fix(server): release agent-device sessions when device_close runs#21478

Approvability

Verdict: Needs Human Approval

Macroscope's review found this PR requires human approval — This focused fix changes device_close from local state removal into an authenticated remote cleanup workflow with retries, failure retention, SSH handling, lifecycle locking, and Android serial retargeting. The cross-file runtime behavior and remote side effects are substantial enough to require human review.

^{You can add or adjust custom eligibility rules. Learn more.}

Review in Change Stack →

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

_{Comment @coderabbitai help to get the list of available commands.}

I don't see how this'll get approved if mine didn't https://github.com/pingdotgg/t3code/pull/18726

thanks @only21mil. the release-before-removal ordering, retryable records, and per-device eas endpoints are preserved, along with your commits. i merged current main (8e1bb0376ce8) into this branch without rebasing or force-pushing.

i simplified shutdown so callers own session removal and discovery, removing the readyOverride flag's extra meanings. i dropped the redundant mcp android-targeting test because main already implements that fix; the service test still verifies release targets the post-boot serial. i also updated response validation for main's new effect typecheck rule. the takeover removes 68 net lines.

verification:

• reproduced main's bug with the focused regression: expected the release rpc, received none.
• vp run check:changed passed through blacksmith-testbox: formatting, lint, both knip passes, graphql generation, server/scripts typechecks, and 78 tests across seven existing files.
• the final helper accepted SESSION_NOT_FOUND over authenticated http from a real agent-device 0.21.31 daemon.
• all 27 ci checks passed on 12879b09adf2eba02fe6feab47f422e463d903f6.
• two independent read-only reviews approved the simplified cleanup. the subsequent effect validation syntax change also passed the checks above.

unverified: live ios/android claim handoff, ssh/eas sessions, and device-panel retry screenshots. this vm has no simulator or connected device; the original screenshots were not recaptured. no client code changed.

the existing bot findings are covered: local endpoints survive hub restart failure; close preserves peer records; ssh reconnect/start handles claims without starting a daemon during close; reopen ordering uses queue/deferred and publication assertions.

@only21mil, a daemon that dies while its local hub survives can still leave a cached endpoint. that existing daemon-recovery issue is separate from claim release and has focused work in #23441; it is not expanded here. maintainer approval of the remote cleanup behavior is still needed.

Verified in a real dev environment

verification is incomplete. the changed claim-release behavior is unverifiable on this vm, not verified by the green checks.

• inspected the diff, description, discussion and commit authors. the pr is open, with no commits by @juliusmarminge or @t3dotgg. exact head: 12879b09adf2eba02fe6feab47f422e463d903f6.
• created ../main at github/main (570d4b03ee) and ran vp i. started main in the background with isolated worktree state. used repo playwright and /usr/bin/google-chrome to add a real git project and send real codex gpt-6.1-sol turns.
• installed jdk 17, android sdk/emulator 37.2.12 and an android 35 x86_64 image. kvm acceleration checked successfully. created the real pr21478 avd. the app discovered it; codex invoked device_list and device_open through t3 mcp. opening timed out with "could not communicate with device support." software graphics, lower resolution, 1536/2048 mib ram, and a restart did not get android through boot. qemu logged hanging main/cpu threads and exited. no claim was acquired, so main's original claim leak was not reproduced.
• checked out the exact pr head, ran vp i, and started its real dev server in the background with explicit isolated .t3 state (web 5733, server 13773). added its project through the app. a fresh codex gpt-6.1-sol thread successfully invoked device_list: local ready, android available, pr21478 stopped, no open records. device_close({shutdown:false}) with no open device succeeded; the server trace confirms DeviceService.close succeeded. discovery and the empty-close path work, but neither proves release of a live claim.
• ran the six changed device test files through blacksmith-check: 76 tests passed. all 27 ci checks are green at this unchanged head. no source changes or pushes.

short runtime transcript

main's failed device-open attempt, not a reproduction of the claim leak:

main device-open blocker

pr head: real provider discovery and empty close:

pr head discovery and empty close

recording of that neighboring provider/client flow, not claim handoff:

codex discovery and empty close recording

Not verified

live claim conflict before close and handoff to another thread after close; release-failure record retention and retry; active-claim shutdown; local hub restart recovery; ssh reconnect/cleanup; eas live sessions. android never completed boot on this vm, so these scenarios could not be exercised. ios needs macos/xcode; native mobile clients and desktop shells were not exercised. uploaded video playback and seeking were checked in chrome; t3's inline media player was not checked.

no merge, close, rebase, force-push, labels, or other prs touched. task dev servers and browser are stopped after evidence capture.

