1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
|
# Changelog
All notable changes to this project are documented in this file.
The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
## [0.3.0] - 2026-08-01
### Changed
- `max_searches` now defaults to 2, up from 1. The empty-reply failure that
forced the lower cap has two causes, both now understood: the model was
never told its last search was spent, fixed below, and llama-server's
quantized KV cache degrades tool calling, which is a server setting rather
than something this client can control. With `f16` KV and the final-round
note, a question needing two searches answered 4 and 5 times out of 6 on
two runs, against 3 of 3 broken before. Single-search turns are solid, 6
of 6. Lower it back to 1 when running a quantized KV cache, or if empty
replies are more annoying than a second search is useful.
### Fixed
- A turn that used its last search no longer ends with an empty reply. The
final round's tool result now states that the search tool is gone and the
answer is due. Withdrawing the tool schema was invisible to the model,
which asked for another search regardless: this model's chat template
defines its own `<tool_call><function=...>` XML dialect, so the request
came back either as literal text in the reply or buried in the thinking
block, and in both cases the reply itself was empty. The note rides on the
tool result because a trailing system message is rejected outright, the
template raising "System message must be at the beginning". Measured on
the question that first showed the bug: 3 of 5 turns broken before, 5 of
5 answered after.
### Added
- Conversation titles are written by the model. Once a reply finishes in a
session still carrying its placeholder title, the exchange is sent back in
a short side request asking for a title of at most six words, and the
answer replaces the opening words of the question the entry was created
with. The trigger is the placeholder rather than the turn number, so a
first reply that comes back empty does not forfeit titling for the rest of
the session, and a one-shot window, which appends every question to the
same entry until New is pressed, is titled from the question that actually
got answered. It runs on its own thread and is silent about failure: an
unreachable router or an unusable reply leaves the original title in place
rather than raising.
Thinking is disabled for that request through the chat template, since a
reasoning model otherwise spends the whole token budget on its thoughts
and returns an empty answer.
- Web search through a SearXNG instance. The model is offered a `web_search`
tool and decides for itself when a question needs current information;
llamachat runs the query and feeds the results back. Off by default, and
enabled only when both `search_enabled` and `search_url` are set, so an
upgrade never starts talking to the network on its own.
- Searches appear in the transcript as a collapsible block above the
thinking block, listing each query and its sources, with a status line
while a search runs. The queries are visible because a bad answer is
usually a bad query, and showing the sources is what lets a poisoned
result be recognised as one.
- Six configuration keys: `search_enabled`, `search_url`, `search_results`,
`search_snippet_chars`, `search_timeout` and `max_searches`. Searches are
capped per turn, after which the tool is withdrawn and the model must
answer.
- Searches are stored per message in a new `searches` column, so reopening a
conversation still shows what was looked up. They are deliberately kept
out of the full-text index, or web text nobody wrote would compete with
the user's own messages.
- The current date is added to the system prompt when search is on. A model
has no clock and falls back on its training cutoff, which it then writes
into the query itself ("latest kernel ... 2025"), poisoning the results
before they are fetched. It is also told not to date its own queries.
- A search where every engine failed is now reported as a failure naming
the engines, instead of as a successful search that found nothing. Rate
limits and CAPTCHAs are the normal way a self-hosted SearXNG stops
working, and "no results" invited a confident answer from stale training
data.
### Security
- Search results are untrusted text entering the model's context. Only
`title`, `url` and the snippet survive, snippets are truncated, all result
text is escaped on display, result links are shown but never fetched
automatically, and a reply forging the search toggle's URL scheme has it
defused the same way the reasoning scheme already was. A poisoned snippet
can still influence what the model says; the sources are shown so that it
can be judged.
## [0.2.1] - 2026-07-31
### Fixed
- Reasoning could be spliced into the reply. A delta carrying both
`reasoning_content` and an empty `content` was classified by truthiness,
so the thinking fell through to the content branch and the tail of the
model's reasoning was stored and shown as the answer. Deltas are now
classified on the presence of the field, and an empty reasoning delta
renders nothing instead of falling through.
