Skip to content

The ultimate harness for coding agents and beyond. All your agents. All your machines. One command center. Start with code, then follow your curiosity and build across disciplines: CAD, circuits, robots, games and music.

4.3 KB
Raw Blame

Harness manager validation

Verified September 22, 2026. Harnesses · Needs input · Minimize keyframe.

All process tests used newly created Claude and Codex harnesses with temporary profiles, synthetic conversation history, and private tmux servers. No existing user harness was paused, stopped, or restarted. Native engines used a local test provider; the checks did not require paid model requests.

Check Result
Native Mac app and CLI builds Passed
Flutter analysis and CLI type checking Passed
Desktop manager, lifecycle, navigation, shortcuts, and terminal regressions 131 passed, including the rendered preview
Lifecycle suite 200 passed; 100% statements, branches, functions, and lines in the five covered lifecycle modules
Related CLI transport, activity timestamps, process, and terminal regressions 188 passed
Native AppKit titlebar checks 472 passed with the File-menu PR
Native Claude and Codex lifecycle fixtures Passed, including three immediate pause/resume cycles per engine
Native Mac UI acceptance Passed, including three manager pause/resume cycles per engine, repeated clicks, background resume, pane reopening, and original history
Final visual render Passed

The suites overlap; these counts are not a combined total. The optional visual render is enabled in the reported desktop run. It is skipped in ordinary runs.

The final integration with the other recent PRs passed the full desktop suite (2,821 passed, 11 skipped), analysis of all changed Dart files, and a fresh macOS build. A separate render run passed all 23 manager checks. This pass caught and fixed delayed input handoff during minimization and clipped keyboard selection in long lists. The combined CLI run passed 3,998 tests with 63 skipped; one timing-sensitive offline hook test failed under load, then its complete file passed all 31 checks in isolation. CLI type checking also passed.

Verified safeguards include confirmed process termination before reporting pause, exact-pane closure that preserves neighboring panes, conversation identity across registry reloads, duplicate-operation prevention, retained rows after refresh failure, and actionable native errors. A late hook may rebuild a registry record without falsely appearing to replace its process; changed process, conversation, or terminal identities still reject the operation.

Pane minimization uses a bounded GPU snapshot and a curved mesh. Closing releases keyboard input immediately; only the captured pixels continue moving, without resizing or stopping the departing process. Keyframes were rendered and the animation was exercised in the native preview. Keyboard navigation, machine search, activity sorting and compact ages, current and replaced questions, separate accessible open/control actions, small windows, larger text, and Reduce Motion are covered. The clock stays on the context line; only a waiting harness adds a help action beside pause/play. The former bell and its shortcut now open Needs input in this same manager. A numbered badge appears at the icon’s top right only for pending questions; large counts use 99+.

Pause/resume controls currently support saved Claude and Codex conversations on both local and connected remote machines. Commands go to the owning machine's daemon; shared machines are view-only. Other engines remain viewable with a tooltip explaining that exact resume is unavailable. The built app and CLI were not installed over or used to restart the user's running app or daemon.

Key repeatable checks are npm run test:resume and npm run test:resume-native from cli, and flutter test test/harness_session_manager_test.dart test/agent_pause_state_test.dart from desktop. The native Mac test is desktop/integration_test/native_resume_e2e_test.dart; its header documents the private fixture setup.

The native integration test lets macOS own accessibility activation instead of retaining a test-only semantics handle across window deactivation. This avoids a Flutter engine bridge reset receiving a partial tree as its first update; ordinary widget tests still verify VoiceOver actions with semantics enabled. The native suite completed in 26 seconds after this correction. Fixture cleanup rescans for shell history files written during shutdown.