_text": _truncate(assertion.get("source_text")),
+        "location_key": assertion.get("location_key"),
+        "normalized_value": assertion.get("normalized_value"),
+        "disputed": assertion.get("disputed"),
+    }
+    return {key: value for key, value in out.items() if value is not None}
+
+
+def _slim_finding(finding: Dict) -> Dict:
+    """The finding as the run recorded it, including how it was checked."""
+    out = {
+        "issue_id": finding.get("issue_id"),
+        "source_stage": finding.get("source_stage"),
+        "agent": finding.get("agent"),
+        "category": finding.get("category"),
+        "severity": finding.get("severity"),
+        "confidence": finding.get("confidence"),
+        "location": finding.get("location"),
+        "disciplines": finding.get("disciplines"),
+        "sheets": finding.get("sheets"),
+        "description": _truncate(finding.get("description"), 1200),
+        "recommended_resolution": finding.get("recommended_resolution"),
+        "code_reference": finding.get("code_reference"),
+        "risk_score": finding.get("risk_score"),
+        "recommended_priority": finding.get("recommended_priority"),
+        "scope_id": finding.get("scope_id"),
+        "evidence": [
+            {
+                "discipline": item.get("discipline"),
+                "sheet": item.get("sheet"),
+                "source_text": _truncate(item.get("source_text")),
+                "asserted_value": item.get("asserted_value"),
+            }
+            for item in (finding.get("evidence") or [])
+            if isinstance(item, dict)
+        ],
+        # Wave 5b / Brain-clarify re-checked some findings against fresh sheet
+        # images + the text layer. When present this is the single best answer
+        # to "did it actually look again?", so it is never dropped.
+        "verification": finding.get("verification"),
+        "clarification_of": finding.get("clarification_of"),
+    }
+    return {key: value for key, value in out.items() if value is not None}
+
+
+def _slim_sheet(sheet: Dict, limit: int = _MAX_SHEET_ASSERTIONS) -> Dict:
+    assertions = sheet.get("assertions") or []
+    out = {
+        "sheet_number": sheet.get("sheet_number"),
+        "sheet_title": sheet.get("sheet_title"),
+        "discipline": sheet.get("discipline"),
+        "level": sheet.get("level"),
+        "page_number": sheet.get("page_number"),
+        "assertion_count": len(assertions),
+        "assertions": [_slim_assertion(item) for item in assertions[:limit]],
+    }
+    if len(assertions) > limit:
+        out["assertions_omitted"] = len(assertions) - limit
+    return out
+
+
+def _discipline_roster(sheet_index: Dict, sheets: List[Dict]) -> Dict[str, List[str]]:
+    """Sheet numbers grouped by discipline - the 'is Civil in here?' answer.
+
+    Built from the classified sheet index when there is one, falling back to
+    raw extraction, so an empty/failed index stage does not read as "no sheets".
+    """
+    entries = (sheet_index or {}).get("sheet_index") or []
+    if not entries:
+        entries = [
+            {"sheet_number": sheet.get("sheet_number"),
+             "discipline": sheet.get("discipline")}
+            for sheet in sheets or []
+        ]
+    roster: Dict[str, List[str]] = {}
+    for entry in entries:
+        if not isinstance(entry, dict):
+            continue
+        discipline = str(entry.get("discipline") or "unknown")
+        number = entry.get("sheet_number") or entry.get("sheet_id") or "?"
+        bucket = roster.setdefault(discipline, [])
+        if len(bucket) < _MAX_SHEETS_IN_ROSTER and number not in bucket:
+            bucket.append(str(number))
+    return roster
+
+
+def _log_excerpt(out_dir: str, terms: List[str], limit: int) -> List[str]:
+    """job.log lines mentioning any search term, newest last.
+
+    The run log is where stage skips, retries, and coverage decisions are
+    recorded ("[Code] gated off", "[Extract] page 12 empty"), which is often
+    the literal answer to "why didn't it look at X".
