opencode run (headless) exits 0 mid-pipeline after a parallel subagent fan-out, before the model's next turn #37548

Summary

In headless mode, opencode run sometimes ends the session (exit code 0) immediately after a parallel task-tool subagent fan-out returns, before the primary model gets its next turn. The run looks successful but the output stops mid-pipeline — the model never processes the subagents' results.

Environment

• opencode 1.16.2 and 1.18.2 (reproduced on both)
• Linux (Ubuntu, kernel 6.17), zsh
• Primary model opencode-go/glm-5.2 (build agent), custom subagents on opencode-go/* models

Repro shape

A custom command (/deep-research) loads a skill that instructs the model to run a multi-phase pipeline, dispatching parallel task-tool subagents at several phases (search → fetch → verify → synthesize).

opencode run "/deep-research <question>"

Observed three times across the two versions:

1. (1.16.2) All 10 parallel verify subagents completed (stderr progress shows every ✓) — the process exited 0 without the model's tally/synthesis turn ever running.
2. (1.16.2) Same run earlier needed a resume mid-pipeline as well.
3. (1.18.2) The run printed the model's Phase-1 text ("Dispatching 6 search agents in parallel"), dispatched them, and exited 0 — no Phase 2 onward.

When it happens, stdout ends mid-pipeline and the exit code is still 0, so scripts can't tell the run failed.

Expected

The session should continue until the primary model ends a turn with no pending subagent results to process.

Workaround

Resuming the same session continues correctly every time:

opencode run -s <session-id> "Continue exactly where you left off…"

(Note: must be run from the same working directory, since sessions are per-project.)

Happy to provide session exports for the affected runs if useful.