@2004/realtime Voice Preview

What this skill does

Creates Starchild Live, the user's realtime voice proxy. It listens, responds, and acts as one natural assistant while using the user's Starchild context, files, memory, tools, and work threads as internal capabilities. OpenAI Realtime handles the live voice experience; ask_starchild remains an implementation detail, not something narrated to the user.

The skill bundles the complete project under template/. Its setup script creates or updates the user's own copy at output/projects/realtime-voice-preview/; it does not rely on this agent's pre-existing project. This integration uses OpenAI Realtime, not Grok.

Prerequisites

1. OPENAI_REALTIME_API_KEY securely configured, with Realtime model access and billing.
2. Starchild runtime at STARCHILD_RUNTIME_URL (default http://127.0.0.1:8000).
3. Node.js for parser tests.

First check whether OPENAI_REALTIME_API_KEY is already configured. If it is missing, call request_env_input for exactly that variable; never request or accept it in chat. The secure prompt's visible guidance must tell the user to:

• create a project key at https://platform.openai.com/api-keys;
• enable billing / add credits on the same OpenAI project;
• confirm that project has Realtime API/model access;
• prefer a project-scoped key without unrelated permissions.

Stop and wait after requesting the key. Do not tell the user to edit .env manually.

Workflow

1. Create the user's own project

python3 skills/realtime-voice-preview/scripts/setup.py

This copies the bundled template to output/projects/realtime-voice-preview/, preserves any existing .env, checks runtime health and runs the 59 parser + 16 background tests. Use --target PATH for another destination and --force to refresh bundled non-secret files.

2. Start and serve

Serve with the Preview tool using these exact values:

• title: OpenAI Realtime Voice Demo (keep constant; it determines Preview ID)
• dir: output/projects/realtime-voice-preview/src
• command: python3 server.py
• port: 8765

Then verify GET http://127.0.0.1:8765/health returns healthy. Tell the user to open the Preview in a separate, preferably full-screen browser tab, grant microphone permission, and test there rather than inside the embedded Preview iframe. Microphone permission is more reliable in a top-level tab.

3. Thread route

Agent Settings exposes one selector. Preserve this ordering:

1. temporary — Temporary — do not enter an existing thread. This is the default.
2. persistent — dedicated voice-only context.
3. thread:<id> — recent normal work threads, newest first.

The browser sends the selected route only when saving /bridge_config. The server maps it to the internal isolated | persistent | selected config. It must ignore any per-call thread_id sent to /agent_bridge.

Temporary uses is_temporary=True, so it does not read or write an existing work thread. A selected work-thread ID must be validated against /bridge_threads. Build that catalog from the authenticated Thread Service (/api/clawd/threads), not runtime /sessions: the Thread Service supplies user-facing titles. Render title first with message count/activity as optional context; keep the thread ID only as the internal option value. Filter temporary/voice/transport threads and cap the result set.

4. ask_starchild execution modes

The Realtime tool accepts question and optional execution_mode:

For HTTP 202 the client must:

1. send exactly one function_call_output acknowledging accepted run_id;
2. persist pending job metadata and a bounded delivered-ID set in localStorage;
3. poll GET /agent_jobs?run_id=... about every two seconds;
4. offer cancel through POST /agent_jobs/cancel;
5. recover active jobs after reload/reconnect with GET /agent_jobs;
6. if disconnected, retain a terminal result as pending;
7. when the data channel is open, inject the verified terminal result once, mark it delivered, then use the existing response/continuation guard to request speech.

Never send a second function-call output for a background completion. Completion is a new conversation message, not another output for the original call_id.

Verbatim relay. ask_starchild.question must carry the user's spoken words as transcribed — the Realtime model must not paraphrase, summarize, translate, expand, or guess intent. Realtime owns speech and turn-taking only; understanding belongs to the Agent. If a request is genuinely ambiguous, the model asks the user out loud to clarify rather than inventing a filled-in question. The tool description and system instructions both enforce this so semantic drift never reaches the Agent.

Temporary rolling context. Temporary mode has no persisted thread history, so elliptical follow-ups ("可以去做了吗", "那第二个呢") cannot be resolved server-side. The browser keeps a rolling transcript (last ~12 turns) and sends it as the optional context array with each /agent_bridge call. The server injects it — only when is_temporary is true — as a compact [Recent voice conversation] block (capped to the last 6 turns / 1200 chars) ahead of the verbatim spoken request. Selected/persistent threads already have real history, so context is deliberately skipped there to avoid duplication. context flows through both the synchronous and background-job paths.

Cancellation is cooperative. cancel_requested is active, not terminal; only report cancelled after the worker observes the flag. The server retains sanitized job state with TTL/cap bounds, allows one active job per effective thread, and reuses a recent identical job's run_id.

5. Verify before delivery

python3 -m py_compile output/projects/realtime-voice-preview/src/server.py
node output/projects/realtime-voice-preview/src/test_function_calls.js
python3 output/projects/realtime-voice-preview/src/test_background_jobs.py

Expected: 59 function-call tests and 16 background-job tests pass. Also verify Preview health and the live HTML title after restarting the service.

Preview ID stability (do not skip). The Preview tool derives the preview ID from the title slug. Keep the serve title constant across every restart so the URL never changes — product/brand renames belong in the HTML <title> and on-page copy only, never in the serve title. Update code by re-serving the same title on the same port; never stop-and-recreate under a new title, which silently mints a new ID and breaks the user's saved URL.

Endpoints

• POST /session, GET /token
• POST /agent_bridge
• GET /agent_jobs, GET /agent_jobs?run_id=...
• POST /agent_jobs/cancel
• GET|POST /bridge_config
• GET /bridge_threads?agent_id=...
• GET /health

Design and safety rules

• ask_starchild is the only capability gateway; Realtime must not invent Agent data.
• Long-lived OpenAI keys stay server-side; only ephemeral client secrets reach browser.
• Preserve function-call event correlation, call-ID dedupe and the response continuation guard.
• Serialize runtime calls per effective thread and retain the short completed-result cache.
• Prefix every relayed runtime user message with a stable [Starchild Live Voice Agent] source marker and label the spoken request; this must be visible in selected work-thread history.
• Never hold /agent_bridge past the Preview gateway window. Use a shorter synchronous budget and promote unfinished wait calls to the existing background-job lifecycle.
• Parse bridge responses as text before JSON so an upstream HTML error becomes a useful diagnostic rather than Unexpected token '<'.
• Base64-wrap SDP in JSON across the Preview proxy.
• UI strings remain English; no hardcoded machine-specific public URL.
• Public publishing is optional and must use the community-publish skill.

Full-screen usage (全屏使用)

Always instruct the user to test in a separate, full-screen browser tab, never inside the embedded Preview iframe:

• The voice UI needs microphone permission (getUserMedia). Browsers restrict or deny this inside embedded/sandboxed iframes, so the embedded Preview panel may silently fail to capture audio even when the page loads fine.
• Use the Preview panel's open-in-new-tab / pop-out action to get a top-level tab, then click Allow when the browser prompts for the microphone.
• Permission is bound to the origin (host + port). A stable Preview ID (constant serve title) means the user grants it once; a renamed title mints a new ID and re-prompts.
• Full-screen also keeps the Connect button and jobs panel unobstructed during testing.

Include this guidance every time you hand the Preview URL to a user.

Troubleshooting