Free profile: a refused session's move picks a Zen model stalled as busy minutes earlier #204

Closed
opened 2026-10-18 10:58:00 -04:00 by cmoriarty · 3 comments
Owner

Follow-up to #200, seen on production in run 78 (c1228b4), openspec.apply[6], ~14:21-14:24Z. Zen's glm-5.3-flash-free was stalled as busy (3 capacity refusals). About 90 s later, when Kilo refused the session (#202), #200's move sent the session back to that Zen model, because the router was dry. It refused 5 more times (about a minute) before the step moved back to the router and parked.

To do: the refusal move and zen_for skip a Zen model stalled as busy, and wait for the earliest reset instead, unless that wait is longer than the model's cool-down.

Follow-up to #200, seen on production in run 78 (c1228b4), `openspec.apply[6]`, ~14:21-14:24Z. Zen's `glm-5.3-flash-free` was stalled as busy (3 capacity refusals). About 90 s later, when Kilo refused the session (#202), #200's move sent the session back to that Zen model, because the router was dry. It refused 5 more times (about a minute) before the step moved back to the router and parked. **To do:** the refusal move and `zen_for` skip a Zen model stalled as `busy`, and wait for the earliest reset instead, unless that wait is longer than the model's cool-down.
Author
Owner

Not seen on production yet. Shipped in #301 (5355515): the refusal move and zen_for pass over a Zen model cooling as busy, and the stall's cause is kept across restarts. Tests cover both.

Run 144 (Free, quick-fix) had no refusal move and no Zen model cooling, so nothing to see. This stays open until a Free run shows a refused session passing over a busy Zen model.

**Not seen on production yet.** Shipped in #301 (5355515): the refusal move and `zen_for` pass over a Zen model cooling as busy, and the stall's cause is kept across restarts. Tests cover both. Run 144 (Free, quick-fix) had no refusal move and no Zen model cooling, so nothing to see. This stays open until a Free run shows a refused session passing over a busy Zen model.
Author
Owner

Looked at with the Free batch of 2026-10-30; left open. The fix shipped in #301. What is missing is proof: production has not hit the case, and the simulated runs cannot reach Zen. Runs reach Zen through opencode's built-in opencode provider, with no URL Braid sets. A proof needs a Zen lane in the sims: a URL override in the rendered config, and the fake answering as Zen does, "busy" refusals included. That is harness work of its own.

**Looked at with the Free batch of 2026-10-30; left open.** The fix shipped in #301. What is missing is proof: production has not hit the case, and the simulated runs cannot reach Zen. Runs reach Zen through opencode's built-in `opencode` provider, with no URL Braid sets. A proof needs a Zen lane in the sims: a URL override in the rendered config, and the fake answering as Zen does, "busy" refusals included. That is harness work of its own.
Author
Owner

Proven in PR 362. The skip itself was already in the code; nothing proved it until now.

The sims now have a fake OpenCode Zen. free-zen-busy-passed-over has a refused session's move pass over a busy Zen model. It passes, and it fails stuck on that model with both the busy check and the cooling order disabled. Either one alone protects the step.

No brief changed, so there is no replay to run.

Proven in PR 362. The skip itself was already in the code; nothing proved it until now. The sims now have a fake OpenCode Zen. `free-zen-busy-passed-over` has a refused session's move pass over a busy Zen model. It passes, and it fails stuck on that model with both the busy check and the cooling order disabled. Either one alone protects the step. No brief changed, so there is no replay to run.
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#204
No description provided.