Repository navigation
Verify native development and recovery end to end in Ace #55
Description
Ace already has native coding tools, lanes, shell/Git/gh access, diffs, durable recovery, and release commands. Qualify the complete development loop so dogfooding MVP is based on working behavior rather than a feature checklist.
Acceptance
- Complete a representative Ace change from issue and lane through edits, required real checks, Ace Diff review, PR/CI, and signed canary delivery using Ace's surfaces and native harness.
- Verify the UI change with channel-owned browser and native tools; preserve the evidence. Follow up on any failing CI or review feedback inside Ace.
- Continue existing work through compaction, worker/helper restart, and a development failure without lost history, unintended side-effect replay, or duplicate lane writers.
- Keep a working installed Ace separate from the source build under development. Validate any stored-format change against an existing-channel backup; never reset the store.
- Record the channel/lane, revision, build/model, outcomes, and any remaining external-tool escape. File observed defects as issues.
This validates existing mechanisms; it does not request another release platform. Shell git, gh, builds, and existing release commands inside Ace count. Human OS/login handoffs are legitimate. #39 owns published-upgrade qualification on another Mac; external-harness continuation (#9), host migration (#49), and mobile are separate promises.
Dogfooding MVP: #5. Depends on the required native/browser verification paths. Review evidence.
Activity
iamnbutler commentedon Oct 6, 2026
Development-loop observation from #148: a React crash left Ace-dev blank. Cmd+R delivered but did not reload (the app has no native reload action); background Cmd+Q was refused as unverifiable by desktop_key. Stopping the exact source dev runner via its SIGTERM lifecycle and restarting it recovered the app, without touching Canary or other dev instances.
iamnbutler commentedon Oct 15, 2026
Context: the two-Mac study (Oct 14–15) covered parts of this loop on a small sample repo (toddler-game-center), not on Ace itself. Scenario 08 went from issue to lane to PR with the Diff reviewed from both hosts. Scenario 14 survived a worker kill, a Helper restart, and a peer leaving the tailnet without lost or duplicated work, and it found #283, #285 and #287. ChatGPT computer use drove those runs, not Ace's own tools, and they didn't include an Ace change through CI and a signed canary. The full acceptance here is still open. The dev-app reload gap from the earlier comment is #97.