+    """
+    path = os.path.join(out_dir, "job.log")
+    needles = [term.lower() for term in terms if term and len(str(term)) >= 2]
+    if not needles or not os.path.isfile(path):
+        return []
+    hits: List[str] = []
+    try:
+        with open(path, encoding="utf-8", errors="replace") as f:
+            for line in f:
+                lowered = line.lower()
+                if any(needle in lowered for needle in needles):
+                    hits.append(_truncate(line.rstrip("\n"), _MAX_LOG_LINE_CHARS))
+    except OSError:
+        return []
+    return hits[-limit:]
+
+
+def _stage_terms(question: str) -> List[str]:
+    """Search terms for the log: quoted sheet-ish tokens plus long words.
+
+    Deliberately crude - this only decides which log lines get shown, and an
+    over-broad match is bounded by REVIEW_CHAT_LOG_LINES anyway.
+    """
+    tokens = re.findall(r"[A-Za-z][A-Za-z0-9.\-]{2,}", question or "")
+    stop = {"the", "why", "did", "not", "and", "for", "was", "were", "does",
+            "this", "that", "with", "from", "what", "how", "you", "its",
+            "it's", "there", "when", "have", "has", "any", "are", "but"}
+    return [token for token in tokens if token.lower() not in stop][:12]
+
+
+def build_context(out_dir: str, review_item_id: Optional[str],
+                  queue: Optional[List[Dict]] = None,
+                  decisions: Optional[Dict[str, Dict]] = None,
+                  question: str = "") -> Dict[str, Any]:
+    """Assemble the evidence bundle for one chat turn.
+
+    ``review_item_id`` selects item scope; None (or an id not in the queue)
+    gives run scope. Missing artifacts degrade to empty sections rather than
+    raising - the model is told what is missing via ``artifacts_available``.
+    """
+    report = _read_json(os.path.join(out_dir, "conflicts.json"), {}) or {}
+    snapshot = _read_json(os.path.join(out_dir, "agent", "memory.json"), {}) or {}
+    summary = report.get("summary") or {}
+    sheets = snapshot.get("sheets") or []
+    sheet_index = report.get("sheet_index") or snapshot.get("sheet_index") or {}
+
+    context: Dict[str, Any] = {
+        "scope": "run",
+        "run": {
+            "source": report.get("source"),
+            "pipeline_mode": summary.get("pipeline_mode"),
+            "agent_status": summary.get("agent_status"),
+            "sheets_analyzed": summary.get("sheets_analyzed") or len(sheets),
+            "by_stage": summary.get("by_stage"),
+            "conflicts_found": summary.get("conflicts_found"),
+            "by_severity": summary.get("by_severity"),
+            "models_used": summary.get("models_used"),
+            # Stage gating is the answer to a whole class of "why didn't it
+            # check X" questions, so it is stated rather than left implied.
