[magnus919]

What changes

Durable research could finish searching and collecting sources, then fail without producing a report. Its construction request allowed only 2,048 output tokens even when the deployment allowed more. In the reproduced failure, the model used that entire allowance for reasoning and returned no answer.

New research runs now use the configured completion allowance, capped at 8,192 tokens. They also select relevant passages before constructing the answer, so the model does not have to process every captured page in full. Captured sources remain available in full, and citations are checked against the passages actually shown to the model.

Older saved runs keep their original request format so recovery can reuse their recorded responses. This change does not retry calls whose outcome is unknown.

Validation

• Construction, model review and planning tests passed locally.
• Ten isolated Valkey recovery tests passed, including old and new policies and recovery without repeated model calls.
• Added tests for source selection across eight documents, exact citation locations, omitted passages and oversized lines.
• Independent code review found no blocking issue. Repository hooks passed.
• CI must also pass the HTTP integration tests with isolated Valkey, PostgreSQL and controlled providers before merge.

A separate real-model confirmation over the eight already captured sources completed construction, review and all three reports in 132.8 seconds using two calls. No searches were repeated. An earlier attempt timed out at 90.1 seconds and is retained as a failed observation; the successful run does not establish consistent latency. The deployment remains unchanged.

[factory-droid]

Droid encountered an error —— View job

Droid Exec exited with code 1:
Tool "Edit" skipped (Nous-GPT-6-Luna uses ApplyPatch instead).
Tool "Create" skipped (Nous-GPT-6-Luna uses ApplyPatch instead).

---

[factory-droid]

Droid encountered an error —— View job

Droid Exec exited with code 1:
Tool "Edit" skipped (Nous-GPT-6-Luna uses ApplyPatch instead).
Tool "Create" skipped (Nous-GPT-6-Luna uses ApplyPatch instead).

---

[factory-droid]

Droid finished @magnus919's task —— View job

---

The candidate set contains no inline findings to validate. The full diff contains bounded evidence selection, preserved line references, citation checks against presented excerpts, and policy-version-specific budgets, with legacy replay behavior unchanged.

0 inline comments posted