OpenCode/Trellis dispatch can still mis-detect child completion in long sessions

[1508324011]

Summary

PR #147 improves the OpenCode dispatch prompt by switching child phase examples from background TaskOutput polling to synchronous Task(..., run_in_background: false) calls.

However, after additional testing, this appears to be only a mitigation, not a root fix.

In long conversations, the dispatch agent can still mis-handle child completion, and the underlying TaskOutput / background_output behavior can still report completion or error before the child session has truly finished.

Why PR #147 is not sufficient by itself

The PR changes Trellis prompts/templates, but it does not change the underlying runtime semantics of background task completion.

We now have evidence that:

1. In real usage, background_output / TaskOutput can disagree with session_info / session_read

   • surface background status may return early or report error
   • the underlying child session can still show continued progress

2. This is especially visible in longer sessions

   • even if the dispatch prompt says "use synchronous child calls"
   • the agent may still drift in long conversations

3. There is at least one environment-level signal that the runtime/wrapper can override the requested mode

   • we observed a system message equivalent to:
     • run_in_background=false was automatically converted to background mode for reliability monitoring
   • this suggests the effective execution semantics may be controlled below the Trellis prompt layer

Current understanding

Based on inspection so far:

• Trellis seems to control prompts, context injection, and workflow guidance
• Trellis does not appear to implement the actual TaskOutput / background_output completion state machine
• Therefore, PR fix(opencode): make dispatch wait for child tasks #147 should be treated as a patch that reduces one bad dispatch pattern, not as a complete fix for early child-completion reporting

Likely root-cause area

The real fix likely needs to happen in one of these lower layers:

1. the OpenCode background task runtime / completion state machine
2. an agent wrapper that maps task completion before the child session reaches a terminal state
3. possibly oh-my-openagent or another environment-level integration layer if it wraps subagent execution

What would be a real fix

A real fix would make background task completion depend on the child session reaching a terminal state, rather than only the wrapper task returning.

In other words, TaskOutput / background_output should not report completed/error while the associated child session is still actively progressing.

Suggested next step

• keep PR fix(opencode): make dispatch wait for child tasks #147 as a mitigation / workflow improvement
• investigate the lower-level runtime semantics for child-session completion
• if needed, document that Trellis alone cannot guarantee correct completion detection until the runtime layer is fixed

Related

• PR fix(opencode): make dispatch wait for child tasks #147