Repository navigation
mobile: worker replacement closes a provider-leased session and drops its device claim #792
Description
Problem
A DeviceProvider cannot hold its device for the whole run on a local agent-device host. A worker replacement drops the device claim in the middle of the run.
agent-device ties host-wide device claims to a session that has an app open. A provider can take the claim by opening the slot's session (e2e-<target>-<slot>) in acquire, and the worker then resumes that session. But:
e2eretires a worker after every failing unit (run/scheduler.js:if (worker.sawFailure) this.retire(worker)).@e2e-dev/mobile'sAgentDeviceSurface.dispose()callsclient.sessions.close()on the slot session.- That close releases the device claim. Until the replacement worker reopens the session, another agent-device session can take the device or install over it.
The provider can't bridge the gap. It cannot hold a second claim-holding session, because that would block the engine's own session. agent-device 0.21.18 also has no lease that is independent of a session.
Versions
[email protected], @e2e-dev/[email protected], [email protected], iOS simulator.
Ask
When a slot is bound to a provider lease, have worker disposal keep the session (and its claim) open, for example by closing the app instead of the session, and leave closing the session to finish/release. Alternatively, add a provider hook that runs around worker replacement.
Context
A per-checkout simulator provider that installs a workspace-specific Release build under the claim: https://github.com/nelearning/chat-app/pull/3012 (e2e/native/checkoutSimulator.ts).
Activity
kvnloo commentedon Oct 3, 2026
I think the useful invariant here is that worker lifetime is an implementation detail; externally leased resource lifetime belongs to the slot/run that owns the lease.
If a failing unit retires a worker, disposal should be allowed to tear down the worker client/app without releasing the provider-backed device claim. The replacement worker should resume the same slot lease, and only provider
release/run finish should end ownership.A regression test that intentionally fails one unit, forces worker replacement, then proves (1) the replacement resumes the same lease/session and (2) a competing session cannot claim the device during the gap would pin the root cause much more strongly than a provider-specific workaround. Does that match where you want the provider/engine lifecycle boundary to live?
okwasniewski commentedon Oct 19, 2026
Fixed in #1209: when a slot holds a provider lease, worker replacement now keeps the session and its device claim, and only
release/ run finish closes it. It's on main and in the nightly (npm i @e2e-dev/mobile@nightly e2e@nightly), and ships in the next stable release.