Opencode relaunches inherit unsupported openai/gpt-5.4 model and look like success

[c-h-]

Summary

Supervisor-triggered agentctl launch opencode relaunches for ENG-8709 never start the coding task because Opencode tries to use openai/gpt-5.4, which is not configured, then exits with ProviderModelNotFoundError. agentctl records the session as status: stopped and the wrapper exit file contains 0, so the supervisor can misclassify the launch as a normal completion.

Impact

This silently blocks automated PR follow-up work. In my case, PR #9271 kept getting relaunched, but the relaunch sessions never actually ran.

Evidence

From ~/.agentctl/opencode-sessions/launch-1773382944447.log:

ProviderModelNotFoundError
 data: {
  providerID: "openai",
  modelID: "gpt-5.4",
  suggestions: [],
}
Error: Model not found: openai/gpt-5.4.

The affected relaunch session state was:

• session id: 677bb005-67d5-4f74-ba9d-b991ec8f1995
• adapter: opencode
• cwd: /Users/ms/code/mono-ENG-8709
• status in ~/.agentctl/state.json: stopped
• exit file ~/.agentctl/opencode-sessions/677bb005-67d5-4f74-ba9d-b991ec8f1995.exit: 0

I confirmed the same launch failure on at least these prior supervisor relaunches for the same PR/worktree:

• d91b396f-21d9-4103-9869-4a19c07fa549
• e999e119-925f-4e95-9990-91c41eaaf5d8
• 677bb005-67d5-4f74-ba9d-b991ec8f1995

Expected

• If the model is invalid, the launch should fail loudly with a non-zero session exit / failure status.
• Supervisor relaunches should not inherit an unsupported openai/gpt-5.4 model override for Opencode.

Questions / hypotheses

• Is agentctl accidentally inheriting the caller session model (gpt-5.4) when launching Opencode?
• Should Opencode launches ignore incompatible caller model defaults unless an explicit compatible model was requested?