livechat

You are entering live voice review mode. The user is speaking to you instead of typing.

STEP 0 — announce the session BEFORE doing anything else

Your very first output in this session must be exactly this line, as plain text:

Listening — go ahead.

This is non-negotiable:

• Do NOT call any MCP tool before printing that line.
• Do NOT skip it, paraphrase it inside a tool result, or hide it inside reasoning.
• The user has no other indicator that the mic is hot. If you call get_voice_input first, the mic is already capturing audio but the user is staring at a blank screen wondering whether it's working. They will start a conversation that they can't tell you've heard.

So: emit the literal line Listening — go ahead. first, then proceed to STEP 1.

STEP 1 — clear stale state once, before the loop

Right after the announcement and BEFORE the first get_voice_input, call the reset_voice_session MCP tool exactly once. The MCP server is long-lived and keeps state across slash commands, so a previous /endlivechat in this same process leaves a shutdown flag set; without resetting, your first get_voice_input call would return __END_SESSION__ immediately and the loop would never start. On a freshly-started server with nothing to reset, the call is a safe no-op and returns "active; nothing to reset" — ignore the message and keep going either way.

Do NOT call reset_voice_session again inside the loop.

The loop

1. Call the get_voice_input MCP tool to receive the user's next utterance.

2. Echo the transcription back to the user before doing anything else. The user has no other way to see what Whisper heard — wrong transcriptions otherwise look like Claude misunderstanding. Format it as a markdown blockquote on its own line, exactly like:

   > "<the exact text returned by the tool>"

   Then on the next line, proceed with your response or action.

3. After echoing, treat the utterance as a normal user prompt and act on it — answer questions, edit files, run commands, whatever the utterance asks for.

4. After completing each action, immediately call get_voice_input again. Do not ask "anything else?" Do not wait for typed input. The loop is the entire point.

5. If the tool returns multiple utterances joined with /, echo the full joined string in the blockquote, then treat them as related sequential requests from the same speaker.

6. If the tool returns the literal string __NO_INPUT__, the long-poll timed out with no speech. Do not echo it — it is not user speech. Call the tool again immediately.

7. If the tool returns the literal string __END_SESSION__, the user has ended the session. Do not echo it. Stop calling the tool, briefly summarize what was accomplished during the session, and stop.

8. If the tool returns a string starting with __ALREADY_RUNNING__: (e.g. __ALREADY_RUNNING__:34291), another livechat session is currently active in a different window (PID shown after the colon). Do not echo it. Tell the user briefly: "Another voice session is active (PID X). Take over from this window?" and wait for a typed yes/no — do NOT call get_voice_input again until the user replies (their mic is not yours yet). If the user agrees, call take_over_voice_session. On OK, print the literal line Listening — go ahead. again before doing anything else (the takeover gives no other indicator that the mic is now hot in this window), then immediately call get_voice_input to begin the loop. On any other return, surface the error and stop.