feat(agent-device): agent-device runs as an optional module on workers

[V3RON]

Closes #597

Status

Implement: done (15/15 green) Review: round 1 passed, 4 text fixes applied unreviewed Hardware: deferred to the end of the run Gate: waiting for the stack
Next: hardware check at the end of the run, then the stack gate

Done when

• Every Tests line passes, and every other test in the suites F11 names passes with only its setup changed and F12's assertions renamed.
• F7's two supervisor tests are deleted. No other agent-device test is deleted.
• docs/CONFIGURATION.md documents capabilities.agent-device.*. Its leftover-daemon line says it applies with capabilities.agent-device.enabled true.
• F14's config writes and deletes use capabilities["agent-device"]; scripts/slow-e2e.sh e2e/slow-agent-device.test.ts and scripts/slow-e2e.sh e2e/slow-agent-device-gateway.test.ts pass on a Mac with an iOS runtime.

Assumptions

• Assumption: the registry joins the existing daemon/http/agent-device import cycle — two entries added to .dependency-cruiser-known-cycles.json and its ratchet raised from 6 to 8, because the front keeps importing src/http (F19) and the cycle check works at folder scope.
• Assumption: a stop-order test in main.test.ts — with hooks running concurrently (ADR 0030), it asserts both stop steps ran before the drivers' disposal, not their order between modules.
• Assumption: module checks run on a gateway too, so a leaseTtlCapMs above lease.maxTtlMs still stops a gateway.

Review

Spec review: 1 blocking, 0 fixed, 0 notes. Code review: 2 blocking, 1 fixed, 0 notes. Claims review: 4 blocking, 4 fixed, 5 notes.
Rounds: 1. Tokens: 12.5M.

Fixed unreviewed: code: src/daemon/optional-modules.test.ts:793 test title; claims: src/agent-device/AGENTS.md:4, docs/internal/agent-rules/architecture.md:78, docs/internal/flows/startup.md:25; after the rebase fix check, docs/internal/flows/startup.md:25 build order.

Rejected:

• spec: the known-cycles ratchet rises from 6 to 8 against architecture rule 14 — the maintainer chose to keep it for now (option A); the cycles go when agent-device stops importing src/http
• code: same finding as the spec one — same reason

Written by an agent.

[V3RON]

Decision taken

The question: the spec keeps agent-device's front importing src/http. With the daemon importing the module registry, that adds two cycles to .dependency-cruiser-known-cycles.json, raising the limit from 6 to 8, and architecture rule 14 says that list only shrinks.

• A: keep the 8-entry limit for now. Users see nothing. This is taken: the maintainer chose to keep going.
• B: break the cycle now by moving what agent-device takes from src/http away from src/daemon. Users see nothing either; it is a larger refactor than the spec asks for.

Reverting means a follow-up that breaks the cycle and lowers the limit back to 6.

Review notes

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

• claims: src/agent-device/AGENTS.md:5 the "Uses" line omits src/modules and the core runtime validators that definition.ts and worker.ts import
• claims: src/daemon/AGENTS.md:85 onSocketClaimed lists agent-device as its own step; it now starts inside the module's afterSocketClaimed hook
• claims: src/agent-device/host/agent-device-supervisor.test.ts:344 the describe "off and starting" and the test "enabled, ..." have no off or enabled case left
• claims: docs/HTTP-API.md:1522 "This build has no optional module of its own" reads as no module registered, but the registry now lists agent-device

Written by an agent.

[V3RON]

Review notes

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

• claims: docs/internal/agent-rules/architecture.md:78 the agent-device lease TTL floor is said to be set in agent-device's config check, but it sits in validator (integerInRange) in src/agent-device/definition.ts:17-27; check only compares the cap to lease.maxTtlMs

Written by an agent.