Repair keeps the backend a run started with, so changing Repair in Settings doesn't reach a run that has failed #325

Closed
opened 2026-10-30 02:21:40 -04:00 by cmoriarty · 1 comment
Owner

What happens
Run 157 (Free, thorough) failed. Asking Repair "why did this step fail?" was refused: "Repair cannot use the Free profile yet: choose another backend for Repair in Settings." The operator then set Repair to vLLM and saved. The same refusal came back. repair_backend is one of the run's snapshot fields (RUN_SNAPSHOT_KEYS), frozen when its first agent step starts, because its opencode server is configured with the Repair provider then. A run started on Free can never be repaired, whatever Settings says after.

Fix
Repair reads its backend from the settings now, not the run's copy. When that differs from what the run's server was configured with, Repair renders the run's opencode config again with the repair provider before its session. Repair works after the run, so restarting the run's server is fine.

Done when
On a failed Free run, choosing vLLM for Repair in Settings and asking again gets an answer. A test covers the backend changing after the run started.

**What happens** Run 157 (Free, thorough) failed. Asking Repair "why did this step fail?" was refused: "Repair cannot use the Free profile yet: choose another backend for Repair in Settings." The operator then set Repair to vLLM and saved. The same refusal came back. `repair_backend` is one of the run's snapshot fields (`RUN_SNAPSHOT_KEYS`), frozen when its first agent step starts, because its opencode server is configured with the Repair provider then. A run started on Free can never be repaired, whatever Settings says after. **Fix** Repair reads its backend from the settings now, not the run's copy. When that differs from what the run's server was configured with, Repair renders the run's opencode config again with the `repair` provider before its session. Repair works after the run, so restarting the run's server is fine. **Done when** On a failed Free run, choosing vLLM for Repair in Settings and asking again gets an answer. A test covers the backend changing after the run started.
Author
Owner

Confirmed on production (add5fd1, PRs #328 and #330); closing. Run 157 started on Free while Repair followed Free. With vLLM then chosen for Repair, its Repair answered from qwen3.8-27b: "The run failed at the agent.code-review step: all five of its attempts died on the very first model call with a 401 provider.auth..." That diagnosis is right; see #331. It took two fixes: Repair now reads its backend when asked (#328), and a Free run's config carries the repair provider (#330).

**Confirmed on production (add5fd1, PRs #328 and #330); closing.** Run 157 started on Free while Repair followed Free. With vLLM then chosen for Repair, its Repair answered from `qwen3.8-27b`: "The run failed at the `agent.code-review` step: all five of its attempts died on the very first model call with a 401 `provider.auth`..." That diagnosis is right; see #331. It took two fixes: Repair now reads its backend when asked (#328), and a Free run's config carries the `repair` provider (#330).
Sign in to join this conversation.
No labels
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
cmoriarty/braid#325
No description provided.