Repository navigation
v0.7.2
Upgrading from 0.7.1 or older
- Start ompweb again after the in-app update. The updater of 0.6.0, 0.7.0 and 0.7.1 installs the new version and reports success, but an
ompwebyou started yourself (no systemd, launchd or tray supervisor) does not come back: the old launcher exits with the server. If the page does not return, runompwebagain. The cause is in the old version's own updater, which this release cannot change. - Then update each host with Force. A 0.7.x host daemon sharing the installation is shown in Settings → Hosts after the restart; its normal update is refused while the package was replaced under it, so use Force update to move it to this release.
Added
- Update omp on a host without a new ompweb release. Settings → Hosts shows "Update omp to …" for a host whose omp is older than the newest published release of the oh-my-pi fork whose runner speaks this ompweb's runner protocol (
RUNNER_PROTOCOL; a release that speaks another one is never offered). The daemon downloads the runner into a fresh staging directory, checks its SHA-256, its--runner-infoprotocol and--smoke-testthere, and only then moves it into place and points new sessions at it; any failure leaves the runner in use and its pointer as they were, and running sessions are not interrupted. A caller'soperationIdnames one omp update: a replay returns the stored operation and installs nothing again, and a conflicting request, or a replay of an attempt a later one superseded, is refused (operation_conflict). An omp update keeps its own record and never touches an ompweb update's: the two never run together, and each is refused with a machine-readable blocker (ompweb_update_active,omp_install_active) while the other runs. - Race-free GitHub watch registration.
watch_githubreturns the metadata snapshot used as a new watch's baseline, so agents can inspect existing comments, reviews, threads and CI before relying on later watch updates. - Clearly attributed automated messages. Cross-session agent-tool messages and GitHub watcher updates now render as source-labelled cards instead of looking like messages typed by the user. Attribution is server-authored, survives queues, retries and reloads, and links to the source session or GitHub item.
Changed
- The omp runner is built and released by the oh-my-pi fork.
runner/and its workflows are gone from this repository: the fork'somlbuild compilesompweb-ompfor every platform and attaches it to its releases withompweb-omp.sha256andompweb-omp.json(the runner protocol). Rootomp-runner.jsonpins the fork release an ompweb build fetches its runner from (ompweb fetch-omp, and in the background when it is missing), SHA-256-checked; source checkouts, CI, smoke tests and binary builds download it (scripts/fetch-runner.ts,npm run fetch-omp) instead of building it, and releases and prereleases record the newest published fork release with a matching protocol.OMP_WEB_OMP_BINstill overrides it. A source checkout needsnpm run fetch-omponce per pin, not a runner build.
Fixed
- In-app updates can replace an npm-global Windows install started by the tray. The update stops the tray program in the package and any node or bun process still using that install, without ending the update itself, and runs the install from outside the package so Windows can rename it. When the tray started the server, it is started again in the logged-on session and starts the one server; otherwise the server is started directly and still told which install to serve.
- Windows updates no longer fail when the update status is read while it is written. Host update state, the server worker's status and upgrade plans retry their atomic replacement for a moment when a concurrent reader holds the file, as the server's own update state already did.
- A host update's failed candidate is stopped on Windows even when it got the old daemon's pid. Windows can reuse a freed pid, so the restarted candidate could receive the pid the update had inspected; the stop then refused the plan as inconsistent ("Invalid upgrade process plan: current instance does not match the inspected process"), left the candidate running and required recovery instead of restoring the previous release. A stop checks the instance it acts on against the live process, so a recycled pid is still refused.
- GitHub workflow watches no longer regress to stale run states. Workflow updates retain GitHub's run timestamp and ignore older snapshots for the same or an earlier attempt, so a delayed
queuedresponse cannot follow a newerin_progressdelivery. - Keep sidebar transcript results aligned with expandable hits on every linked host. Each daemon now searches its own active conversations, ignoring abandoned branches and pre-compaction messages.
- Restore HTTP LAN and VPN access. Browser UUID generation now works without secure-context-only
crypto.randomUUID(), preventing the "Something went wrong" crash when opening the app or creating a session. - GUI server updates work for owned binary releases again. Settings → System prepares a server-and-host compatibility plan before approval, pins the release artifacts and checks service stop/start permissions before incompatible hosts move. Incompatible hosts must have a verified migration or recovery path; the local host participates in the initial migration and a shared-installation cutover. Progress survives a server restart, and failed replacement retains the previous installation for rollback. Uncertain shutdowns retain their update claim and report recovery rather than starting a second update.
- Old supervised hosts have an explicit forced recovery path. Settings → Hosts offers Force update when recovery is available. Confirmation warns that sessions and subagents may end. A fixed, integrity-checked bootstrap uses the historical runner channel without enabling old application traffic. Windows uses native PowerShell/.NET and process/service/task controls; Linux and macOS retain their own supervisor semantics. Force never bypasses release integrity, ownership, process identity, update disablement or service permissions.
- Production starts its missing local host automatically. It prepares the loopback link, reuses any answering daemon and otherwise starts an independent host. Concurrent starts keep one daemon, and a web-server exit leaves it running. Development does not take over production's host.
- Coordinated upgrades stage every participant before committing. A later host's download or preflight failure leaves earlier hosts on their existing release. Approval inventory is rechecked before host commits and server cutover. Native supervisor discovery prevents slow managed startup from being replaced by an unmanaged daemon, and macOS system LaunchDaemons retain their actual supervisor identity.
- Exited foreground hosts no longer reserve recycled process IDs during an update. A confirmed exit retires only that instance; unreadable liveness remains an error, and uncertainty after a stop signal retains the recovery fence.
- Package servers supervised by a shell restart loop stop their server child before replacing the launcher. Child discovery waits for the same launcher identity to reach a confirmed stopped state, including when it was still waiting to spawn. Unconfirmed freezes retain the recovery fence. The loop restarts the correct launcher, the old server releases its port, and independent host daemons keep running.
- Approved host updates keep their selected release and operation ownership. A newer published release does not block pinned admission. Commit decisions become visible only when complete, and active prepared or committing workers retain their update claims past the preparation deadline. A worker that dies during commit retains its recovery claim and blocks new work.
- Interrupted server upgrades remain bound to the approved server. Recovery verifies the installation, restart command and listener, supervisor and process ownership before claiming an attempt. Another installation or source checkout sharing the user’s update directory cannot resume or fail it. Older attempts without that binding are not resumed automatically.
- Unknown legacy host activity never grants Force permission. A shared local daemon can be interrupted only when the approved plan explicitly authorizes Force; otherwise newly busy hosts refuse cutover.
- Windows update status publication tolerates brief file-sharing conflicts. Atomic replacement retries sharing-related errors for at most one second without deleting the prior state. Persistent failures still stop the update and report the error.
- Local-host startup ignores unrelated services. Installation paths match at directory boundaries, and daemon detection distinguishes the host subcommand from directory names or server argument values.
- Open live subagent transcripts without waiting forever. Transcript pages now read the subagent's on-disk artifact directly instead of waiting for its busy agent process to answer first.
- Select reasoning effort for fallback models. Settings → Models → Native OMP Retry & Fallback now stores each fallback as
provider/model:effort, while leaving the effort unset uses the model default.
Upgrade
npm install -g @kahme247/ompweb@0.7.2Full changelog: v0.7.1...v0.7.2
omp: ompweb-dev/oh-my-pi v18.22.0-int.202611172157.ga645f02