The ultimate harness for coding agents and beyond. All your agents. All your machines. One command center. Start with code, then follow your curiosity and build across disciplines: CAD, circuits, robots, games and music.
Validation — 2026-09-23
Source changes only. No commit/push, daemon restart, live CLI installation, robot deployment, Blender installation, paid model task or airplane render was performed.
Verified:
- TypeScript typecheck: passed.
- CLI build (
npm --prefix cli run build): passed. - Device suite: 117 tests passed in 8 files (includes the added restart/readiness regression).
- Broader Store/install/update/catalog/materialization, project folder, creation receipt, backend DSH and device regression run: 234 passed in 16 files, before the one additional device restart/readiness test. These counts overlap; do not add them.
- Final schema/fixture tests: 2 passed after checking malformed requestId error responses.
- Generated request/response JSON schemas match runtime Zod validators. Request defaults are optional inputs; invalid/missing requestId can be echoed/omitted in error responses.
git diff --check: passed.
The first broader run used /bin/sh and one pre-existing materialization test saw an extra
logout line from that interactive shell. Re-running with /bin/zsh and an empty temporary
ZDOTDIR passed the full selected regression suite. Product shell behavior was not changed.
Runtime adapter tests use real temporary setup/doctor scripts and workspace materialization;
only their shell selection is isolated from the operator's profile. Engine creation is mocked.
Other device tests use isolated temporary durable journals/creation receipts and mock package
or engine boundaries. Transport tests verify role, encryption and hello gating at the existing
relay seam; this is not a physical Lamp acceptance run.
Reproduce the broader selection from cli/ with an empty temporary ZDOTDIR:
SHELL=/bin/zsh ZDOTDIR=/path/to/empty-test-zdot npx vitest run \
src/lib/autonomous-device \
src/lib/agentCreationReceipt.spec.ts src/lib/projectFolder.spec.ts \
src/dsh/catalog.spec.ts src/dsh/install.spec.ts src/dsh/materialize.spec.ts \
src/dsh/service.spec.ts src/dsh/update.spec.ts src/backendSocket.dsh.spec.ts \
--maxWorkers=1
npm run typecheck
npx tsx scripts/device-store-contract.ts --check
npm run build
Next joint acceptance: install the new CLI through the machine's authorized lifecycle helper, verify existing LAN connectivity, confirm all four capabilities in a real Lamp hello, then run discover → inspect → prepare/poll → separate turn.send against Blender. Record actual engine login/permission prompts and generated artifacts. OS must not report a rendered airplane from this validation or from preparation reaching ready.
Desktop reveal and first-turn readiness follow-up
- Reproduced the reported Blender block on a real running agent: its terminal displayed the Claude
input prompt, the registry had
active=true,launch.state=ready, andsessionId=""; preparation remainedrunning/launch. No test prompt was sent to force a native session into existence. - Runtime readiness now reuses confirmed launch state, with explicit
engineAuthentication=unknown. - Follow-up CLI device/local WebSocket selection: 135 tests passed in 9 files; TypeScript typecheck, CLI build and generated schema check passed.
- Desktop dial/preparation selection: 21 tests passed. Targeted Flutter analysis passed.
- UI tests use a fake machine/terminal and assert concurrent opens, existing-tab reuse and one viewer; durable delivery tests use temporary journals and simulated disconnect/restart/acknowledgement. They do not establish a physical robot-to-Desktop end-to-end acceptance run or a rendered airplane.
- This follow-up changes readiness semantics, but does not change OS JSON schemas or capability names. OS integrators must read the handoff update in both language versions of the contract.