agentctl launch opencode defaults to OpenAI-backed build model and exits when OPENAI_API_KEY is unset

[c-h-]

Summary

In the Kindo mono workflow, agentctl launch opencode can exit immediately even though the workspace default opencode model is Kimi via Ohm. The launch log shows it selecting build · gpt-5.4 and failing with OpenAI API key is missing, so no agent work actually starts.

Repro

Environment:

• workspace: /Users/ms/code/mono-conversation-compaction
• opencode config default model: ohm/moonshotai/Kimi-K2.5 (~/.config/opencode/opencode.json)
• no OPENAI_API_KEY in the launch environment

Command:

agentctl launch opencode \
  --cwd /Users/ms/code/mono-conversation-compaction \
  --callback-session agent:kindo:main \
  --callback-agent kindo \
  -p "any prompt"

Observed launch log (~/.agentctl/opencode-sessions/launch-*.log):

> build · gpt-5.4
Error: OpenAI API key is missing. Pass it using the 'apiKey' parameter or the OPENAI_API_KEY environment variable.\n```\n\nThe generated session exits immediately (exit file written, no code changes land).\n\n## Expected\n`agentctl launch opencode` should either:\n1. respect the workspace/default opencode model for the actual run, or\n2. fail clearly before launch with a message explaining that the selected build model requires an unavailable provider key.\n\nIt should not silently route onto an OpenAI-backed build path in a mono workspace that is supposed to use Kimi/Ohm by default.\n\n## Workaround\nExplicitly passing the model works:\n```bash\nagentctl launch opencode --model ohm/moonshotai/Kimi-K2.5 ...\n```\n\nOnce I did that, the session launched successfully (`> build · moonshotai/Kimi-K2.5`).\n\n## Notes\nThis was especially confusing because `agentctl launch` itself reported a launched session id, so the failure only became obvious after checking the opencode session logs and seeing that the PR branch had not changed.