Claude parent shows Waiting while async subagents are still working

[kouyichi]

Bug description

When Claude Code is waiting for asynchronously launched Agent subagents, abtop reports the parent session as Waiting even though the subagents are still actively working and updating their transcripts.

This makes the top-level status misleading when monitoring multiple long-running Claude Code sessions: ◌ Wait normally means the session needs user input, but in this case no input is needed and work is still in progress.

Steps to reproduce

1. Start a Claude Code session.
2. Ask Claude to launch one or more Agent subagents.
3. With current Claude Code, observe that the Agent tool receives an immediate tool_result saying that an async/background agent was launched, including an agentId and output file.
4. Let the parent emit a text-only assistant message while it waits for the subagent results.
5. Open abtop (optionally enable the subagent tree with T).

Expected behavior

The parent session should remain active, preferably Executing, while any of its subagents is still working.

Actual behavior

The parent session changes to Waiting while its subagent transcripts continue to update. The subagent tree may simultaneously show working subagents under a parent marked Wait.

Reproduction evidence

I reproduced this from a real session with two anonymized subagents:

• The first Agent tool received its async result about 176 ms after launch, but its subagent continued writing for another 14m 59s.
• The second received its async result about 61 ms after launch, but continued writing for another 13m 40s.
• After the parent's later text-only assistant message, the two subagents remained active for about 14m 02s and 13m 08s respectively, during which the parent could be classified as Waiting.

A minimal regression fixture using this transcript sequence also reproduces it:

user prompt
assistant Agent tool_use
user tool_result: async agent launched, running in background
assistant text-only message: waiting for research result
subagent JSONL still recently updated

The expected parent status was Executing, but the result was:

left: Waiting
right: Executing

Root cause analysis

The Claude collector currently determines parent status using only:

• active descendant CPU,
• a non-empty current_task (pending tool), or
• a trailing real user message (last_user_ts_ms).

Relevant code:

• Parent status calculation
• current_task is cleared on every assistant entry
• Subagents are collected only after parent status has already been calculated
• collect_subagents already derives working/done

For an async Agent call, Claude Code writes the tool result immediately. A subsequent text-only assistant entry clears current_task, and the subagent may not appear as a CPU-active descendant process. The status therefore falls through to Waiting even though collect_subagents() can see a recently active subagent.

PRs #84 and #101 fixed a closely related case where a normal tool_use remains pending, and their existing regression test passes. However, async Agent calls return their tool_result immediately, so that pending-tool signal cannot represent the remaining subagent lifetime.

Possible fix

As a minimal fix, collect subagents before calculating parent status and classify the parent as Executing when any subagent is working.

The existing 30-second transcript-mtime heuristic may still flicker during a long subagent model generation. A more robust implementation could track the async lifecycle by linking the toolUseId in *.meta.json with the agentId returned by the parent Agent tool result.

Environment

• abtop: 0.5.3
• abtop commit: a3328db20fe189887986e0147080954bc07178a1 (current main and v0.5.3)
• Claude Code: 2.1.212
• OS: Ubuntu 24.04, Linux 6.8.0-117-generic x86_64