- An unbalanced backtick reflowed the rest of a reply as code. Leaked
thinking is dense with backticks, and an odd count left a run open to the
end of the message. An unterminated fence is now closed before parsing,
with a closer matching the opener's length, and a lone dangling inline
backtick is dropped. This also covers streaming, where a fence is
unclosed on nearly every frame.
### Added
- Context meter in the top bar showing how much of the active model's
context window the next request will use. It estimates while typing and
switches to the server's exact token counts after each reply, turning
amber at 75% and red at 90%. The limit comes from `ctx-size` in
`presets.ini`, since the router does not report it.
- System prompts stored as markdown files in `~/.config/llamachat/prompts/`.
A global `default.md` applies to new conversations, named presets replace
it, and any conversation can take a one-off custom prompt or none at all.
A dialog behind the picker adds, edits and deletes them; the files stay
editable outside the app.
- `default_prompt` in `config.toml` chooses what new conversations start
with.
- Streaming requests now ask for usage statistics, which is what makes the
meter exact without an extra round trip.
- A toggle for the history panel, on the `☰` button and `Ctrl+\`. The
panel's width and hidden state are remembered between runs in
`~/.config/llamachat/state.ini`.
- `Ctrl+N` starts a fresh conversation, `Ctrl+F` jumps to the search field.
### Fixed
- Escape while filtering history dismissed the whole window. It now backs
out of the search field first: clear the text, then return to the input,
and only then hide.
### Changed
- Sessions record the system prompt they were built with, so reopening a
conversation restores it rather than applying whatever is selected now.
## [0.1.0] - 2026-07-31
First working version.
### Added
- Persistent background process with a Unix-socket control channel, so a
window toggle costs a socket round trip instead of a process start. The
control commands (`--toggle`, `--show`, `--hide`, `--ping`, `--quit`) do
not import Qt and answer in well under a tenth of a second.
- Two chat modes. One-shot asks a single question with no context carried
over; Chat keeps a resumable multi-turn conversation.
- Model picker populated at runtime from the router's `/v1/models`, with
vision-capable models marked.
- Streaming replies rendered token by token.
- Markdown rendering for replies: headings, emphasis, lists, tables, inline
code and tinted fenced code blocks, via `QTextDocument` with no markdown
dependency. User input is shown literally so attached file contents are
never reflowed.
- Collapsed reasoning. Models with a reasoning budget return their thinking
in a separate field; it shows as a one-line summary that expands on click,
tracked per reply.
- File attachment by drag-and-drop or file dialog. Text and code files are
inlined into the prompt and truncated to fit the model's context with a
warning. Images require a vision model, and dropping one on a text model
offers to switch.
- SQLite history with FTS5 full-text search. Chat sessions reopen and
continue with context intact; one-shot entries reopen read-only.
- Attachment provenance: original path, size, SHA-256, and a thumbnail for
images, so a reopened conversation still shows what was sent even if the
file has since moved.
- System tray icon for show/hide/quit.
- Configuration in `~/.config/llamachat/config.toml`, written with defaults
on first run.
- `presets.ini` parsing for the two things the API does not report: which
models have an mmproj file, and each model's context size.
- Hyprland Lua snippets for autostart, keybind and floating window rules.
### Fixed
- `presets.ini` opens with a bare `version = 1` before any section, which
made `configparser` reject the whole file.
- Sessions sharing a one-second timestamp sorted unpredictably in the
history list and in search results.
- A venv PySide6 bundles almost no Qt plugins, so `QT_QPA_PLATFORMTHEME`
found nothing and every window fell back to Fusion, ignoring qt6ct and
Kvantum. The system plugin directory is now used when its Qt version
matches the bundled one exactly.
- A reply containing a markdown link to the internal reasoning scheme
produced a working toggle, letting model output drive the interface.
The scheme is namespaced, forged anchors are rewritten before display,
and only `http`, `https` and `mailto` links reach the desktop opener.
- Venv discovery compared resolved interpreter paths, so a venv created
with `--system-site-packages`, whose `bin/python3` symlinks to the system
interpreter, was discarded as "already running".
### Security
- Markdown is parsed with the `MarkdownNoHTML` flag, so markup in a reply is
displayed rather than interpreted.
- The control socket is created mode 0600.
- Search input is tokenised and quoted before reaching FTS5, so punctuation
cannot be read as query syntax.
|
