voice input: ask takes a WAV, the model transcribes it

An ask with audio sends a transcribing state, has the loaded model
transcribe the clip (thinking off, no tools, the prompt naming Italian
or English), then answers the transcript as a typed question, so the
user event, history and search hold the words. A stop while
transcribing ends the turn like any reply.

Refused: a relative path, a file that is not a WAV, over 60 s (the
ubatch bound, untried for audio), a model without audio input, and an
empty transcript. The length comes from the file size, since a
recording stopped by a signal leaves the header's length unset.

Live: 10/10 clips heard right, the right tool each time.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

1 files changed, 8 insertions, 1 deletions

diff --git a/AGENTS.md b/AGENTS.md
index f0f433b..357e04e 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -148,7 +148,14 @@ no `data:` prefix), no whisper. The 10 clips in
   20/20. ~70 more prompt tokens per clip. **Decision: transcribe first,
   then ask the transcript as text.** History and search need the words
   anyway.
-- Not tried: a clip long enough to cross the 2048 ubatch (~80 s).
+- Not tried: a clip long enough to cross the 2048 ubatch (~80 s), so
+  the daemon refuses clips over 60 s.
+- **Through the daemon** (`ask` with `audio`, a second instance on a
+  scratch socket): 10/10 heard right, `user` event 0.36-0.87 s after the
+  ask, and the right tool every time (weather, updates, mail, calendar,
+  now; none for the time in Italian and "Ciao"). "What's the weather
+  tomorrow?" was answered in Italian: the reply after the weather tool,
+  not the transcript, which was English.
 
 ### Attachments, live check (2026-10-16)