[V3RON]

Closes #477

Status

Implement: done Review: round 1 + one review of the hardware fix loop, passed (text fixes unreviewed) Hardware: pass (run 20261013-164634, 30/30) Gate: waiting for the stack
Slow lane: the maintainer allowed it for every PR of this run.
Next: gh stack sync, then merge-pr.sh 509 for stack #508

• On a Mac with an iOS runtime, Apple silicon, at least 7 GiB of RAM and SIMLOCK_E2E_AGENT_DEVICE set as the toolchain row says: scripts/slow-e2e.sh e2e/slow-agent-device.test.ts passes every test in the iOS describe. Checked: run 20261013-164634, 15/15.
• On a machine with emulator and adb, at least 12 GiB of RAM and SIMLOCK_E2E_AGENT_DEVICE set the same way: the same command passes every test in the Android describe. Checked: run 20261013-164634, 15/15.
• With SIMLOCK_E2E_AGENT_DEVICE unset, the same command writes <log>.exit 0 and shows both describes skipped. Checked: run 20261013-150830.
• With SIMLOCK_E2E_AGENT_DEVICE=/nonexistent, the same command fails, naming /nonexistent in the beforeAll failure. Checked: run 20261013-150900.

The real-device runs found five simlock defects, fixed here: the front refused the backend that connect proxy sends; it read agent-device's replies one level too high (result.data), which also affects #474's lease-end code; /health answered 503, so the CLI hid simlock's reason; a sharded test named devices by id, not name; upstream failures were not logged. The agent-device build also needs pnpm build:android.

Assumptions

• Assumption: release-all ordering check — a session sample is a session list CLI call of about a second, so "the first no-session sample comes before the wipe" would fail on correct behaviour; the test asserts no sample that saw the session started after the wipe began, plus the first and last samples, which fails the same wrong behaviour.
• Assumption: the running-command test allows 1 s between the CLI's exit and the wipe start — the CLI process exits a few ms after the daemon answers.
• Assumption: JSON shapes of agent-device (data.sessions, data.devices[].id) and simlock record fields (requesterId, deviceId) are read as the code shows them; the green stage confirms them on hardware.
• Assumption: each agent runs connect proxy --daemon-base-url <worker>/v1/agent-device once, with its own AGENT_DEVICE_STATE_DIR and token in env, and every command with --json; the output envelope {success,data,error} is read as agent-device's source shows it.
• Assumption: waitForAgentDevice polls GET /v1/agent-device/health and accepts status 200 with ok: true; the front relays the daemon's own /health and adds no upstream field in code, so the spec's upstream check cannot be made.
• Assumption: backend on allocate — accept only the backend connect proxy sends for the platform (ios-instance on iOS, android-instance on Android) and forward it unchanged; any other backend is still refused.

Review

Spec review: 2 blocking, 2 fixed, 1 notes. Code review: 4 blocking, 4 fixed, 3 notes. Claims review: 2 blocking, 2 fixed, 1 notes.
Rounds: 1. Tokens: 20.8M.

Fixed unreviewed: the round 1 test fixes, and the review of the fix loop's text fixes (run artifacts removed, .agent-device ignored, two tests, one comment). The fix loop's commits had no review per fix, at the maintainer's request; one review covered them all.

Rejected:

• code: e2e/helpers/env.ts restart() has no timeout of its own — duplicate of a spec finding, fixed under it
• spec: e2e/slow-agent-device.test.ts short sample gap stays green — not blocking: Assumption 1 accepts it
• code: e2e/slow-agent-device.test.ts session list exit code unchecked — not blocking: weakens no stated check
• claims: toolchain.md package.json names only a short hash — not blocking: still true via vendor/README.md
• code: e2e/helpers/env.ts:384 restartSteps keeps running after the 90 s limit — not blocking: slow-lane helper, teardown only
• code: src/agent-device/front/request-rules.ts:236 emulator name fallback may match another agent's emulator — not blocking: speculative

Written by an agent.

[V3RON]

Review notes

Not blocking, not verified. Each is one reviewer's claim.

• spec: e2e/slow-agent-device.test.ts:375 a session ending within one ~1.25 s sample period after the wipe begins stays green (Assumption 1)
• code: e2e/slow-agent-device.test.ts:359 no sample checks the session list exit code, so a failed sample counts as "no session"
• claims: docs/internal/agent-rules/toolchain.md:77 says rev-parse HEAD prints the commit in package.json's @agent-device/proxy dependency; package.json holds a short hash, the full commit is in vendor/README.md
• code: e2e/helpers/env.ts:384 when the 90 s restart limit wins, restartSteps keeps running; a late startDaemon can leave a daemon behind after teardown
• code: src/agent-device/front/request-rules.ts:236 if agent-device cannot read an emulator's AVD name it falls back to the shared model name, so that --device name could match another agent's emulator

Written by an agent.