Tensor Dynamics Watch assistant pilot 0.2.4
Test date: 2026-11-03

Result: suitable for a guided, read-only desktop pilot.
The full 'install and configure everything' journey is NOT ready.

Passed
- 53 automated plugin checks, including OS/GPU inventory fixtures, theoretical
  TOPS arithmetic, capacity bounds and performance false-pass regressions.
- Independent ZIP/installer checks and retry-setting preservation.
- Eight read-only tools are available, including hardware, TOPS, current-workload
  performance observation and capacity planning.
- The guide reader covers 31 public customer guides and Event API pages.
- Removed developer topics are rejected before any documentation request.
- Settings, 360 viewing and database guides are retrieved without credentials.

Platform packaging in 0.2.4
- Portable macOS, Windows PowerShell and Linux/POSIX launchers are included.
- Windows npm-shim/path resolution has isolated tests; no command shell interpolation.
- Windows and Linux full-machine installation remains unverified.

Earlier integration baseline, plugin 0.2.1
- 7 API contract checks.
- Real synthetic RTSP transport decoded at 320 x 180 in the 0.2.1 integration
  baseline. Version 0.2.3 narrows the public guide catalog and aligns pairing
  guidance with customer builds; performance logic is unchanged.
- Local signed webhook delivery, event/delivery identity, API-key read scopes,
  denied configuration access, and revoked-key rejection.

Test scope
RTSP/API tests used isolated synthetic input and a deterministic analysis fixture.
The test harness configured those resources. The current plugin did not do so.
No customer camera, customer footage, or real AI detection accuracy was tested.
Complete ChatGPT/Claude desktop click-through and clean-machine OS approval
remain unverified.

What the plugin can do
Check a running Watch 0.39.0, read current guides, list active camera status,
read requested recent event summaries, inventory the connector-host hardware,
calculate theoretical TOPS, observe current per-camera metrics and plan capacity.

Capacity/performance limits
No benchmark means no camera recommendation. A qualifying short observation
reserves headroom and is capped at the camera count already observed. Incomplete
timing coverage, stale telemetry or per-camera failures cannot pass. This is not
maximum-capacity, thermal-endurance, physical camera-to-alert latency or accuracy
certification. Hardware TOPS remains unknown unless independently established;
the calculator only evaluates supplied MAC-unit/clock/precision arithmetic.

What still needs implementation
Automatic Watch installation/launch, camera enrollment, API-key creation,
settings changes, and deployment of a customer's integration.
The Mac Watch pilot also needs signing/notarization work for broad distribution.

The portable ZIP provides macOS, Windows and Linux launchers with Node 20+ and Codex CLI.
Mac installation is exercised. Windows path handling is unit-tested; full Windows/Linux
Watch/client installation remains unverified.
Claude Desktop supplies the runtime for its separate MCPB download.

No public ChatGPT directory listing has been published.
Installation instructions: https://watch.tensordynamics.org/assistant/watch-support/INSTALL.txt