Related: #430. This change gives a room member its configured Local VM; it does not create a room-owned desktop or a room Computer panel.
The pre-existing direct-turn stale lease after a missing terminal event is tracked separately in #860. That follow-up needs distinct direct invocation identities and tests protecting a replacement turn from an old watchdog callback; it is outside this room slice.
runGroupMemberTurn now mounts the speaking bot's computer MCP and prompt after
claiming the same target-scoped lease used by direct turns. Setup cancellation
and terminal cleanup release the lease and setup ownership. Timed-out or stalled
providers keep the desktop through the existing interrupt grace period; the
fallback then removes lease bookkeeping before releasing the bot. Computer
capabilities are bound to the exact VM claim and stop working on expiry or
member handoff.
Prepare the managed desktop image on an explicitly selected Podman test machine.
Build the renderer (node node_modules/vite/bin/vite.js build), set
OMB_VERIFY_PODMAN to the absolute executable path and OMB_VERIFY_MACHINE to that connection,
then run:
node --experimental-strip-types scripts/verify-group-vm.tsThe launcher creates a temporary home, data directory, explicit loopback server
and fake Claude engine. It grants that fixture access to the selected engine,
creates only fixture-specific desktop targets and removes them afterward.
Do not substitute a live application URL or data directory. The JSON receipt
records the exact MCP target for each speaker, the computer-off case, and
revocation of each settled speaker's computer capability. It is written to
.omb-scratch/verification-logs/group-vm-routing.json only after cleanup succeeds.
The script removes the previous dump before every send and requires every Goal,
including computer-off, to settle. Cleanup errors fail the run.
2026-09-07 on Windows/WSL2, rootless Podman 5.8.3:
- Two Goal speakers received different matching desktop targets and settled.
- A speaker switched to
computer: offreceived no computer MCP. - Each settled speaker's computer-control capability returned HTTP 401.
- Six regression scenarios run a real isolated server and fake Claude, replacing only the container boundary: failed readiness, Stop during readiness, shared member handoff/impersonation, lease expiry, timeout, and stall cleanup.
- Before the follow-up fix, the same readiness, expiry, and timeout regressions
failed on
a9ec061c; after the fix they pass. - After integrating upstream browser engine step 2 (
b3c2f74b): Goal, browser, lease, and fixture-launcher coverage passed 68 tests with 1 existing TODO. The dedicated real-server browser-engine room-mount test also passed. - Server TypeScript checking and server bundling passed.
This fake-engine fixture proves routing, not model-driven clicks or screen streaming. Firefox sandbox compatibility and Japanese guest fonts are separate changes. Full repository coverage is provided by PR CI; this local recipe is targeted.
Direct-turn regressions additionally send two different requests through a Local VM-pinned conversation while the bot default is Cloud, checking the actual MCP descriptor, prompt, matching preview surface, and capability expiry. A channel member's explicit This computer and Cloud (Boat or VPS) destinations mount exactly as its bot thread mounts them: the same checks, the same control gate, and the same consent boundary for waking or creating a cloud computer. The fixture proves the host descriptor reaches the speaker behind a working gate that closes when the turn settles, that a missing CUA driver is reported instead of dispatching without the promised tools, and that a Cloud speaker wakes its own archived Boat, runs there, and gives it back for the next turn. The VPS mount shares the bot-thread code path but has no channel fixture yet. Channels still have no conversation pin and no Auto fallback to the host.
The same isolated server now covers chat-driven computer selection: the agents tool discovers ready, startable, and provisionable destinations without mutating them. Selecting one ends the old turn, immediately blocks its previous computer and browser capabilities, and resumes the original request with fresh tools and one user-history entry. Auto prefers a ready destination over creating a cloud computer. The fake Boat boundary verifies wake, create, reuse, and a computer disappearing between discovery and dispatch; no real paid computer is created. Stop, provider failure, Off, a new queued request, and rejected in-flight manual surface changes have regression coverage. The VPS fixture separately proves starting a stopped container and creating a missing one only after selection.
One limit remains explicit: a stopped Local VM is not destructively rebuilt to make selection succeed. Every engine keeps its own model on Cloud, with the Boat as its computer tools; no engine runs a turn on Boat's own agent.
Run the regression coverage without a container engine:
node node_modules/vitest/vitest.mjs run server/group-local-vm.e2e.test.ts server/vps-routing.test.ts server/local-vm-lease.test.ts server/group-goal-run.test.ts server/group-goal-run.e2e.test.ts server/group-goal-wait-cap.e2e.test.ts server/control-omb.test.tsThe test-only Node loader in server/testing/group-local-vm-hooks.mjs replaces
container status and controls lease expiry and watchdog/deadline timing inside
that child process.
Production launchers never import it. No provider or live desktop is used by
these regression tests; the opt-in Podman acceptance above covers real routing.
The restart regression creates a per-bot VM through the fixture's lifecycle boundary, preserves a workspace file while removing the container, and restarts the server in that same disposable home. Auto recovers the prior VM once capacity is free, preserves its file, and neither probes nor provisions a bot that never had a VM. The occupied-cap case remains unprovisioned.
The pool regressions start two cold seats concurrently while one seat's
inspection is blocked, then show a third turn waiting and reusing a released
seat. They repeat with a full existing per-bot inventory: the per-bot cap
still rejects its next desktop, while switching to pool mode permits its own
configured seats. maxInstances limits the selected mode, not the total
containers retained across mode changes. Both scenarios use the isolated
container boundary above and require the actual fake-engine MCP descriptors
to match the assigned seats; no real desktop or model is contacted.