Subagents dispatched sequentially instead of in parallel

[PointFiveBool]

Description

When I ask the model to run multiple independent tasks in parallel, I only ever see one subagent running at a time. Each one finishes before the next starts, even though they should be independent and could run concurrently.

This is a re-report of #14195 (closed due to inactivity) — the bug is still present.

I poked around the code a bit. The session loop in src/session/prompt.ts around line 1734 does tasks.pop() then calls handleSubtask which blocks until the subagent finishes — so it naturally processes one at a time. Seems like it should collect all subtasks from tasks and run them with Effect.forEach(..., { concurrency: "unbounded" }) instead. Something like:

const subtasks = tasks.filter(t => t.type === "subtask")
const remaining = tasks.filter(t => t.type !== "subtask")
if (subtasks.length > 0) {
  yield* Effect.forEach(subtasks, (task) =>
    handleSubtask({ task, model, lastUser, sessionID, session, msgs })
  , { concurrency: "unbounded" })
  continue
}

Each handleSubtask call has its own local state (assistantMessage, part, summaryUserMsg) so they shouldn't step on each other.

There's also a minor thing — the background param on the task tool is hidden from the LLM behind an experimentalBackgroundSubagents flag (task.ts:332), so even if the model wanted to dispatch non-blocking tasks it can't. Might be worth exposing that or at least defaulting it on.

Plugins

None

OpenCode version

v1.15.11

Steps to reproduce

1. Start an opencode session with a model that supports parallel tool calls
2. Ask a task that the LLM will decompose into 2+ independent subtasks (e.g. "Implement these 3 features in parallel: A, B, and C")
3. Observe the subagent dispatch — only 1 subagent runs at a time, each completing before the next starts
4. Expected: all subagents start concurrently and run in parallel

Screenshot and/or share link

N/A

Operating System

Linux (Ubuntu 22.04)

Terminal

Alacritty

[github-actions]

This issue doesn't fully meet our contributing guidelines.

What needs to be fixed:

• No issue template was used. Please use one of the three official templates: Bug Report, Feature Request, or Question. For this report, the Bug Report template is appropriate — it requires a Description field, steps to reproduce, OpenCode version, and Operating System.

Please edit this issue to address the above within 2 hours, or it will be automatically closed.

---

This issue might be a duplicate of an existing issue. Please check:

• Multiple Task tool calls in a single LLM response execute sequentially instead of in parallel #14195: "Multiple Task tool calls in a single LLM response execute sequentially instead of in parallel" — describes the same root cause (tasks.pop() in prompt.ts processing subtasks one at a time), and was previously closed after no activity. If the bug is still present, this new report should reference that prior issue.

If you believe this was flagged incorrectly, please let a maintainer know.

[github-actions]

Thank you for updating the issue and using the Bug Report template — the required fields (description, steps to reproduce, version, OS) are now present.

However, the description still doesn't meet our guidelines. The current write-up reads as AI-generated: it includes an exhaustive root cause analysis, four numbered proposed fixes with code diffs, a structured evidence list with file paths and line numbers, and a level of detail that goes well beyond what a reporter would typically know without having already investigated and written the fix themselves.

What still needs to be fixed:

• Please rewrite the description in your own words. Focus on what you observed (the bug behavior), how to reproduce it, and what you expected to happen. If you have a hunch about the root cause, a brief note is fine — but the extensive proposed-fix code blocks and multi-section analysis should be removed.

A simpler description is perfectly acceptable. Something like:

When I ask the model to run multiple independent tasks in parallel, I only ever see one subagent running at a time. Each one finishes before the next starts, even though they should be independent.

…along with your steps to reproduce and system info is all that's needed.

Please edit the issue to address this, or it may be closed.

[github-actions]

Thanks for updating the issue! Everything looks good now — the description, reproduction steps, and system information are all filled in. We appreciate you taking the time to improve the report.

[JunJieLiu51520]

Hi @jlongster 👋 — I've opened a PR with the parallel dispatch fix described in this issue (Effect.forEach with concurrency: "unbounded"). Would you be able to take a look when you have a moment?

PR: #29819