agentctl opencode launch can pick invalid default model openai/gpt-5.4 in isolated supervisor sessions

[c-h-]

Summary

agentctl launch opencode from an isolated OpenClaw supervisor session picked a broken default model and the session exited immediately.

Repro

From an isolated supervisor session on macOS:

agentctl launch opencode \
  --callback-session "agent:kindo:mattermost:channel:4tcg8sejp3bxfxuokh354hjreh" \
  --callback-agent "kindo" \
  --cwd /Users/ms/code/mono-charlie-ci-self-hosted-pilot \
  -p "Reply with exactly OK"

Observed session IDs:

• 92d185b9-0eac-4fc6-a004-4d0878eabcfa
• 8a303fe6-c8cf-4aec-ad78-988147da61f9

Both stopped almost immediately with exit code 1.

Diagnosis

Running opencode directly in the same worktree showed:

ProviderModelNotFoundError: Model not found: openai/gpt-5.4

The generated wrapper script was invoking:

/opt/homebrew/bin/opencode run -- '...prompt...'

with no explicit model, so the environment/default config resolved to invalid openai/gpt-5.4 instead of the expected Kimi default.

Expected

agentctl launch opencode should either:

1. inherit the configured working default model correctly, or
2. fail fast with a clearer error surfaced via agentctl status/peek, ideally including stderr.

Actual

The session is marked stopped, agentctl peek <id> says session not found, and the only easy breadcrumb is the .exit file under ~/.agentctl/opencode-sessions/.