[rekram1-node]

The subagent tool told the model not to "ask the subagent for status", with no exception. A strict model could read that as "never message a running child", even when the user asks what it's doing. The point of the rule is to stop the model checking in on its own, so this says that instead.

 tool output:
- DO NOT sleep, poll for progress, ask the subagent for status, or duplicate this subagent's work; avoid working with the same files or topics it is using.
+ DO NOT sleep, poll, proactively check on its progress, or duplicate this subagent's work.

 description:
- … Do not sleep, poll for progress, or ask the child for status.
+ … Do not sleep, poll, or proactively check on the child's progress.

I also dropped "avoid working with the same files or topics it is using". "Topics" is vague enough to stop the parent from doing related work like reviewing what the child changes, and the next line ("Work on non-overlapping tasks") already covers stepping on the child's files. The "you'll receive a <subagent> message" and "call subagent again with this sessionID" lines are unchanged, so the model still knows results arrive on their own and that messaging the child is allowed.

The mini TUI demo's copy of the launch text is updated to match.

Checked: bun test test/tool-subagent.test.ts (15 pass) and bun typecheck in packages/core.

[opencodejr]

The wording change looks good: it's consistent across the tool description, the background result and the mini demo copy, and no copies of the old phrasing are left. The prompt still says that <subagent> messages arrive on their own and that the child can be messaged with its sessionID, so dropping the "topics" clause doesn't remove any guidance the model needs. tool-subagent.test.ts passes and tsc type-checks packages/core.

Public API is unchanged: only model-facing prompt text changes.