feat(subagents): state the delegation on its row, not the child's reply

[ericleepi314]

What

A delegation's ⎿ row carried the child's entire answer. That text is addressed to the model — the parent reads it and writes its own reply — so the row restated work the reader had not asked to see, at whatever length the child felt like.

before                                          after
⏺ Subagent(List the sources)                    ⏺ Subagent(List the sources)  (/agents to monitor)
  ⎿  echo: PROBE bash {"command": "sleep 4         ⎿  Done (1 tool use · 27 tokens · 4s)
     && ls -1 src | head -3", "description":
     "peek at src"}                             ⏺ Subagent(Audit the tests)  (/agents to monitor)
                                                  ⎿  Backgrounded agent

Claude Code renders ⎿ Done (2 tool uses · 48.0k tokens · 11s) for a foreground Task and ⎿ Backgrounded agent for a backgrounded one.

Where the counts come from

The child's own session log, which #19 started reading — and only when the join is unambiguous: exactly one delegation in flight and exactly one child settled under it.

The harness's own prompt tells a model to "start independent delegations together in one assistant message", so attributing one child's tokens to another's row would be worse than reporting only the time this call took — which is always this call's own. A parallel delegation therefore still gets Done (4s), with a test pinning that.

The child's reply stays one ctrl+o away. The parent model read it; so can the reader.

Test

• npm run typecheck, npm test — 150 files / 1938 passing
• node scripts/verify-boundary.mjs
• three new cases: foreground summary + reply behind ctrl+o, backgrounded wording, and the two-in-flight fallback
• live captures via scripts/tool-gallery.py subagent subagent_working subagent_background

🤖 Generated with Claude Code