| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
ask carries paths; the daemon reads them as the user. Text files and a
PDF's text go inline within half the free context, shared equally, and
are stored that way, so a reopened conversation still has them. Images,
and PDF pages with no text, go as image parts to a model the router
reports as seeing images, and are stored as a path and hash, resolved
when sent: a file changed or gone since is reported as no longer
available. turns gains a files column for the chips, added to an older
database on open. A refused file is one log line, not a traceback.
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
|
|
Every client sees every event, so a command error from one client, such as
opening a conversation another just deleted, reached a canvas watching a
reply and was taken as that reply's end: the canvas stopped rendering
tokens the daemon kept streaming. An error from the reply thread now
carries turn: true; command errors carry none.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
A startup probe that failed with anything but OSError (llama-server not in
router mode, an empty model list) crashed the daemon at boot, against its
own comment; it is caught like the per-request failures now. llama-server's
reasons reach the client: an HTTP error carries its body and an error chunk
mid-stream its message, where the client used to see 'choices'. A daemon
that refuses to start, or fails before serving, exits 1 instead of 0.
assistant.sh no longer mistakes a --test run for the daemon.
AGENTS.md said every section from Hardware to Kokoro was voice-only, which
included findings the daemon still applies; it now names the voice ones.
It also claimed a stopped reply emits nothing after done, while a late ctx
can. The spec lists every error message, the example config no longer
shows a non-default temperature as the default, and the updates tool stops
telling the model not to read lists aloud. The spec is marked implemented.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
Measured live: with reasoning on, a two-line question took 21 s of
thinking before its first token, against about 5 s for questions that
needed little. thinking = false in config.toml skips it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
From review of the protocol commit. An exception inside a command, such as
an id SQLite cannot bind, killed that client's reader and left a deaf
connection still receiving events; commands now run under a try that
answers error. Ids must be integers: delete with no id compared None to an
unsaved conversation's None and wiped it, and a string id deleted the row
through SQLite's affinity while the Python comparison missed, leaving the
current conversation pointing at nothing so every later answer failed to
save. ask, search and their arguments are type-checked instead of coerced.
serve() refuses to start over a socket that answers, since unlinking it
orphaned the live daemon's clients, and main() removes the socket only if
it served it. The accept loop logs and continues on errors instead of
dying silently, and a client line over a megabyte is dropped. The two
known ceilings, a stalled client delaying everyone once for 2 s and a
stopped reply holding the llama-server slot until its next chunk, are
marked. New checks cover each of these and a client connecting mid-reply.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
Each client gets a reader thread; commands are ask, stop, new, list, open,
search and delete. A reply streams think, token and tool events through
say(), which checks the turn counter under the same lock that stop takes,
so a stopped reply cannot send a token after its done. A finished turn is
saved; a stopped one is not. hello carries the current conversation and
any reply in flight, so a reconnecting canvas redraws both. Errors from
llama-server become an error event and the daemon stays up; every exit
path of main() logs a line, which the voice version's silent death lacked.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
|
|
From review of the storage commit. makedirs and connect leave an existing
directory or file at whatever mode it had, so a history dir made earlier
with the default umask stayed 0755; both are now chmodded. A NUL in a
search ended the FTS string and raised instead of returning nothing. The
index on turns(conv, seq) serves the seq lookup per save, load_conv and the
cascade, which otherwise scan every turn. The self-check now also asserts
that an OR query matches nothing, which a non-raising but unquoted query
would not.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
One file under XDG_DATA_HOME, 0700 directory and 0600 file. turns.messages
holds what the model is sent for the turn, user/reply/tools what the canvas
draws. FTS5 with remove_diacritics finds città from citta. The query is
quoted as one phrase so FTS syntax typed by the user is matched as text,
and the self-check's integrity-check proves a conversation delete cascades
into the index.
