[danielehrhardt]

Paused runs (on a usage limit, a budget or a question) could only be stopped with the Stop button. When a stand-in Mac hit its weekly Claude limit, 16 paused runs were set to resume days later and repeat outreach and posting, and no agent could stop them.

What changes

• New Godmode tool run_cancel({ runId, reason? }) for manager agents (same gate as run_continue): stops a queued, running or paused run for good, together with the work it delegated. A paused run never resumes by itself afterwards.
• Runs of chats that work on another computer of the group are stopped there over the link (/api/link/runs/:id/stop, falling back to the plain cancel for older computers). An offline computer gives a clear error and nothing is closed locally.
• Refuses the caller's own run and the runs that handed it its task, keeps the reveal/computer/VM guards of the other manager tools, and is audited (run.cancel).
• On a secondary it runs on the primary like runs_list.
• Orchestrator prompt, activity labels, chat tool meta and ARCHITECTURE.md mention it.

Tests

• packages/core/test/run-cancel.test.ts: paused and queued runs close, ended runs stay, own/parent runs refused, unknown runner refused, non-managers don't get it.
• Core and desktop typecheck pass. Full core suite: the only failures (maintenance, mods, updates, VM lifecycle) fail the same way on main without this change.