Inquire whether the v2 backend mode subagent has truly improved compared to v1 and if there is relevant evidence to support this.

[Hisir0909]

Because I have used Codex's backend subagent before, I found that it often has issues:

1. The main agent frequently does not wait for the subagent and proceeds to work on its own.
2. The main agent has no sense of time and does not know when the subagent will finish.
3. Multiple subagents returning in quick succession frequently interrupt the main agent's workflow.

I also tried the beta version of omo-slim. After submitting a task, I reverted to v1 within a few tens of seconds (it still had the same problems as Codex).

So I want to know: does omo-slim v2 combined with opencode's backend subagent truly offer an improvement over v1, or is it just a new paradigm that does not guarantee a positive improvement — or could even be a negative one?

[mhenke]

Short answer: yes, v2 is genuinely better than v1 — and the beta you tried isn't representative of v2.

The three problems you hit are all solved architecturally in v2, not papered over:

1. Orchestrator not waiting for subagents — v2 uses hook-driven completion instead of polling. OpenCode's runtime injects the finished task's result into the orchestrator; it never sits in a poll loop.
2. No sense of time — the Background Job Board tracks each task's lifecycle explicitly: running → completed / error / cancelled / timed_out. The orchestrator prompt references that state, so it knows what's still in flight.
3. Interruptions — completion events are batched. Results accumulate between the orchestrator's turns, so it gets one consolidated update rather than a stream of interruptions.

The beta you reverted from predated OpenCode's native background subagent support (added in v1.14.51+). v2 is built on top of that native infrastructure — not a reimplementation of the old polling approach. For the full picture, see docs/background-orchestration.md and src/hooks/task-session-manager/.