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
|
|
The assistant becomes a text daemon behind the quickshell canvas. STT,
TTS, the mic button loop, the presentation gate, the audio levels and the
overlay config table go; the socket, the tool runner, the never-swap model
rule and the config merge stay. The reply loop goes with the voice and
returns with the text protocol. The voice version is tagged voice-v1, so
later voice work starts from the tag instead of from dormant code.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
It was the only child started without PR_SET_PDEATHSIG. The assistant
died with no log line and left arecord recording into the unlinked raw
file for ~76 minutes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
History kept only what was said, so after a few factual answers the
model saw no tool behind them and invented the next one: asked for
pending updates it made up counts without calling the tool (1/10 on a
replayed session). Keeping each turn's tool calls brings it back to
10/10. Their results are replaced by a placeholder, which scores the
same and keeps the context close to what it was; a follow-up on an old
result calls the tool again. The window is now the last 10 whole turns.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
Whisper wrote "Che ori sono?" twice in one day and the model once
answered with the weather. On 10 recorded short questions beam search
(beam_size=5) gets 9 right against 8 for greedy, for 0.06 s more; an
initial prompt also got 9 but lost one question entirely, so it is not
used.
The default system prompt now says the user's words come from speech
recognition and should be read by how they sound: on "Che ori sono?"
stray tool calls went from 1 in 10 to none, other tools unaffected.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
With the clock only in the system prompt, "what time is it?" asked again
an hour later copied the old answer from history. The current question
now carries the date and time too; history stays unstamped, because
stamping every message made the model drop tool arguments (the finance
account filter went from 6/6 to 0/6).
Gemma garbles numbers it spells out in Italian ("trent settler"), 3 in
10 time answers. The default system prompt asks for digits, which
Kokoro reads correctly: 20/20 clean.
temperature in config.toml is sent with every request; unset keeps the
llama-server preset's, which other clients share.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
speaking comes from the reply thread, not the speaker: the tool
announcement would otherwise replace the tool state while the tool is
still running. arecord gets a 33 ms period, because it writes its file
in periods and the default left the user's waves at 8 levels a second;
the context size is read at startup so the first hello carries it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
Asks llama-server for usage in the stream, so the log shows how full
the model's context is; history, tool definitions and tool results all
count.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
- Game controllers report battery through their kernel driver, not
BlueZ: bluetooth.sh falls back to the power supply named after the
device address.
- Asked for a VM's IP after a listing without one, the model made one up
without calling a tool. The default system prompt now says facts about
the user's computer, devices, accounts, mail, calendar or money come
only from a tool, or it says it does not know.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
The status registry's effective presentation is the OR of the manual
flag and status.gaming, which the shell's game detector writes;
statusctl presentation get reads only the manual half and printed 0
with a game running. presenting() now reads both files.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
Every press bumps a turn counter and kills the running aplay; queued
audio from an older turn is dropped, and reply() now runs on its own
thread and returns once its turn is stale, so the button loop is never
blocked. A cut-off exchange is left out of history so user and
assistant messages still alternate.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
- mail printed threads and the model called them emails; it now
reports messages and threads.
- finance spending takes a since date (default the first of the month)
instead of a day count, and an optional asset account. The
per-account listing needs an end date or the API rejects it; an
unknown account name gets the user's accounts back.
- Tool errors are capped at 2000 characters like normal output.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
- The model has no clock and invented a 2025 start date for the
calendar; the system prompt now carries the current date and time.
- It also derived a wrong weekday from a date; agenda.py prints the
weekday itself.
- An announcement ran straight into the answer; a 0.6 s pause follows it.
- The calendar days hint asks for 30 or more when looking for the next
event.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
kokoro-onnx is now packaged (python3-kokoro-onnx, without
espeakng-loader) and finds the system espeak-ng with its default
config, so the explicit EspeakConfig and the .venv setup go away.
Tested live: two Italian questions answered and spoken correctly.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
|
The microphone's mute button is the push-to-talk: the ALSA Mic Capture
Switch starts and stops a recording. whisper-server (large-v3-turbo, Vulkan
on the Arc) transcribes it, the model already resident in llama-server
answers, and Kokoro speaks the reply sentence by sentence.
Only Italian and English are spoken, so detection is restricted to those
two; the voice follows the language of the reply, not of the question.
Presses are ignored while presentation mode is on, so calls keep the mic.
AGENTS.md records the backend, model and voice comparisons behind these
choices. Licensed GPLv2 only.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
