[magnus919]

The durable research endpoint currently returns a scripted example instead of researching the submitted question. This change connects it to the configured search engine, scraper and model, and saves the actual sources and reports in PostgreSQL.

It also records external requests and their completed responses. After a restart, completed responses can be reused. A request that may have reached a provider but has no saved response fails with outcome_unknown rather than silently spending again. Cancellation and deletion prevent an old worker from publishing a late result.

This is the first part of the 1.0.1 cleanup. Its initial limits are one search, three newly acquired sources, eight total sources including selected evidence, and 180 seconds. Multi-query research, callbacks and time evaluation follow in #674. Workspace integration, storage migration and existing-job recovery are separate PRs.

Validation: 4,686 local unit/service tests and 42 subtests passed, with 21 skips. Type checks and documentation checks passed. CI uses controlled search, capture and model services with real PostgreSQL and Valkey. The latest CI rerun includes a correction to the test container's SQL inputs; final required checks remain pending. These tests verify integration, not live answer quality. No deployment was performed.

Related: #666. Supersedes the guard-only change in #667. This PR alone does not complete the release checklist.

[factory-droid]

Droid finished @magnus919's task —— View job

---

The changes connect admitted research runs to durable external-call receipts and retained evidence, with scoped coordination and recovery coverage. No high-confidence actionable issues were identified in the validated candidates.

0 inline comments posted