Escalation triggers

The coordinator owns every phase by default. Escalation is a bounded subagent call to the strongest available model for one specific judgment call — not a handoff of the phase. The coordinator frames the question, gets the answer, and stays the one accountable for the result.

Escalate when one of these is actually true, not on a general feeling that the task is hard:

• Root cause can't be pinned with confidence after actually reading the code — not after re-reading the report again.
• The change is architectural: it crosses a module boundary, touches a public API/interface, or introduces a cross-cutting abstraction.
• The code is security-, auth-, crypto-, payments-, or data-migration-sensitive.
• The operation is irreversible or high-blast-radius (schema migration, deletion, force-push, production config).
• An implementer has reported a spec gap or contradiction more than once on the same task.
• Reviewing a diff leaves the coordinator unsure whether the fix is correct or merely plausible.
• The user explicitly asked for a second opinion or an adversarial review.

None of these fire on routine work. Most tasks should complete without ever calling Escalation tier — the coordinator is capable enough to do the analysis, review, and verification itself, and the strongest model gets paid for only on the calls that actually need it.

How to escalate

Frame the specific question, not the whole task. Hand over the pinned context needed to answer it — the relevant code, the hypothesis so far, why it's uncertain — get the answer, and resume the phase. Escalation never becomes the new owner of the task, and it never decides scope; it answers the question it was asked and returns control to the coordinator.