Repository navigation
v1.4.0
This release makes fallback handling steadier and easier to follow.
Fallback models
A failed model run is now sorted into three kinds: out of credits or a billing problem, a temporary problem, or a permanent problem. Out of credits or a billing problem stops using that model at once, with no retry on it. A temporary problem, such as a rate limit, a server error, a timeout, or a network problem, gets one more try in the same session, two tries in total, waiting 1 second, or following the provider's own hint, never waiting more than 15 seconds. A permanent problem, such as an invalid or unsupported request, stops and shows the error, with no retry and no switch.
If a retry succeeds, the model stays in use but is flagged, so the next fresh failure moves to the fallback at once. If the second try fails, the model is set aside for 5 minutes and the work runs once on the fallback. A set-aside model stays aside for a fixed 5 minutes, where it used to double up to 60 minutes, and then the primary is used again. A successful run no longer clears that 5 minutes. Only time passing or the manual restore clears it, and each new failure restarts the 5 minutes.
This is tracked per model, so one exhausted model affects every subagent that uses it, and one restore clears it for all of them. A subagent picks its model when it starts and is never moved in the middle of a run, and a resumed session is not reassigned. When every fallback is also set aside, the last fallback is still tried once instead of failing. A background subagent that fails inside its own run is not re-run on the fallback, but later subagents use the fallback once the model is set aside. An agent with no fallback now keeps OpenCode's own retry behaviour instead of having it switched off.
Sidebar
A subagent that will use a fallback shows it with a small arrow before the model, in orange. Clicking that model cell once arms switch primary?, and a second click on the same cell brings the primary back for every subagent sharing those models, each to its own primary, while other rows are unaffected, and clicking elsewhere disarms it. If the restore fails, a warning reads Could not restore primary model: server plugin unreachable and the row stays on the fallback.
Status tool
fallback_status now reports whether an agent's primary is flagged and lists models set aside with a reason (quota or retries) and remaining wait in milliseconds, replacing the old backoffMs field.
Install
opencode plugin add openchoreographer
npx openchoreographer --no-plugin
Upgrade
opencode plugin update
npx openchoreographer --force
Full Changelog: v1.3.0...v1.4.0