# 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). ## [Unreleased] ### Added - External OpenAI-compatible providers. The local llama.cpp router stays the default; adding a `[providers.*]` table brings cloud models into the picker as `provider:model`. The generated config ships the provider block commented out, so a fresh install never reaches an endpoint nobody configured. - Three ways to supply a provider API key: `pass:name` reads from the password store, `env:VAR` reads from the environment, and a literal `sk-...` string is used as-is. `pass:` is preferred because it keeps secrets out of `config.toml`. - Lazy key resolution. The password store or environment is only consulted on the first request to a provider, so a local-only session never triggers a pinentry. - Model filtering per provider. The `filter` key keeps only models whose id contains one of the configured substrings, case-insensitively. Providers list hundreds of models, and without a filter the picker is unusable. - A per-model settings dialog for context size, vision support and input/output prices. Values are saved to `models.ini` beside `config.toml`, so provider defaults can be overridden per model. - Per-conversation cost readout with projection. The top bar shows what has been spent so far plus what the current draft would add, based on hand-entered prices. This exists to make visible that a long conversation on a cloud provider is billed for its whole history on every turn. - Three new columns on `messages`: `prompt_tokens`, `completion_tokens` and `model`. Token counts come from the stream's final usage chunk and power the cost readout; the model column records which model actually replied, so a switched conversation prices each turn correctly. Tool calling and reasoning output vary between providers. Web search on a cloud model may not work as reliably as it does with a local router. ## [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 `` 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.