Rescue foreground calls can time out before OpenCode finishes large tasks #37

Summary

The rescue flow can invoke opencode-companion.mjs task in the foreground. For large tasks, the Claude Code Bash/tool timeout can expire before OpenCode finishes, leaving the result only in background state/logs or losing the user-visible handoff.

This is separate from #13, which tracks a missing wall-clock timeout inside runTrackedJob itself.

Local evidence

plugins/opencode/agents/opencode-rescue.md says:

• prefer foreground for small, bounded rescue requests
• prefer background for complicated/open-ended work

That guidance relies on the forwarding agent estimating duration correctly. If it chooses foreground for a task that exceeds the host Bash timeout, the command may terminate before output is delivered to the parent conversation.

Expected behavior

Large rescue work should reliably return either:

1. final output within the foreground timeout, or
2. a job id immediately with clear instructions to use /opencode:status and /opencode:result.

Suggested fix

Options:

1. Prefer opencode-companion task --background for any rescue request that may exceed the Bash timeout.
2. Add a companion-level mode that starts asynchronously and returns a job id immediately, rather than relying on Claude-side run_in_background behavior.
3. Teach /opencode:rescue to inspect task size/risk and choose async mode more conservatively.
4. Include the job id in all timeout/failure paths when a worker was started.

Upstream reference

Derived from openai/codex-plugin-cc#122.