+            "code_review_enabled": config.ENABLE_CODE_REVIEW,
+        },
+        "sheets_by_discipline": _discipline_roster(sheet_index, sheets),
+        "sheet_reconciliation": report.get("sheet_reconciliation"),
+        "missing_expected_sheets": (sheet_index or {}).get("missing_expected_sheets"),
+        "suppressed_by_the_run": [
+            {
+                "issue_id": item.get("issue_id"),
+                "category": item.get("category"),
+                "description": _truncate(item.get("description"), 300),
+                "verification": item.get("verification"),
+            }
+            for item in (snapshot.get("suppressed") or [])[:_MAX_SUPPRESSED]
+            if isinstance(item, dict)
+        ],
+        "artifacts_available": {
+            "conflicts.json": bool(report),
+            "agent/memory.json": bool(snapshot),
+            "job.log": os.path.isfile(os.path.join(out_dir, "job.log")),
+        },
+    }
+
+    item = None
+    for candidate in queue or []:
+        if candidate.get("review_item_id") == review_item_id:
+            item = candidate
+            break
+    if item is None:
+        context["log_excerpt"] = _log_excerpt(
+            out_dir, _stage_terms(question), config.REVIEW_CHAT_LOG_LINES)
+        return context
+
+    context["scope"] = "item"
+    payload = item.get("payload") or {}
+    context["review_item"] = {
+        "review_item_id": item.get("review_item_id"),
+        "kind": item.get("kind"),
+        "blocking": item.get("blocking"),
+        "review_triggers": item.get("reasons"),
+    }
+    saved = (decisions or {}).get(review_item_id) or {}
+    if saved:
+        context["reviewer_decision_so_far"] = {
+            "decision": saved.get("decision"),
+            "reason_code": saved.get("reason_code"),
+            "comment": _truncate(saved.get("comment")),
+        }
+
+    if item.get("kind") == "clean_cluster":
+        context["cluster"] = {
+            "key": payload.get("key"),
+            "location": payload.get("location"),
+            "disciplines": payload.get("disciplines"),
+            "kind": payload.get("kind"),
+            "assertions": [_slim_assertion(a)
+                           for a in (payload.get("assertions") or [])[:_MAX_CLUSTER_ASSERTIONS]],
+        }
+        cited_sheets = [a.get("sheet_number") for a in payload.get("assertions") or []]
+    else:
+        context["finding"] = _slim_finding(payload)
+        cited_sheets = list(payload.get("sheets") or [])
+        cited_sheets += [e.get("sheet") for e in payload.get("evidence") or []
+                         if isinstance(e, dict)]
+        scope_id = str(payload.get("scope_id") or "")
+        if scope_id.startswith("conflict:"):
+            cluster_key = scope_id.split(":", 1)[1]
+            cluster = next((c for c in snapshot.get("clusters") or []
+                            if c.get("key") == cluster_key), None)
+            if cluster is not None:
+                assertions = cluster.get("assertions") or []
+                context["originating_cluster"] = {
+                    "key": cluster.get("key"),
+                    "location": cluster.get("location"),
+                    "disciplines": cluster.get("disciplines"),
+                    "kind": cluster.get("kind"),
+                    "disputed_attributes": cluster.get("disputed_attributes"),
+                    "assertion_count": len(assertions),
+                    "assertions": [_slim_assertion(a)
+                                   for a in assertions[:_MAX_CLUSTER_ASSERTIONS]],
+                }
+                cited_sheets += [a.get("sheet_number") for a in assertions]
+        issue_id = payload.get("issue_id")
+        brain_decisions = [
+            decision for decision in snapshot.get("decisions") or []
+            if isinstance(decision, dict) and (
+                decision.get("kept_issue_id") == issue_id
+                or issue_id in (decision.get("finding_refs") or []))
+        ]
+        if brain_decisions:
+            context["brain_decisions"] = brain_decisions[:10]
+
+    # The sheets the finding actually rests on, with their raw extraction -
+    # this is what lets the model say "it read 'MOUNTED ON GRADE' off M2.1".
+    wanted = {str(number) for number in cited_sheets if number}
+    if wanted:
+        context["source_sheets"] = [
+            _slim_sheet(sheet) for sheet in sheets
+            if str(sheet.get("sheet_number") or "") in wanted
+        ]
+
+    context["log_excerpt"] = _log_excerpt(
+        out_dir,
+        _stage_terms(question) + sorted(wanted),
+        config.REVIEW_CHAT_LOG_LINES,
+    )
+    return context

backend/review/chat_prompts.py

@@ -0,0 +1,36 @@
+"""Prompts for the review-screen chat (read-only run explainer)."""
+
+REVIEW_CHAT_SYSTEM_PROMPT = """You are the explainer for a completed automated construction-drawing review run. A human reviewer is working through the review queue and is asking you why the run reached a particular conclusion.
+
+Your ONLY job is to explain what the run did and why, using the run's own artifacts, which are supplied to you as a JSON context bundle. You are a witness to the run, not a participant in it.
