feat: speak answers sentence by sentence and never speak reasoning

Segment the final answer as it streams and synthesize each completed
sentence immediately, so a listener hears the reply begin while the model
is still writing it. Sentences share the progress clips' FIFO chain and
output-sequence counter; only text left unspoken (a hit budget) is
synthesized after settle, so audio is never sent twice. A message that
grows a tool call stops answer-streaming without re-speaking what already
played.

Reasoning models inline their chain of thought as <think> blocks in the
same delta stream. Strip it while streaming -- carrying a partial tag
across delta boundaries -- so it is neither spoken nor shown as answer
text.

The fixed "Hang on while I work on that." is now a rotation of short
natural phrases that never repeats twice running, and a long silent think
gets a spoken "still on it" roughly every 45 seconds until real speech
begins. Turn timeout rises to five minutes to match the client.

Co-Authored-By: Claude Fable 5 <[email protected]>

docs/voice-api.md

@@ -93,7 +93,13 @@ Azure push-stream recognition emits `transcript.partial` and `transcript.final`
 
 ### Turn output and lifecycle
 
-For either input form, the server emits `agent.working`, then `agent.accepted` after Pi accepted the prompt. If work is still in progress after two seconds, it sends one `agent.progress` event with a short spoken status clip (for example, “Hang on while I work on that.”). It never repeats this spoken update within the same turn; ordinary screen-only `agent.working` heartbeats continue every 15 seconds. While it waits for Pi's native `agent.settled` event (not merely `agent.end`), it repeats the existing `agent.working` frame every 15 seconds; those progress frames stop before `assistant.final` and never overlap synthesized audio. The wait has a two-minute turn timeout. `assistant.delta` is streamed while Pi answers; `assistant.final` contains only assistant text, never STT transcript.
+For either input form, the server emits `agent.working`, then `agent.accepted` after Pi accepted the prompt. If work is still in progress after two seconds and nothing has been spoken yet (see below), it sends one `agent.progress` event with a short spoken status clip, picked pseudo-randomly from a small set of natural filler phrases (“One sec.”, “Let me take a look.”, “Hmm, checking now.”, “Working on it.”, “Give me a moment.”, “Let me look into that.”). The server never repeats the same phrase on two consecutive turns. Ordinary screen-only `agent.working` heartbeats continue every 15 seconds regardless.
+
+During a long turn, Pi may emit several assistant messages between tool calls before its final answer — each one ends with Pi's `message.end` event. Whenever a completed message includes a tool call (i.e., it is not the turn's final answer), the server also speaks it: it sends `{"type":"agent.progress","text":"<message text>","audio":true}` — `text` is always the message's full, untruncated text — immediately followed by that clip's binary audio frames, unless every word of that text was already spoken live as answer sentences (see below), in which case the clip is skipped entirely, or only its unspoken remainder is synthesized if part of it was spoken live and part was not. `agent.progress` may therefore occur multiple times per turn, always in the order the messages completed, and each is optionally followed by its own clip's binary frames. For speech only, a message longer than about 300 characters is truncated at a sentence boundary where possible (the JSON `text` field is never truncated). Progress clips are sent one at a time — a clip's audio always finishes sending before the next `agent.progress` frame is sent — and a turn's total spoken progress narration is capped at roughly 60 seconds of audio or 8 clips, whichever comes first; the fallback phrase counts toward this same cap. Once the turn's answer has audibly begun (the first `answer.sentence` has been sent — see below), a brand-new message that had no live narration of its own is left quiet rather than starting fresh progress narration that would talk over the answer; a message that was already mid-narration when a tool call appeared in it still gets its own unspoken leftover finished so it isn't cut off abruptly. `agent.progress` frames, spoken or not, always stop before `assistant.final`.
+
+While the final answer streams in as Pi generates it, the server does not wait for it to fully finish before speaking it: as each complete sentence of the in-progress message accumulates (split on `.`/`!`/`?` — plus a closing quote or bracket — followed by whitespace; a decimal like “3.5” never triggers a false split, since the period is not followed by whitespace), the server sends `{"type":"answer.sentence","text":"<sentence>"}` immediately followed by that sentence's own binary audio frames, serialized through the same ordered chain as `agent.progress` clips so nothing interleaves or reorders on the single audio output. Markdown-decoration-only fragments (e.g. `---`, `` ``` ``, `**`) are skipped silently. Because a message might still turn out to include a tool call (making it narration, not the final answer), every message is optimistically streamed this way until either a tool call is detected in it or it ends; once a tool call appears, sentence-streaming for that message stops (already-spoken sentences remain spoken — that's fine acoustically) and any unspoken remainder is handled by the `agent.progress` mechanism above instead. A message's trailing fragment with no terminal punctuation yet is held back until the message ends, then flushed as one final `answer.sentence`. Spoken answer audio is capped at roughly 120 seconds or 40 sentences per turn, whichever comes first; once agent.settled fires, `assistant.final` (the full, untruncated answer text) is always sent next, and any part of the answer that was never spoken live because a cap was hit is synthesized once, as ordinary audio frames, between `assistant.final` and `audio.end` — a turn where every sentence streamed live sends zero audio frames in that window. `audio.end` always ends the turn's audio either way.
+
+While it waits for Pi's native `agent.settled` event (not merely `agent.end`), it repeats the existing `agent.working` frame every 15 seconds; those progress frames stop before `assistant.final` and never overlap synthesized audio. The wait has a two-minute turn timeout. `assistant.delta` is streamed while Pi answers; `assistant.final` contains only assistant text, never STT transcript.
 
 Azure synthesis output is signed little-endian 16-bit, 24 kHz, mono PCM. Every binary server frame has this eight-byte header followed by PCM: