Problem

The subagents tool currently only exposes action='list'. When a child run goes ghost (parent waiting on a child that will never complete), there is no in-tool way to terminate it. The only recovery path is a full gateway restart, which is destructive — it kills every healthy subagent and session along with the stuck one.

Requested actions

action='kill'

Args: runId (or sessionKey, or taskName as a stable alias). Effects:

1. Terminate the underlying child process / session.
2. Emit a synthetic completion event to the parent so sessions_yield unparks.
3. Remove the run from the active list.

action='reap-stale'

Optional helper. Args: olderThanMs (default 600000). Effects:

1. Iterate active runs whose last activity is older than olderThanMs.
2. Apply 'kill' semantics to each.
3. Return a summary { reaped: [...], skipped: [...] }.

Motivation

Today (2026-05-27) two stuck subagent chains in the same workspace forced two gateway restarts (TSK-20260527-0006 and TSK-20260527-0010 v1). The restart disrupts other healthy work, breaks session continuity for unrelated channels, and is the wrong granularity of recovery.

Related: upstream issue for gateway-enforced runTimeoutSeconds (filed separately) would prevent most ghost cases; subagents kill is the manual escape hatch when prevention isn't enough.

Acceptance

• subagents action='kill' target=<runId|sessionKey|taskName> terminates the target and emits a completion event to its parent.
• subagents action='reap-stale' (optional) does the same in bulk.
• Tool schema documents both actions in the allowed values enum.