| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
A searched turn on a cloud reasoning model ended at the thinking: the
model emitted the tool call, but finish_reason: "tool_calls" landed on
the same SSE line as the include_usage block, so the single-event parser
returned that line as a usage chunk and the loop never saw the tool
finish. The parser now emits every event a line carries, so the search
fires.
Also in this change:
- Replay each round's reasoning_content on the assistant tool-call
message, which interleaved-thinking models require to keep going.
- Add a per-provider replay_reasoning option (DeepSeek, SiliconFlow
GLM-4.7+) to carry prior turns' reasoning_content when search is on.
- Add an on-demand diagnostic log gated by $LLAMACHAT_DEBUG_LOG.
- Record provider, usage_json and reported_cost_usd per reply, so a
searched turn keeps every round's billed usage for external consumers.
|
|
- Send max_tokens = 32768 for non-local providers so reasoning models have room for both chain-of-thought and answer.
- Add per-provider thinking_budget option for endpoints (currently SiliconFlow) that cap reasoning tokens separately.
- Accept bare JSON arrays from /v1/models; some OpenAI-compatible endpoints omit the {"data": [...]} envelope.
|
|
Metadata now resolves models.ini over provider defaults over presets.ini,
so a local model keeps getting its context and vision from the preset
while remaining overridable. A model whose vision support is unknown no
longer blocks an attachment: unknown is not the same as no, and the API
will reject an image if it really cannot take one.
Key resolution moves to the worker thread so a pass: key's pinentry
never freezes the GUI. KeyResolver gains a lock since StreamWorker and
TitleWorker can now overlap. Skipped providers surface in the status bar
rather than only on stderr. model_info uses dataclasses.replace to avoid
mutating the ModelInfo that resolve() may return from the store.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Three states stay distinct: blank for a free local model, ? for a cloud
model whose prices were never entered, and a figure when they were.
Collapsing the first two would make an unpriced cloud model read as free.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Field conversion is separated from the widget so the part with the edge
cases is testable without a running Qt application. Vision is tri-state
through a binary checkbox: an unchecked box with an unknown prefill stays
unknown rather than writing False, which would shadow a provider-level
vision=True through models.resolve's pick().
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
A provider that is unreachable or whose filter matched nothing is
reported rather than fatal: local models must stay usable when the
network is down. Clients are built on demand so a key is resolved only
when that provider is really used.
The model listing carries its own timeout, separate from the per-turn
one: the picker has to populate fast, and a dead cloud provider must
not hang it for minutes. Client.models() takes list_timeout (default 30)
and uses httpx.Timeout(list_timeout, connect=5), and MultiClient threads
its own list_timeout through to each listing call. The fan-out stays
serial on purpose: KeyResolver._cache is unlocked and only safe while
every caller resolves on the GUI thread.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
The local router needs none, so the header is omitted entirely rather
than sent empty.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Cost has to survive reopening a conversation, which means storing what
each reply used. The model goes on the message rather than the session
because sessions records only the current one, and a conversation that
switched models would otherwise be priced entirely at whichever is
selected now. All three columns are nullable so old rows read as unknown.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
A nan or inf in [providers.*] survived float() and reached the cost
arithmetic, rendering $nan or, for -inf, a negative cost. Worse than a
display wart: is_priced answered True, so the readout took the priced
branch and showed a poisoned figure in the one state the ? is there to
admit it cannot say.
Guarded at both ends. providers._number mirrors models._get, which
already covered the models.ini writer. is_priced states its own
predicate completely rather than trusting its callers, since the dialog
is a third writer and the file is documented for hand-editing.
Records why float money stays, and that exact-equality assertions in the
cost test hold for the chosen prices rather than in general.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Resolution merges per field, not per source, so an entry setting only
ctx_size still inherits the provider's prices. Cost prices each reply at
the model that produced it, and rows predating the token columns
contribute zero rather than a fabricated number.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
float() accepts "nan" and "inf", so a hand-edited models.ini could feed
either into the cost arithmetic and render "$nan". Rejected once in _get
rather than at each downstream call site.
A model whose id is literally DEFAULT wrote a [DEFAULT] section, which a
stock configparser reader treats as defaults inherited by every other
model. Since the file is documented for hand-editing, that reintroduces
the contamination the renamed default_section exists to prevent, so the
id is escaped on write and mapped back on read.
Also drops the marker branch in save(), leaving mark_skipped() as the
sole writer of the "marker present iff no real keys" invariant, and
prunes two test blocks that could not fail.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Cloud models have no presets.ini entry, so context size, vision and
prices are recorded per model in a file the app writes. A cancelled
dialog leaves a marker so a model tried once never asks again, which is
distinct from an absent section meaning never asked.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
A malformed local entry was the one skip that said nothing, on the
reasoning that local still ends up working. It does, but the merge
rebuilds it from the bare base_url, so an api_key, filter or ctx_size
set on that entry is dropped without a word. The app then comes up
looking healthy, which is the strongest possible signal that nothing is
wrong, making this the case that most needs saying, not least.
Local surviving at all depends on DEFAULTS supplying base_url, since
that is what the rebuild reads. Noted in config.py so the key is not
removed as redundant, and pinned by a test.
