fix(playbook): reconstruct full provenance for manual review

reflexio/client/client.py

@@ -63,6 +63,7 @@
 IS_TEST_ENV = os.environ.get("IS_TEST_ENV", "false").strip() == "true"
 
 BACKEND_URL = "http://127.0.0.1:8000" if IS_TEST_ENV else "https://www.reflexio.ai/"
+REVIEW_USER_PLAYBOOKS_TIMEOUT_SECONDS = 600
 
 from reflexio.models.api_schema.domain.entities import (
     UpgradeProfilesRequest,
@@ -1630,7 +1631,11 @@ def review_user_playbooks(
         """Re-review current user playbooks created within a bounded window.
 
         The default ``report_only=True`` runs inline and returns one decision per
-        selected playbook.
+        selected playbook. Each playbook is reviewed from its full finalized
+        extraction-run window plus any cited consolidation evidence, independent
+        of the current extractor window size or source filter. Rows with missing
+        or invalid persisted provenance are skipped. Only cited interactions are
+        reviewer evidence; other run-window interactions provide chronology.
 
         With ``report_only=False`` the server accepts the run and applies it in
         the background, so the response carries only ``run_id`` — not
@@ -1650,6 +1655,11 @@ def review_user_playbooks(
             "POST",
             "/api/review_user_playbooks",
             json=request.model_dump(mode="json"),
+            timeout=(
+                max(self.timeout, REVIEW_USER_PLAYBOOKS_TIMEOUT_SECONDS)
+                if report_only
+                else self.timeout
+            ),
         )
         return ReviewUserPlaybooksResponse(**response)
 

reflexio/models/api_schema/domain/entities.py

@@ -1737,9 +1737,10 @@ class ReviewUserPlaybookEdit(BaseModel):
 class ReviewUserPlaybookResult(BaseModel):
     """One persisted user playbook's re-review outcome.
 
-    ``skip`` means the row could not be reviewed at all (its cited evidence is
-    no longer reconstructable), not that the reviewer chose to leave it alone —
-    that is ``accept``. A skipped row is never written to.
+    ``skip`` means the row could not be reviewed at all because its finalized
+    generation-window or cited-evidence provenance is absent, missing, or
+    invalid, not that the reviewer chose to leave it alone — that is ``accept``.
+    A skipped row is never written to.
     """
 
     user_playbook_id: int = Field(gt=0)

reflexio/server/middleware.py

@@ -25,6 +25,7 @@
 SYNC_REQUEST_TIMEOUT_SECONDS = (
     600  # Longer timeout for synchronous processing (wait_for_response=true)
 )
+SYNC_REQUEST_PATHS = frozenset({"/api/review_user_playbooks"})
 SUSPICIOUS_USER_AGENTS = ["bot", "crawler", "spider", "scraper", "curl", "wget"]
 ALLOWED_EMPTY_UA_PATHS = ["/health", "/"]  # Paths that allow empty user agents
 DEFAULT_MAX_BODY_BYTES = 10 * 1024 * 1024
@@ -128,7 +129,10 @@ async def dispatch(
 
         # Use longer timeout for synchronous processing requests
         timeout = REQUEST_TIMEOUT_SECONDS
-        if request.query_params.get("wait_for_response", "").lower() == "true":
+        if (
+            request.url.path in SYNC_REQUEST_PATHS
+            or request.query_params.get("wait_for_response", "").lower() == "true"
+        ):
             timeout = SYNC_REQUEST_TIMEOUT_SECONDS
 
         try:

reflexio/server/services/playbook/README.md

@@ -66,8 +66,19 @@ do not use this reviewer.
 ### Persisted Review (`review_service.py`)
 
 `POST /api/review_user_playbooks` selects the newest current user playbooks in
-an inclusive creation-time window, capped by `top_k`. A row whose cited evidence
-can no longer be reconstructed yields a `skip` decision and the run continues.
+an inclusive creation-time window, capped by `top_k`. A row whose original
+generation window or cited evidence can no longer be reconstructed yields a
+`skip` decision and the run continues.
+
+Manual review reconstructs context from the full interaction window persisted on
+the row's finalized playbook-extraction run, plus any extra cited interactions
+retained through consolidation. It never substitutes the current extractor
+window or silently falls back to the smaller cited-evidence subset. Automatic
+post-generation review continues to use the generation call's configured window.
+Only the playbook row's cited interaction IDs become candidate evidence units;
+the rest of the generation window is ancillary chronology. Evidence spans are
+rebuilt from those exact stored interactions instead of requiring every cited
+span to remain duplicated in the row's bounded combined `source_span`.
 
 Report mode runs inline and returns `accept`, `edit`, `reject`, and `skip`
 decisions without writes. Apply mode runs in the **background** (one LLM call

reflexio/server/services/playbook/components/reviewer.py

@@ -41,6 +41,10 @@
 }
 
 
+class PlaybookCandidateEvidenceError(ValueError):
+    """A persisted candidate's cited evidence cannot be reconstructed."""
+
+
 class CandidateRevision(BaseModel):
     """Complete replacement fields for one narrowly revised candidate."""
 
@@ -253,42 +257,21 @@ def _candidate_evidence_units(
         units_by_candidate: dict[str, list[CandidateEvidenceUnit]] = {}
         for candidate_index, candidate in enumerate(candidates, start=1):
             candidate_id = f"C{candidate_index}"
-            combined_span = candidate.source_span or ""
             units: list[CandidateEvidenceUnit] = []
             for interaction_id in candidate.source_interaction_ids:
                 mapped_units = by_interaction_id.get(interaction_id)
                 if not mapped_units:
-                    raise ValueError(
+                    raise PlaybookCandidateEvidenceError(
                         f"Reviewer cannot map {candidate_id} evidence to a local turn"
                     )
-                matching_units = [
-                    unit
-                    for unit in mapped_units
-                    if unit.source_span
-                    and (
-                        unit.source_span in combined_span
-                        or combined_span in unit.source_span
-                    )
-                ]
-                if not matching_units:
-                    span_fragments = [
-                        fragment.strip()
-                        for fragment in combined_span.split("\n\n")
-                        if fragment.strip()
-                    ]
-                    matching_units = [
-                        unit
-                        for unit in mapped_units
-                        if any(
-                            fragment == unit.source_span or fragment in unit.source_span
-                            for fragment in span_fragments
-                        )
-                    ]
-                if not matching_units:
-                    raise ValueError(
-                        f"Reviewer cannot resolve {candidate_id} validated source span"
-                    )
-                for mapped_unit in matching_units:
+                # Exact persisted interaction IDs are the durable evidence
+                # identity. Rebuild their bounded visible spans from storage;
+                # do not require every span to survive in the playbook's
+                # combined ``source_span``, which may be truncated or may have
+                # been assembled from multiple rows during consolidation.
+                for mapped_unit in mapped_units:
+                    if not mapped_unit.source_span:
+                        continue
                     if any(
                         unit.turn_ref == mapped_unit.turn_ref
                         and unit.source_span == mapped_unit.source_span
@@ -304,7 +287,9 @@ def _candidate_evidence_units(
                         )
                     )
             if not units:
-                raise ValueError(f"Reviewer received {candidate_id} without evidence")
+                raise PlaybookCandidateEvidenceError(
+                    f"Reviewer received {candidate_id} without evidence"
+                )
             units_by_candidate[candidate_id] = units
         return units_by_candidate