+
+HARD RULES - never break these:
+- You do NOT write, propose, suggest, or output code, patches, diffs, file edits, configuration changes, prompt changes, or shell commands. If the reviewer asks for any of those, say that this chat only explains findings, and answer the underlying question in construction-review terms instead.
+- You do NOT change, re-decide, confirm, reject, or re-score any finding. The reviewer owns that decision; the radio buttons on their screen are the only thing that changes a finding. You may explain what the evidence supports, and you may say plainly that a finding looks wrong, but you never state that a finding "has been" changed.
+- You answer ONLY from the supplied context bundle. You have no access to the PDF, to sheets that were not extracted, or to anything outside the bundle. Never invent a sheet number, a quotation, a dimension, or a stage that is not in the bundle.
+- Separate what the run RECORDED from what you INFER. Attribute recorded facts to the artifact they came from ("the extractor recorded ... on M2.1"). Mark reasoning of your own as inference.
+- When the bundle does not contain the answer, say so directly and name what is missing and which artifact would have held it. "The Civil sheets were never extracted, so there are no Civil assertions to compare" is a good answer. Guessing is not.
+
+HOW TO ANSWER "why does it think X":
+Trace the chain backwards through the bundle and quote it: the finding's evidence, the assertions in the originating cluster, the source_text the extractor pulled off each sheet, any verification verdict from the re-check pass, and the Brain's merge or drop decision. If a value is marked disputed, or the verification status is refuted or unverified, say so - that is usually the real answer.
+
+HOW TO ANSWER "why didn't it pick up X":
+Work through the bundle's coverage material in this order and report which one explains it: (1) sheets_by_discipline - was the discipline in the set at all? (2) sheet_reconciliation - did the cover sheet's own index declare sheets that were never identified (declared_not_in_set)? (3) run.by_stage and run.code_review_enabled - was the responsible stage gated off or did it produce nothing? (4) suppressed_by_the_run - was something found and then dropped? (5) log_excerpt - did the run log record a skip, a retry, or an empty page? Name the specific reason. If several are possible, say which is best supported and what would confirm it.
+
+Be direct and concrete. Quote verbatim source_text when it carries the answer. A short, specific, evidence-anchored answer is worth more than a thorough hedge. Use plain ASCII. Respond only with valid JSON."""
+
+REVIEW_CHAT_USER_PROMPT = """A reviewer is asking about this run. Answer from the context bundle only.
+
+Respond ONLY with a valid JSON object - no markdown fences, no prose outside the JSON:
+{"answer":"your direct explanation to the reviewer, plain text, no markdown headings","findings":["one short factual determination per item - what you established about this question, each standing on its own"],"evidence_cited":[{"artifact":"which part of the bundle, e.g. 'finding.evidence' or 'source_sheets[M2.1]' or 'log_excerpt'","sheet":"sheet number or null","quote":"verbatim text from the bundle","why_it_matters":"one sentence"}],"answerable":"yes | partial | no","missing_information":"what the bundle would need to answer fully, or null if fully answered","assessment_of_finding":"looks_supported | looks_unsupported | cannot_tell | not_applicable","suggested_category_correction":"if the reviewer is telling you the run misidentified an object, the object they say it actually is, e.g. 'power floor box'; otherwise null","confidence":"high | medium | low"}
+
+Set assessment_of_finding to not_applicable for run-scope questions that are not about one finding. Set suggested_category_correction to null unless the reviewer is asserting a correction - do not invent one.
+
+{scope_line}
+
+Reviewer's question:
+{question}
+
+{history_block}
+Context bundle (the complete set of artifacts you may reason from):
+{context}"""

backend/review/feedback.py

@@ -1,9 +1,19 @@
-"""Feedback labels: one label artifact per human-review decision, for metrics."""
+"""Feedback labels: one label artifact per human-review decision, for metrics.
+
+Labels are written twice: job-locally under ``<out_dir>/review/`` (the
+auditable record for that run) and, via ``append_shared_feedback``, to the
+cross-job store at ``config.REVIEW_FEEDBACK_DIR``. The shared store is
+append-only and nothing reads it yet - it exists so that a later pass can prime
+a run with what reviewers corrected on previous sets without having to walk
+every job directory.