The not-a-table wording now follows what was written: a list gets the
single-bracket fix by name, while a scalar entry does not, since telling
someone to fix a [[...]] they never typed points at the wrong line.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0176tzAW6H1i2Kz8vm2XXVGV
|
|
Eight config shapes, all valid TOML, crashed config.load() before any
window existed, so the user got a traceback on a terminal a desktop
launch does not have. Skip malformed providers instead, keeping the
local path working, and say on stderr which one was dropped and why.
The guards are isinstance tests rather than a try/except because the
shapes raise three different exception types, and the local entry is
checked the same way: [[providers.local]] is a truthy list that raised
during the bare base_url merge, before the loop could skip it.
A config that cannot load at all now reports the file and exits rather
than half-starting on defaults the user never configured.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0176tzAW6H1i2Kz8vm2XXVGV
|
|
base_url stays populated so existing callers are untouched; the provider
table is built alongside it. models.ini sits beside config.toml and
state.ini, following the same pattern as the window layout.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
A TimeoutExpired chained without from None already prints no secret,
since its __str__ shows only the command and the duration. The suppressed
display buys a tidier traceback, not a narrower leak, and the comments
claimed otherwise. The from None calls stay.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Raise KEY_TIMEOUT to 120s: a graphical pinentry plus a hardware token the
user has to find and touch makes 30s a plausible successful unlock, and
failing one converts a slow success into an error the user can only retry
against the same clock.
Point the timeout message at a pinentry that never appeared, which is the
silent failure, rather than at one the user is already looking at. Name
the binary when pass is missing instead of reporting a bare Errno 2.
Raise from None so a partial secret on the failed process stdout stays
out of printed tracebacks, and record on the cache that it is unlocked
only because every caller resolves on the GUI thread today.
Rename KeyError_ to KeyResolutionError before later tasks import it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
One prefix-dispatched field. Resolution is lazy so a local-only session
never triggers a pinentry, cached for the process lifetime, and bounded
by a timeout so a stuck pinentry surfaces as an error instead of a frozen
send. Failures name the provider.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
A provider name containing a colon builds ids that split back to a
different provider, and an empty name builds ':model'; both routed to the
local router under a nonsense name with no error, so parse() now skips
them. split() stays non-injective by design, since bare local ids are the
premise here, so the residual risk is recorded as a ponytail comment with
its upgrade path rather than hidden. Also pin the trailing-colon form Task
9 relies on, and the empty-needle filter, which were correct but untested.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Cloud models are addressed as provider:model; local ones stay bare so
existing sessions keep resolving. An unknown prefix is treated as part of
the model name rather than a provider. Filters are case-insensitive
substrings because provider ids capitalise inconsistently.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
The synthesis guard tested key membership, so a [providers.local] that set
only an api_key claimed the slot, blocked the bare base_url from filling
it, then failed the URL check and vanished. Losing the local provider is
the one outcome this feature cannot have. The guard now tests the URL and
merges, so the table adds detail to the local provider rather than
replacing it.
A filter given as a bare string was iterated character-wise, turning
filter = "qwen" into four needles that match almost every model id. Both
failures were silent, which is what made them worth fixing now.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Providers come from a [providers.*] table. A bare top-level base_url
synthesizes the local provider so existing configs keep working, and an
explicit [providers.local] wins over it. Unset numbers stay None so
'unknown' never collapses into zero.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Titles: once a reply lands in a session still wearing its placeholder
title, the exchange goes back to the model in a short side request asking
for six words or fewer. Keyed on the placeholder rather than the turn
number, so an empty first reply does not forfeit titling for the session
and a one-shot window is titled from the question that was answered.
Thinking is disabled for that request, or a reasoning model spends the
whole budget thinking and returns nothing.
Empty replies: the final search round now says so in the tool result.
Withdrawing the tool schema is invisible to the model, which asks for
another search regardless; the request then surfaces as literal
<tool_call> text or vanishes into the thinking block, leaving the reply
empty either way. The note rides on the tool result because Qwen3.5's
template rejects a trailing system message outright.
max_searches now defaults to 2. The other half of that bug was
llama-server's quantized KV cache: with cache-type-k/v = q8_0 three of
five turns broke, and f16 answered six of six.
Measured against a live router, not mocked.
|
|
A second tool call in the same turn, issued after the model has seen the
first set of results, does not arrive as a tool_calls delta. It comes back
as literal <tool_call><function=web_search> text inside reasoning_content,
with finish_reason stop, no tool_calls and empty content. There is no
structured call for the loop to act on, so the turn ends and the user is
shown a blank reply.
Reproduced 3/3 with Qwen3.5-9B through the router; the first call of a turn
parses correctly every time, so this is specific to a call that follows a
tool result.
A retry with the tool withdrawn was tried and rejected: it produced an
answer only some of the time and could surface the raw XML in the reply,
which is worse than the blank it replaces. Capping at one search avoids
reaching the broken state at all. Empty replies drop from 3/3 to about 1/3
and the XML no longer leaks, but the failure is not eliminated.