Verified on Android emulator (API 35) against a real dev server

verified head 12879b09adf2eba02fe6feab47f422e463d903f6 on the t3 android 35 x86_64 emulator with kvm. ran vp i, a real isolated dev server (13773), metro (8081), and the shared-cache native dev client. paired the native app and created a real project and two threads. both ran real codex gpt-6.1-sol turns through the local proxy. no mocks or demo data.

• a acquired session t3-0ea41a91aa7f6ada15421ae7 by opening the app through the returned agent-device command.
• b attempted the same app with its own session t3-8cc47649d17786c1fc77cd6d and received DEVICE_IN_USE naming a. the conflict is visible in the native client.
• a invoked the real mcp device_close with shutdown:false. it succeeded, removed only a's panel record, and left android booted (sys.boot_completed=1). no manual agent-device close was used for this handoff.
• b retried with its original session, without force, stealing, changing session, or restarting anything. app open and snapshot -i both succeeded. the native client shows the result.
• repeated close of a succeeded and preserved b's record. b then invoked mcp device_close(shutdown:true) with its active claim: success, no panel records left, emulator stopped.

76 focused tests across the six changed device test files passed through blacksmith-check. all 27 ci checks remain successful on the unchanged head. no source changes or pushes were needed.

before close, b cannot claim the device:

android native client shows b blocked by a's device claim

after a's server close, b opens and snapshots with its own session:

android native client shows successful claim handoff after device_close

recording of the native conversation during the close and retry:

android native client recording of device claim conflict and successful retry

Not verified

ios requires macos/xcode and cannot run here. ssh/eas hosts, unavailable-host retry, release-failure retention, hub-restart recovery, concurrent close/reopen, and desktop/web client controls were not exercised live. this pass did not reproduce the original leak on github/main; the before capture shows the expected active-owner conflict on the pr head. an initial android snapshot-helper parse failure was transient; b's later snapshot succeeded. uploaded images loaded at full resolution and the 58-second video played to completion and sought successfully in t3 preview. git.cafe renders these external media embeds as links, so inline media rendering there is unverified.

no merge, close, rebase, force-push, labels, or other prs touched. verification processes are stopped after capture.

merged main 812c0a71c785 into the existing contributor branch without rebasing. head is f62735c649d3153e2c9da998eaa39c6f824db286.

preserved retryable daemon release and main's launcher cleanup before power-off. shutdown releases peer claims before removing their records; a failed release or shutdown keeps the records for retry. preserved main's android eas routing and device-location tests. the existing cancellation test now waits through the state write lock rather than rediscovery, which would wait on the lifecycle lock held by close.

vp run check:changed passed through blacksmith: format, lint, both knip checks, graphql generation, affected typechecks, and 97 tests across eight existing files. two independent read-only reviews approved the final sha. no new test files or client changes. ci is running.

live android/ios/ssh/eas interactions were not rerun for this merge. the earlier android handoff evidence remains in this discussion; it covers the prior head, not this merge. no merge, close, rebase, force-push, or label changes performed.

final head f62735c649d3153e2c9da998eaa39c6f824db286 is mergeable with no conflicts. git.cafe reports 29 successful contexts, zero pending, zero failing; github confirms all executed checks passed (conditional preview/native checks and the paused review bot are skipped). no unresolved inline findings remain.

blacksmith check:changed passed, including 97 tests. both independent reviews approved this final head. no live device interactions were rerun after the merge; the prior android evidence remains above. ready for maria's review. no pr merge performed.