Task tool fails with exclusivity error when invoked during a prompt

[astrobot-houston]

Originally reported by @khoi in #175.

All community-submitted pull requests are automatically converted to issues (bugs) & discussions (feature requests, enhancements) where they can be triaged and prioritized. Once prioritized, a PR implementation is created automatically.

Describe the Bug

When the model-facing task tool is invoked during a prompt() call, it fails with an exclusivity error. The task tool internally calls the public session.task() API, which is guarded by runExclusive('task'). Since tool execution happens while the parent session is already inside runExclusive('prompt'), the exclusivity check rejects the call with:

[flue] Session "default" is already running prompt. Start another session for parallel conversation branches.

This means any agent that uses the task tool during a prompt cannot spawn child tasks, making the tool effectively unusable in its primary context.

Expected Behavior

The model-facing task tool should be able to create and run a child task while the parent session is inside a prompt() call, since spawning child tasks from within tool execution is the intended use case.

Steps to Reproduce

1. Create a session with a createTaskSession callback configured.
2. Start a prompt() call on the session (which enters runExclusive('prompt')).
3. During tool execution within that prompt, invoke the task tool to spawn a child task.
4. The call fails with Session "default" is already running prompt. Start another session for parallel conversation branches.

Minimal reproduction:

const session = new Session({
	name: 'default',
	storageKey: 'session',
	affinityKey: 'affinity',
	config,
	env,
	store: new InMemorySessionStore(),
	existingData: null,
	createTaskSession: async () =>
		({
			prompt: async (text: string) => ({ text }),
			getAssistantText: () => '',
			getLatestAssistantMessageId: () => undefined,
			close: () => {},
			abort: () => {},
		}) as Session,
});

(session as unknown as { activeOperation: string }).activeOperation = 'prompt';

await (
	session as unknown as {
		runTaskForTool: (
			params: { prompt: string },
			tools: [],
			role: undefined,
			model: undefined,
			thinkingLevel: undefined,
		) => Promise<unknown>;
	}
).runTaskForTool({ prompt: 'inspect the diff' }, [], undefined, undefined, undefined);
// Throws: Session "default" is already running prompt...

---

Original implementation from #175 by @khoi

[khoi]

Some context:

I was building a code review agent using flue and ran into the problem where

[flue] tool:start  read  .github/workflows/goodreview.yml
[flue] tool:done   read  (478 chars)
[flue] thinking:start
  **Considering finalization options**
  I’m in the process of finalizing things, but before jumping to a decision, I think it might be useful to involve a child reviewer. Their perspective could bring fresh insights, which might help me make a better choice. I wonder how their views would influence the overall outcome. Let’s take that step before wrapping things up! It could really make a difference.
[flue] tool:start  task
[flue] tool:error   task  [flue] Session "default" is already running prompt. Start another session for parallel conversation branches.
[flue] thinking:start
[flue] tool:start  glob  AGENTS.md
[flue] tool:done   glob
[flue] thinking:start
[flue] tool:start  finish
[flue] tool:done   finish
{
[flue] Done.
  "reviewPosted": true,
  "mode": "full",
  "findings": 0,
  "overallCorrectness": "patch is correct"
}

[sidpalas]

Looks like a duplicate of #169

[sidpalas]

This is fixed in https://github.com/withastro/flue/tree/v0.8.0!

(specifically b1609ac)

[FredKSchott]

Resolved in b1609ac and covered by the current runtime tests. Closing as a duplicate of #169.