Raise max_searches once the chat template parses follow-up calls.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
The model is offered a web_search tool and decides when a question needs
current information. llamachat runs the query against SearXNG's JSON API,
feeds the results back, and the model answers from them. Off unless both
search_enabled and search_url are set, so an upgrade never starts talking
to the network on its own.
The loop lives in backend.stream_chat: a round ending in tool_calls is
searched for, the result appended, and the request re-sent. After
max_searches rounds the tool is withdrawn, which forces an answer rather
than letting an uncertain model search forever.
Searches show as a collapsible block above the thinking block, listing
each query and its sources. The queries are visible on purpose: when an
answer is wrong it is usually the query that was wrong, and without
seeing it a bad search and a bad answer look identical.
Two failures found while testing against the live stack shaped the
design:
The model has no clock, so it falls back on its training cutoff and
writes that year into the query itself ("latest kernel ... 2025"),
poisoning the results before they are fetched. The current date now goes
into the system prompt whenever search is on, with an instruction not to
date its own queries.
A SearXNG whose engines are all rate-limited or CAPTCHA'd returns a
valid response with zero results. Reporting that as "no results" tells
the model the web is empty and invites a confident answer from stale
training data, so a search where every engine failed is now an error
naming the engines.
Results are attacker-influenced text entering the model's context. Only
title, url and the snippet survive, snippets are truncated, result text
is escaped on display, result links are never fetched automatically, and
a reply forging the search block's URL scheme has it defused as the
reasoning scheme already was. None of that stops a poisoned snippet from
influencing the answer, which is why the sources stay visible.
Adds eight test groups, 23 to 31, all hermetic behind a fake HTTP layer.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
A delta carrying both reasoning_content and an empty content was
classified by truthiness, so the reasoning fell through to the content
branch. The tail of the model's thinking was stored and displayed as the
answer, splitting mid-sentence across the two fields.
Classify on the presence of the field instead, and return nothing for an
empty reasoning delta rather than falling through.
The visible symptom was a code block swallowing the message: leaked
thinking is dense with backticks, and an odd count leaves a run open to
the end of the reply. Close an unterminated fence before parsing, with a
closer matching the opener's length, and drop a lone dangling inline
backtick. Streaming needs this regardless, since a fence is unclosed on
nearly every frame.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Ctrl+N starts a fresh conversation and puts the cursor in the input box.
Ctrl+F jumps to the search field.
The search field sits in the top bar, but its results render in the history
panel, so focusing it reveals a hidden panel rather than leaving the search
with nowhere to show its hits.
Escape now backs out of the search field before it hides the window: first
press clears the text, second returns to the input, third hides. Previously
a stray Escape while filtering dismissed the whole window.
That check reads focusWidget() rather than hasFocus(), so it still holds
when the window is not the active one.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
The history panel now hides, on the ☰ button in the top bar and on Ctrl+\.
Its width is captured before hiding and reapplied on show, so toggling does
not snap the panel back to a default. Both the width and whether it was
hidden persist between runs.
Layout state lives in state.ini beside the config rather than at QSettings'
default path, which keeps it obvious where it is and lets the checks point
it somewhere temporary instead of writing to the real one.
The button's checked state, not isVisible(), is the authority for whether
the panel is shown: isVisible() is False for every child of a window that
has not been mapped yet, so consulting it during startup would disagree
with what the user sees.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Two additions to the chat window.
A context meter in the top bar shows how much of the active model's window
the next request will occupy: system prompt, prior turns, attachments and
the current draft. The router does not report a usable context size in
router mode (/props returns n_ctx 0), so the limit comes from ctx-size in
presets.ini. Streaming requests now ask for usage statistics, which makes
the figure exact after each reply at no extra round trip; before that it is
an estimate marked with a leading ~.
The estimate reads low on models with a reasoning budget, since thinking
tokens cannot be known before the reply arrives. The tooltip says so.
System prompts are markdown files in the prompts/ directory beside the
config, one per file, so they can be edited in an editor and kept in
version control. default.md is global, any other file is a named preset
that replaces it rather than adding to it, and a conversation may instead
carry one-off text or opt out entirely. The choice is stored per session so
reopening a chat restores the prompt it was built with.
Prompt names are reduced to safe filenames, so a name like ../../etc/passwd
cannot write outside the prompts directory, and a name that reduces to
nothing is rejected rather than creating a dotfile.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
A native PySide6 chat client for a local llama.cpp server in router mode.
Runs as a single persistent process with a Unix-socket control channel, so
a Hyprland keybind toggles the window with a socket round trip rather than
a process start. The control commands do not import Qt and answer in under
a tenth of a second.
Features: one-shot and multi-turn chat modes, runtime model discovery from
/v1/models, streaming replies rendered as markdown, collapsible reasoning
for models with a thinking budget, drag-and-drop file and image attachment
with context-aware truncation, and SQLite history with FTS5 search.
Markdown is parsed with MarkdownNoHTML so markup in a reply is displayed
rather than interpreted, and links a reply produces cannot drive the
interface.
Includes the Hyprland Lua snippets for autostart, keybind and window rules,
and a self-check suite covering everything except the GUI.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
