Repository navigation
OpenWiz 0.0.12
Pre-release0.0.12 · 2026-11-16
Scripts and agents can drive OpenWiz
What's new for you
- Run
openwiz help --agentto get every command, option and result format in one list. It reads nothing and changes nothing. - Add
--jsonto any command for one clean result, and a clear error when something goes wrong.
Live-checked in this release's smoke test, run by agents: installing this release, openwiz help --agent, ls --json and show --json, one send to a throwaway thread, and one new thread in a throwaway project. Everything else here is tested on synthetic fixtures only; the terminal app was never live-checked on a Mac.
Under the hood
- Every
--jsoncommand prints exactly one document, its result or a structured error, and the command list is a written contract. - Fleet candidate checks can return JSON too.
- A visual terminal app (
openwiz ui) was built in this version and is being retired.
0.0.11 · 2026-11-16
Make every decision in one place, and see the whole plan
What's new for you
- A Decisions page: answer grill questions, confirm specs and approve plans. Each answer goes to your orchestrator as one message.
- Anything live, release or publishing asks you twice before it goes.
- A Plan page shows each version's tickets by state, what holds up the most work, and what any ticket waits on.
Tested on synthetic fixtures only; a whole day of real decisions on these pages is still unverified.
Under the hood
- A plan API reports ticket states, dependencies and the critical chain, and marks reads that failed instead of hiding them.
- The planning skills now ask for decisions in a fixed, structured form.
- Each answer is sent with the question and revision it answers, and its text is kept exactly.
- Malformed decision requests are reported, not dropped.
0.0.10 · 2026-11-16
A cockpit for your phone and PC
What's new for you
- A Cockpit page: what needs you, oldest first, and every running agent by project and fleet.
- Answer an item right there when its exact proposal is known, after a preview and a confirm. Risky ones ask twice.
- A command bar to move around fast:
ls,cd,next,blockers, or an item's number. - Each item shows its question, a link, when it was asked and how risky it is.
Tested on synthetic fixtures only; the cockpit on a real phone and PC is still unverified.
Under the hood
- The local server answers the whole cockpit in one read-only request, and marks any part it couldn't read or that is stale, instead of showing it empty.
openwiz fleetshows the same details as the page.
0.0.9 · 2026-11-16
All your loops in one place
What's new for you
- A Loops page on your PC lists every scheduled job: next run, last run and cost over the last day, problems first.
- Switch any loop off or on. It records who did it and when, and stays that way after a restart.
- No 100-loop limit: tests plan 200 loops within the shared limits.
- OpenWiz tells you when a loop check is missed or stuck, and stays quiet until that changes.
Tested on synthetic fixtures only; a whole day of real loops run from the phone is still unverified.
Under the hood
- The service runs loop checks itself, and never retries an agent start whose result is unclear.
- Review and CI follow-ups go to the agent already on that work, through durable records; GitHub webhook events are routed once each.
- A loop runs only when its exact definition is approved and it has its own capacity limits; other service work carries on.
- Loop health is watched separately from the scheduler.
0.0.8 · 2026-11-16
OpenWiz keeps running on your PC
What's new for you
openwiz-serviceruns on your PC: it serves the web page, runs scheduled jobs and reports its own status.- It checks its own health and recovers its loops, fleet runs and the web server when they fail.
- It updates itself from tested builds, and goes back to the last good one if a new one won't start.
Tested on synthetic fixtures only; a whole day with the Mac shut is still unverified.
Under the hood
- A GitHub App gives the service tokens scoped to each repository, renewed and shared across its jobs; the private key stays in owner-only storage.
- The service runs the fleet tick on a schedule and records how each run ended.
- The tick picks builders by ticket label and reviewers from the other model family, with one fallback that keeps the branch when a Claude usage limit stops a builder.
- A service update can resend a record that never went out.
0.0.7 · 2026-11-16
Helpers for builders and reviewers, and a "where are we?" report
What's new for you
- Builders and reviewers can bring in helpers for side jobs, while the lead keeps every decision.
- Ask "where are we?" and get one screen of current work, open decisions and the next move.
- New skills to map an unclear project (wayfind), research with cited sources (research), and track down a hard bug before fixing it (diagnose).
Helper swarms ran for real and "where are we?" was used for real; whether helpers make work faster or cheaper is still unproven.
Under the hood
- A written guide sets how many helpers to use and which ones.
- A check catches a helper that pushes code; helpers never push.
- The skills check now covers the new skills.
0.0.6 · 2026-11-16
Fleets run inside OpenWiz
What's new for you
openwiz fleet tickmoves your fleets forward: it works out what is due and does it.openwiz fleet due --github-onlytells a loop whether there may be work for it.- See your fleets with
openwiz fleetsand on the web page's fleets screen.
Tested on synthetic fixtures only; a real fleet run by OpenWiz itself, followed on a phone, is still unverified.
Under the hood
- Fleet views refuse conflicting GitHub facts and show "unknown" instead of guessing progress or who is building.
- Larger fleet lists finish within the time budget: merge history is read in batches.
- Inventory commands report a failed write, and send previews refuse thread details with repeated keys.
- On a PC, OpenWiz finds the running CloudRoom app, and asks you to restart CloudRoom if an update leaves two copies running.
- Keyboard focus stays put while pages load and fleets refresh.