# PoC verification record

## Automated evidence

Run:

```sh
corepack pnpm install
corepack pnpm test
corepack pnpm run typecheck
corepack pnpm run build
npm pack --dry-run
```

The deterministic tests cover:

- exact Bearer authentication;
- strict streamed Chat Completions admission;
- text and OpenAI text-part parsing;
- request and prompt limits;
- SSE role, token, finish and `[DONE]` frames;
- explicit cancellation and reset;
- Harness message-to-turn correlation;
- reasoning and stale-turn suppression;
- barge-in ordering and late-chunk fencing;
- fresh session creation after reset;
- the empty tool allow-list applied at agent creation.
- configuration-route origin and custom-header checks;
- redacted credential status and write-only credential replacement;
- strict App ID, App Certificate, bridge-token and public HTTPS URL validation;
- dynamic missing-token rejection before model admission.

## Exact-artifact smoke

The delivery gate is the packed `.tgz`, not the source checkout:

1. Produce `dsh-shengwang-voice-bridge-0.2.0.tgz` with `npm pack`.
2. Inspect the file list; it must contain compiled `lib/`, type declarations, the Bundle patch, documentation, security policy and license. It must not contain tests, source files, environment files, credentials or local logs.
3. Set a temporary `DSH_HOME` and install the exact tarball through `dsh plugin --profile voice-poc add <absolute-tarball-path>`.
4. Inspect the composed config and confirm the `shengwang-voice-bridge` row resolves to the installed package and defaults to `enabled: false`.
5. Boot the profile and verify `/api/shengwang-voice/config` repeatedly returns redacted defaults.
6. In a real browser, open Settings → Plugins → Voice and verify both configuration cards, password inputs, invalid-URL blocking, saved-state refresh, console output and desktop/mobile overflow.
7. Save an ephemeral token and enable the bridge; verify the status changes to listening without restarting Harness, `/health` is stable, an unauthenticated chat request returns `401`, and shutdown is clean.

## Evidence boundaries

Passing the deterministic suite proves local protocol and lifecycle mechanisms. Packing/installing/booting proves the exact artifact composes in a clean compatible Harness profile. Neither proves:

- a live Shengwang cloud agent reached the endpoint;
- the account has Conversational AI entitlement;
- a live DeepSeek provider credential or model response works;
- microphone capture, ASR accuracy, TTS voice quality or end-to-end latency;
- safe production exposure to the public internet.

A genuine live E2E additionally requires user-supplied Shengwang credentials/entitlement, a publicly reachable HTTPS chat path and a configured Harness model route. Starting a browser RTC call and hearing a model reply is the acceptance event; record it separately from the deterministic and packed-browser evidence.