+"""
 
 import json
 import os
 from datetime import datetime, timezone
 
+from backend import config
+
 
 def _as_dict(value) -> dict:
     return value if isinstance(value, dict) else {}
@@ -21,6 +31,7 @@ def decision_to_label(queue_item: dict, decision: dict, job: dict) -> dict:
     payload = _as_dict(queue_item.get("payload"))
     summary = _as_dict(_as_dict(job.get("report")).get("summary"))
     return {
+        "kind": "review_decision",
         "review_item_id": queue_item.get("review_item_id"),
         "job_id": job.get("job_id"),
         "pipeline_mode": job.get("pipeline_mode"),
@@ -30,6 +41,11 @@ def decision_to_label(queue_item: dict, decision: dict, job: dict) -> dict:
         "confidence": payload.get("confidence"),
         "decision": decision.get("decision"),
         "reason_code": decision.get("reason_code"),
+        # The reviewer's structured corrections. Carried here (and into the
+        # cross-job store) so "wrong category" survives as data rather than
+        # only as free text on the suppressed issue.
+        "category_correction": decision.get("category_correction"),
+        "severity_correction": decision.get("severity_correction"),
         "location": payload.get("location"),
         "disciplines": payload.get("disciplines"),
         "sheets": payload.get("sheets"),
@@ -40,7 +56,12 @@ def decision_to_label(queue_item: dict, decision: dict, job: dict) -> dict:
 
 
 def write_label(out_dir: str, label: dict) -> None:
-    """Append one label as a JSON line; never raises on I/O failure."""
+    """Append one label job-locally and to the cross-job store.
+
+    Never raises on I/O failure: a lost label must not fail the save that
+    produced it.
+    """
+    append_shared_feedback(label)
     try:
         review_dir = os.path.join(out_dir, "review")
         os.makedirs(review_dir, exist_ok=True)
@@ -49,3 +70,49 @@ def write_label(out_dir: str, label: dict) -> None:
             f.write(json.dumps(label) + "\n")
     except OSError as e:
         print(f"[Review] feedback label write failed: {e}")
+
+
+def append_shared_feedback(record: dict) -> None:
+    """Append one record to the cross-job feedback store; never raises.
+
+    One JSONL file per record ``kind`` so a reader can pick up decisions and
+    chat turns independently. Failure here is logged and swallowed: the
+    cross-job roll-up is a convenience, and losing a line must never fail the
+    review action that produced it.
+    """
+    try:
+        kind = str(record.get("kind") or "misc")
+        os.makedirs(config.REVIEW_FEEDBACK_DIR, exist_ok=True)
+        name = "chat_turns.jsonl" if kind == "review_chat_turn" else "decisions.jsonl"
+        path = os.path.join(config.REVIEW_FEEDBACK_DIR, name)
+        with open(path, "a", encoding="utf-8") as f:
+            f.write(json.dumps(record) + "\n")
+    except OSError as e:
+        print(f"[Review] shared feedback write failed: {e}")
+
+
+def read_shared_feedback(kind: str = "review_decision") -> list:
+    """Read the cross-job store for one record kind, oldest first.
+
+    Corrupt lines are skipped so a partial write cannot hide the rest.
+    """
+    name = "chat_turns.jsonl" if kind == "review_chat_turn" else "decisions.jsonl"
+    path = os.path.join(config.REVIEW_FEEDBACK_DIR, name)
+    if not os.path.isfile(path):
+        return []
+    records = []
+    try:
+        with open(path, encoding="utf-8") as f:
+            for line in f:
+                line = line.strip()
+                if not line:
+                    continue
+                try:
+                    value = json.loads(line)
+                except json.JSONDecodeError:
+                    continue
+                if isinstance(value, dict):
+                    records.append(value)
+    except OSError:
+        return []
+    return records