From the gpt-5.6-sol eval (P1, effort M-L, multi-session). AppDelegateShortcutRoutingTests (9k+ line class per the eval) resets global defaults and runs the main loop in every teardown (programaTests/AppDelegateShortcutRoutingTests.swift:44), has ~58 fixed 50ms waits, and creates real windows for many routing cases (:80). CI already isolates it, runs it serial, skips one test, and retries it as timing-sensitive. The routing BEHAVIOR is worth keeping; 87 headless AppKit integrations + retries are not.
First step: extract a pure ShortcutRoutingDecision over event/window/responder snapshots; port the layout/chord/precedence matrix to table-driven tests; keep ~6-10 real AppKit integration tests + 1-2 WindowServer UI tests for actual focus repair. This is the finish line for the 'faster CI' work.
From the gpt-5.6-sol eval (P1, effort M-L, multi-session).
AppDelegateShortcutRoutingTests(9k+ line class per the eval) resets global defaults and runs the main loop in every teardown (programaTests/AppDelegateShortcutRoutingTests.swift:44), has ~58 fixed 50ms waits, and creates real windows for many routing cases (:80). CI already isolates it, runs it serial, skips one test, and retries it as timing-sensitive. The routing BEHAVIOR is worth keeping; 87 headless AppKit integrations + retries are not.First step: extract a pure
ShortcutRoutingDecisionover event/window/responder snapshots; port the layout/chord/precedence matrix to table-driven tests; keep ~6-10 real AppKit integration tests + 1-2 WindowServer UI tests for actual focus repair. This is the finish line for the 'faster CI' work.