kj-aiml/heli-harness/verify-premise

verify-premise

Trigger: "fix X", "Y is broken", review comment claims, bug reports, or requests based on asserted behavior.

Scope:

• Restate the claim.
• Read cited files and nearby code.
• Reproduce, trace, or disprove the behavior.
• Decide whether the premise is confirmed.
• Record the observed boundary separately from the interpretation in the active diagnosis state.

Outcomes:

• confirmed as stated
• confirmed but broader
• confirmed differently
• not reproducible or premise broken

Rules:

• Do not implement if the premise is false or unclear.
• Do not silently switch scope.
• If confirmed but unexplained, route to debug.
• If confirmed and shared surface is affected, route to impact before editing.
• Manual smoke / UX "fails": open existing e2e/Playwright/smoke contracts first. If the manual path disagrees with the contract (wrong entry URL, wrong CTA), treat it as agent/operator error until the product path is proven broken.
• Authorized ops request: before mutating, map the request to a supported control-plane action. If no action exists (e.g. "rebind pin" but service only has pin-on-create), stop with a typed blocker such as ACTION_UNSUPPORTED — do not invent DB bypasses.
• Do not write product-defect claims into plan Evidence until this skill confirms them.
• A premise check does not establish root cause by itself; record supporting/contradicting evidence and a falsifier, then route to debug when confirmed but unexplained.