Multiple Task tool calls in a single LLM response execute sequentially instead of in parallel

[gareys]

Description

When an LLM generates multiple Task (subtask) tool calls in a single response, they execute one at a time instead of concurrently. In packages/opencode/src/session/prompt.ts, the session loop does tasks.pop() to grab a single subtask, awaits it, then continues the loop — so even 3 simultaneous Task calls run sequentially.

Regular tool calls already run in parallel via Promise.all in the batch/processor path, but subtask parts bypass that entirely.

Steps to reproduce

1. Create an agent that dispatches 3 Task tool calls in one response (e.g. an orchestrator that sends work to 3 subagents)
2. Observe that each subagent runs only after the previous one completes
3. Expected: all 3 subagents run concurrently

OpenCode version

Latest dev branch (commit c6bd320)

Operating System

macOS 15.5 (darwin-arm64)

[rekram1-node]

I dont think this is quite accurate. Did you read what is considered a task in that file? It's only compaction or subtask invocations. A subtask invocation only occurs from manual slash command invocation that specifies a subagent or is instructed to run as a subtask.

In both cases we want these to occur in order.

[rekram1-node]

The llm can call the task tool multiple times at once and those will all be evaluated in parallel. you can test this easily by editing the task tool to sleep for like 30 seconds if the agent is a particular one then have any model that can call parallel tools and tell it to call 3 agents at once.

You will see the other agents complete before the sleep one, even if sleep one is called first.

[gareys]

I guess in my experience I'm not seeing that.

It could be my agent setup

I have an orchestrator agent called "review-team" who specifically checks out a PR then farms out review to 3 different model agents (opus, codex, gemini).

- agents/
   - codex-reviewer.md
   - gemini-reviewer.md
   - opus-reviewer.md
   - review-team.md (main agent)

review-team is mode: primary

The issue arises in this setup. My main agent can't spin up the *-reviewer subagents in parallel. Even though my prompt specifically instructs this, it is being routed through a sequential task.pop() procedure.

My PR addresses this #14196 and now my agents get spun up in parallel.

Perhaps the structure of my agents is not ideal and specifically prevents me from parallelizing sub-agent tasks.

What do you think @rekram1-node?

I can share the agents with you so you can see the issue if you'd like.

[rekram1-node]

if u read this code:

opencode/packages/opencode/src/session/prompt.ts

Line 1744 in c162074

export async function command(input: CommandInput) {

The command function is the only place we are assigning "subtask"

So the task tool shouldnt be hitting this code path at all.

[gareys]

I'm not encountering the issue anymore, but I promise you that they were being invoked sequentially earlier today 🤣

I can try to grab a video next time it happens. I run this tool quite often so please leave the issue open through the end of the week.

Thanks for the quick replies today.

[github-actions]

To stay organized issues are automatically closed after 60 days of no activity. If the issue is still relevant please open a new one.

[bendtherules]

@rekram1-node Quick q: If a model response has multiple tool_calls (like read 2 files), it seems like opencode is doing only first tool_call and ignoring the rest.
Is this expected behaviour?