[magnus919]

Attached research can now supply evidence when a user asks a session question. The session resolves the saved source text when it is used, checks that it has not changed, and refuses references whose research has been deleted or expired.

The API, CLI and MCP also gain explicit follow-up execution. Previewing a follow-up does no research. Executing it validates the selected evidence and starts a separate run with a stable retry key.

The public scrape test exposed a related gap: a normal request could accept challenge text before trying browser recovery. The shared scraper now escalates those pages. A verified browser solve can clear a narrow false alarm caused by a page mentioning CAPTCHA; unresolved widgets and other challenge markers remain errors and are not cached.

Validation:

• 4,707 workspace unit/service tests and 42 subtests passed, with 21 skips; agent type checks passed.
• Isolated Valkey tests cover attachment ownership, revision checks and fresh-connection reads.
• A document test uses real Parse/API HTTP servers and Valkey, with CLI subprocesses and MCP tools. It checks upload, questions, exact citations, deletion, foreign-owner rejection and a relevant passage beyond character 8,000.
• The capture correction passed 168 focused tests and type checks across 53 scraper files. Its unchanged public API container probe passed in runtime CI, including normal-request recovery and unresolved-challenge refusal without caching.

Model responses in these tests are controlled. This proves integration and evidence handling, not live answer quality. Final PostgreSQL and container checks remain pending.

Depends on #670 and contributes to the 1.0.1 cleanup; it does not complete the release checklist by itself.

[factory-droid]

Droid encountered an error —— View job

Droid Exec exceeded the maximum of 100 turns and was stopped. This usually means the agent got stuck in a loop.

---