[JasonW404]

Target corrected to develop (aa713dab0cc42ae388dc25bc815010e3d411f270). The task commits were rebased so this PR contains only its 21 intended files, without unrelated main backports. Current head: c1c8e56b3e6e5a080e15cf98c1f54f541f4845da. The VM still runs the previously accepted main-based image; develop-based CI and container acceptance must be repeated before treating that head as verified.

Dreaming used the observer class as its default, so semantic summaries raised a missing-argument error and fell back mechanically. Create an isolated observer instance while preserving explicitly supplied observers.

Require automatic memories to declare durable scope and quote exact source evidence. Reject local answer constraints, indirect references to existing preferences, invented paraphrases and assistant-derived preferences. Route extraction through the monitored model wrapper so the actual call is visible in Phoenix.

Previous main-based validation (before the develop rebase):

• 105 focused memory/extraction/Dreaming tests passed; extraction modules have 100% line coverage. 491 Agent service tests, 24 observer/retry tests and 141 OpenAI model tests passed.
• Reran all 14 files that failed in the initial full CI with coverage: all passed. The earlier main-based pytest compatibility adjustment was dropped during the rebase; develop's existing pytest 9.1.1 / pytest-asyncio 1.4.0 pins are preserved. Registered the new service module and updated the four deliberately changed, reviewed prompt hashes.
• Built main/web images from final source e2010ab1404d323ba39becc95b9bdbc082fde5be and deployed through repository scripts on the existing Nexent-Test VM after a clean snapshot restore. Live model/extraction source hashes match HEAD; no dev source mounts.
• Formal real-provider D3 AGT-DREAMING-SEMANTIC-01 passed again on the final containers: actual worker, semantic active version, no fallback.
• Actual browser conversations preserved durable rules while rejecting 200-/120-character table requirements and English/75-word/JSON constraints. Independent real queries provided recall evidence; no statistics were fabricated.
• UI-triggered Dreaming created V1 from four sources, then incrementally merged a newly learned release-record policy into V2. Both semantic summaries were accepted on their first attempt without fallback. V2 preserves the original rules and the new policy, excludes temporary requirements and assistant-added procedures, and is injected into a new conversation with zero history.
• Phoenix project nexent-pr4249-verified: final V2 summary trace 02dcfc3fb724297b810a591416086516; fresh-conversation trace e9f158fd1994ae2cdba9dd0c10b3129a.

The full formal-asset validator has an existing unrelated missing test reference in CHAT-AGENT-DETAIL-20261026; this change's scoped schemas/execution mapping passed. One task-owned synthetic failure record from the earlier indirect-reference test was soft-deleted after retaining evidence, then the request was replayed against the corrected build. Original user data and real recall statistics were preserved.

[codecov]

Codecov Report

❌ Patch coverage is 98.73418% with 1 line in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...end/management/services/memory/extraction_scope.py	98.38%	0 Missing and 1 partial ⚠️

📢 Thoughts on this report? Let us know!