dsh-openai-live
GPT-Live-1 full-duplex voice for DeepSeek Harness.
Status: pre-release. Verified against the live API, not against documentation. 168 tests, a per-file 100% coverage gate that is enforced rather than reported, and a built-artifact smoke that runs the published output under plain
node. Not yet published to npm.
Why this exists
DeepSeek Harness ships no full-duplex speech subsystem at all, and no plugin in the ecosystem
implements gpt-live-1 — the only live, non-sunsetting full-duplex voice model. The nearest existing
plugin targets the retiring gpt-realtime family at a different endpoint with a different event
vocabulary. This is that port.
The interesting part is not the audio. It is that the model can delegate: mid-conversation it raises a task for the host agent and waits to be answered, without stopping the conversation. Getting that envelope right — and failing closed when it cannot be answered — is the domain of this project.
The seam, not a monolith
Three packages, mirroring how DSH itself models ctx.llm:
| Package | Role |
|---|---|
dsh-realtime |
The seam. Registers the realtime service; owns providers, routes, and session lifecycle. Deliberately refuses to own turn-taking, endpointing, or any auto-answer path for a delegation. |
dsh-realtime-openai |
The GPT-Live-1 adapter. Wire types, transport, translation, and the session — four separable layers, with the translation layer pure and tested against recorded frames. |
dsh-realtime-replay |
The keyless backend. Plays a recorded session through the real adapter and session, so conformance runs in CI with no credential, no network, and no spend. |
A fourth package, dsh-openai-live (this repo's root), is a bundle: it carries the
cordis.patch.yml that mounts the two rows above and depends on all three.
Install
dsh plugin --profile <name> add dsh-openai-live # once published
dsh --profile <name> --dump-config # composes?
dsh --profile <name> web
Then supply the credential from the environment — it is a validated config field, never a file this code reads:
export OPENAI_LIVE_API_KEY=…
A profile composes fine without one. The route registers, and opening a session fails with
MISSING_CREDENTIAL — an error that names the setting and never the value. That is deliberate: an
unconfigured machine should boot and then say precisely what is missing, rather than break the harness
at startup.
Properties held to, and tested
- Fail-closed delegation. An unresolvable or erroring delegation never auto-approves.
- The provider owns the turn boundary. There is no hand-rolled VAD, and the seam refuses
end-of-utterance calls outright —
session.input_audio.commitdoes not exist in the API, which the server itself confirmed by enumerating its vocabulary. - Credentials are a tested invariant, not a convention. The credential error names the setting. No value reaches a log, a fixture, a commit, or a published tarball.
- Keyless replay. The full conversation path is exercised with no API key present.
- Drift canary. A scheduled check fails loudly when the provider changes shape, so the first signal is a red build rather than a user bug report.
- Appends resolve on the provider's acknowledgement, correlated on the echoed
client_event_idwith a FIFO fallback — an undocumented echo must degrade to ordering, never to a hung promise. - Usage is cumulative, so a consumer replaces its figure rather than adding to it. The same value
legitimately arrives twice, on
session.usage.updatedand again onsession.closed. - Every registry proves disposal — a contribution is made from a child fiber, that fiber is disposed, and the contribution is observed gone.
Verified against the live API
The docs/ directory is the primary record. Everything in it was measured, and the traps each cost a
run:
| Document | What it holds |
|---|---|
| docs/protocol.md | GPT-Live-1 over WebSocket: the full client and server vocabulary, the delegation envelope, audio formats, and the traps |
| docs/design.md | architecture, scope, session lifecycle |
| docs/decisions.md | ADRs — why client delegation, why a seam, why this repo is non-governed |
| docs/w1-entitlement-probe.md | entitlement, and why a zero balance is not an entitlement failure |
| docs/w1-delegation-envelope.md | the two delegation modes, measured |
Working on it
pnpm install
pnpm gate # build → typecheck (src AND tests) → coverage gate → built-artifact smoke
pnpm canary # live drift check; self-skips without OPENAI_LIVE_API_KEY
The gate enforces per-file 100% statements, branches, functions and lines. That is enforced with
perFile: true, and the enforcement was proven to bite: one injected uncovered branch exits 1 naming
the file and each axis. Beyond unit tests there are two tiers a change cannot skip — a real
composition test that boots a cordis.yml through the actual Loader, and a built-artifact
smoke that runs the published lib/ under plain node.
See CONTRIBUTING.md.
Security
No credential is stored in this repository. .gitignore was committed before the first tracked
file, and the full history is secret-scanned in CI — a secret removed in a later commit is still
published. Report a vulnerability per SECURITY.md.
Licence
MIT — see LICENSE.
No comments yet. Be the first to write one.