dsh-plugin-trtc-conai

Turning a text-only agent into a voice conversation operator normally means hand-rolling room provisioning, UserSig signing, TC3-HMAC-SHA256 authentication, and task lifecycle management against a vendor pipeline. dsh-plugin-trtc-conai collapses all of that into seven structured agent tools built on Tencent RTC Conversational AI, so the model can mint a room, drop an AI robot in, steer the live spoken exchange mid-call, and wind down cleanly. Audio stays entirely inside the TRTC pipeline and never touches the Harness process.

The seven tools span every critical point of a call: credential issuance (locally signed, no billing), starting and stopping a session, querying task status across Idle / Preparing / InProgress / Stopped, hot-swapping voice, model, persona, welcome line, and interruption behaviour while a call is in flight, and two direct-control channels that bypass the LLM entirely. One makes the robot recite an exact line; the other feeds content into the LLM as though the user had spoken it. The billing boundary is explicit: with allowStartConversation set to false the agent retains every read-only and control tool but can never open a billable task, so it can observe and close out existing calls at zero cost.

This plugin is aimed at developers who run DeepSeek Harness text agents and want real-time voice interaction on top without rebuilding the Tencent Cloud integration from scratch. Secrets stay in the configuration layer and are asserted absent from the model-facing tool schemas. The signing stack is built on node:crypto instead of pulling in the vendor SDK, keeping the runtime dependency footprint to a single package. Twenty-nine offline tests verify UserSig and TC3-HMAC-SHA256 output byte-for-byte against the official implementations, catching the non-standard base64 alphabet and UTC-midnight credential-date traps that would otherwise produce tickets the backend silently rejects.