agent-device sharding with no --device fails: the front rewrites --device to the UDID

What happened

The slow-lane case in e2e/slow-agent-device.test.ts (iOS, "with the device limit raised to 2 > with both devices booted, agent A holding one and agent B the other: A's test --shard-all 1 with no --device passes, and A's test --shard-all 2 with no --device fails with test sharding requires 2 devices, but only 1 matched") fails on origin/main (3de46d1). It fails on the --shard-all 1 step:

INVALID_ARGS: Session "…" is already bound to apple device "simlock-<uuid>" (<UDID>), but this request selected --device=<UDID>

Command: scripts/slow-e2e.sh e2e/slow-agent-device.test.ts -t "shard-all 1 with no --device passes" --bail=1, with agent-device 0.21.28. Log on the slow-lane Mac: /var/folders/tb/j39z273s5kn7lb9tz1dhk8l00000gn/T/simlock-slow-e2e-20261023-234115.log.

Expected

--shard-all 1 with no --device runs on the agent's one leased device.

Likely cause (not confirmed)

ruleSharding in src/agent-device/front/request-rules.ts:256-261 rewrites --device to the device's name when the open reply carried one, and to its address (UDID) otherwise. replyFacts in src/agent-device/front/allocate.ts:419 keeps the name only when agent-device's open reply has data.device.name. In this run the rewrite used the UDID, and agent-device refuses a UDID against a session it bound by name.

Found during the #957 delivery (PR #967), which does not touch this path.

Written by an agent.

feat(core): a standalone worker serves agent-device on localhost by d…

Reproduction

Confirmed, without devices. Branch bug/971-repro, src/agent-device/edge.test.ts, "rule 7, sharding": "test --shard-all with no --device names each of the token's leased devices by the name agent-device's inventory gives it when the allocate reply named no device, for an agent holding one device and one holding two".

pnpm exec vitest run --project unit src/agent-device/edge.test.ts -t "inventory gives it"

AssertionError: expected 'fake-ios-1-addr-1' to be 'simlock-aaaa'
 ❯ src/agent-device/edge.test.ts:3994  expect(flagsOf(one)["device"]).toBe("simlock-aaaa");

Root cause

Defect. The issue's guess is right, and it hits every run. replyFacts (src/agent-device/front/allocate.ts:419) takes the name from data.device.name. agent-device's allocate reply never has that field, so name is always undefined. ruleSharding (src/agent-device/front/request-rules.ts:257-260) then falls back to the UDID. agent-device checks --device against the session's device by name only, so it refuses the UDID. The unit tests missed it because FakeAgentDevice (src/agent-device/testing.ts:524-527) adds device: { id, name } to every allocate reply.

# agent-device v0.21.28 (main is the same)
packages/daemon/src/sessions/handlers/lease.ts   data: { lease: allocated, ...(providerData ? { provider } : {}) }
dist/src/sdk-device.js (installed 0.21.28)       function Y(e,t){return X(e.name)===X(t)}   // --device vs bound session: name only
packages/replay-port/.../session-test-shard-devices.ts   device.id === selector || name match  // shard lookup takes either

Simplest fix

When an allocate reply names no device, the front calls agent-device devices once on that lease (its udid/serial) and stores the entry's name in the front map. The fake's default allocate reply drops device, as agent-device does.

Alternatives rejected

Send --udid/--serial instead of --device: the shard lookup reads only --device, so with no list it picks any booted device, including another agent's.
Have simlock's driver report the simulator name: that copies agent-device's naming, and Android emulators get no name this way.
Ask agent-device to put the device in its allocate reply: this needs an upstream release, and the bug stays until then.

Risk

The existing test "names a device … by its id when … the reply named none" (edge.test.ts:3962) pins the bug and must change.
Allocate gains one devices call; through a gateway it must reach the lease's worker.
ruleDevice/ruleShardList (namesDevice, request-rules.ts:176) also refuse an agent that names its own device by its agent-device name. The same fix covers that.

Written by an agent.

Decision

Maintainer: fix it on the simlock side now. Once per lease, the agent-device front reads agent-device's device list and takes the leased device's name, so ruleSharding rewrites --device to the name (simlock-<uuid>, unique per device) instead of the UDID. The test fake's allocate reply (testing.ts:524-527) is fixed to match the real one.

Later, an agent-device change: accept the bound device's UDID as --device (tracked in agent-device).

Written by an agent.

agent-device side: callstack/agent-device#3945.

test(agent-device): the front reads a lease's device list once, even …

fix(agent-device): a named --device on any command reads the leased d…

fix(agent-device): sharding names a leased device by agent-device's d…

Delivery run

Merged: #986 (#971)
Parked: none
Waiting: none
Decided: #971 — when to read the device list → lazily, once per lease, the first time a request rule needs the name (Assumption in PR #986)

Written by an agent.

Navigation Menu

Repository navigation

agent-device sharding with no --device fails: the front rewrites --device to the UDID #971

What happened

Expected

Likely cause (not confirmed)

Reproduction

Root cause

Simplest fix

Alternatives rejected

Risk

Decision

Delivery run

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

agent-device sharding with no --device fails: the front rewrites --device to the UDID #971

Description

What happened

Expected

Likely cause (not confirmed)

Activity

V3RON commented on Oct 24, 2026

Reproduction

Root cause

Simplest fix

Alternatives rejected

Risk

V3RON commented on Oct 25, 2026

Decision

V3RON commented on Oct 25, 2026

V3RON commented on Oct 25, 2026

Delivery run

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions