Chat with local models (Ollama, LM Studio, OpenAI-compatible) and Markdown answers (Tauri build) - #173
Open
AinzDerErste wants to merge 14 commits into
Open
Chat with local models (Ollama, LM Studio, OpenAI-compatible) and Markdown answers (Tauri build)#173AinzDerErste wants to merge 14 commits into
AinzDerErste wants to merge 14 commits into
Conversation
added 14 commits
October 3, 2026 13:26
With NVIDIA's proprietary driver, WebKitGTK's DMABUF renderer commits frames to a Wayland surface after announcing explicit sync without an acquire point. The compositor then closes the connection with a protocol error and the app dies on its first frame. Switch the DMABUF renderer off when the NVIDIA driver is loaded, unless the user set WEBKIT_DISABLE_DMABUF_RENDERER themselves. Split prepare_environment into one function per concern.
Wayland has no main display, and the layer surface was never assigned an output, so the compositor picked one. - Pin the layer surface to the chosen monitor (gtk_layer_set_monitor). - Settings -> Island lives on lists every display by name and size. - "Main display" now asks XWayland's RandR for the primary output, which KWin and Mutter set from the user's choice, and falls back to the first display.
- Report the compositor's pointer leave (GTK leave-notify) to the page. WebKitGTK sends no mouseleave when the pointer leaves the surface, so the island never learned the mouse was gone and never auto-closed. - Restart the collapse timer when an alert reaches an island that is already open. forceHome cancelled it and only a state change started it again. - Start the drop sequence after the island has opened. Waking a closed island passes through its default view, which stopped the sequence and left "uploading" stuck. - Let clicks through #content and the hidden DOM view to the drop canvas' buttons; Cancel and Ask were covered by invisible elements. - Give the ticker's dim text top:0. Without it WebKit places the absolute box one line below the text it should overlap. - Let notes dismiss themselves after five seconds.
…there is no editor - The relay passes Konsole's D-Bus names. The app switches to the tab over D-Bus and raises its window with a short KWin script, found by Konsole's process id and the tab's title. - Look for code, code-insiders, code-oss, codium, vscodium and cursor in that order. - When no editor is installed the island says so instead of silently opening the file manager. - open_in_vscode returns an enum instead of a bool.
Settings -> Language (system, English, Deutsch, Français). An explicit choice tells
the chat to always answer in that language, whatever the question or the search
results are in. The step labels ("Runs", "Führt aus", "Exécute") follow it, where
they were fixed to French.
Without an API key the chat runs `claude -p` on the user's own login instead of failing with "API key missing". Read, WebSearch and WebFetch are pre-approved, the user's hooks are left out (--setting-sources project, run from a temp directory) so the helper does not show up as a session, and follow-up questions resume the same session.
Embed the handful of Font Awesome Free 7 (solid) paths in use, with the CC BY 4.0 notice, and drop the hand-drawn header icons. The tabs get a little more room.
Clicking the overview's left card opens an editor view while the session has edited a file or run a command: Mochi and the Read / Edit / Bash / Done phases on the left, on the right the changed lines in red and green with a few lines of context, and under them the last command and what it printed. Nothing opens by itself. - snippet.rs reads the lines around an edit, only from a regular file inside the session's folder, up to 2 MB. - The relay forwards the last three lines a Bash command printed, without colour codes; the rest of tool_response stays out.
CSS cannot transition between two gradients, so the halo's colour flipped in one step. Ease colour and opacity per frame, like Mochi's position.
GTK's leave-notify reaches the page over IPC and can overtake the last mouse-move WebKit was still handing it. That move put the mouse "inside" again after the leave, and since nothing follows a pointer that is gone, the island never started its auto-close timer; it stayed open until OK was pressed. Ignore page mouse-moves for 400 ms after a leave. A real re-entry is picked up by the next moves right after.
The view read its data from the one shared "Claude Code" pill. With several sessions running, a new prompt in one of them wiped what the view showed and the project name jumped between them. Each session now has its own record, keyed by session_id, and the view shows the one that was active last and has something to show. A session remembers the folder it started in: its name, the paths in the diff and the folder the file is read from no longer change when the shell cd's. A finished session is dropped after its view has been shown; one that never says so is forgotten after 30 minutes.
As on the Mac (docs/INTEGRATIONS.md, "Jauge de forfait Claude"), for the Tauri build:
- A pill in the header of the overview ("Claude 73%", green below 50 %, orange up to
80 %, red above, grey without data). Clicking it puts the plan card in place of the
left card (5 hours and week, reset times, how old the numbers are), and Mochi wears
the plan's colour while it is open. It closes when the view, the mode or the focus
changes. Off by default.
- The numbers come from Claude Code's own status line (`rate_limits`), like on the Mac,
and are checked the same way: 100-200 % is shown full, anything else is dropped, a
reset more than 400 days away is milliseconds in disguise, and a window whose reset
has passed counts as 0 %.
- The relay runs as the status line (`coucou-hook --statusline`), passes the limits on
and runs the status line the user had before with the same input, printing its output
untouched (10 s limit). That one is kept in statusline-previous.json beside the relay.
- Installing and removing the relay is its own step, apart from the hooks: Settings ->
Plan usage shows the diff of the `statusLine` key only, backs settings.json up and
writes after a click. Only `command` is swapped, so padding and the like stay;
removing puts the previous status line back, or the key goes if there was none, and a
status line the user changed since is left alone.
- Settings -> Plan usage -> Show in the notch turns it on and starts the install first
when the relay is not in yet.
Calls using the older `coucou-hook StatusLine` form keep working.
Claude Code 2.1.85+ sends AskUserQuestion as a PreToolUse. The island now shows the choices and hands the answer back, as the Mac app does (docs/INTEGRATIONS.md, "Répondre aux questions"). - A dedicated hook entry (matcher `AskUserQuestion`, `coucou-hook --ask`, 130 s) is added by the installer next to the general PreToolUse one, which ignores that tool. Hooks installed before this show "Hooks outdated" in Settings with an update button; the usual diff, backup and confirmation apply. - The relay waits up to 125 s and prints the documented PreToolUse output: allow, with `updatedInput` carrying the questions whole plus `answers`. Single-choice values are a string, multi-select an array. No answer, or an unusable one, prints nothing and Claude Code asks in the terminal. If Coucou is closed it exits at once, as before. - One question at a time with a 1/N counter, the options as chips, "Other…" for free text, and "Reply in terminal". A single question with a single choice is answered by the click; several questions or a multi-select use Next / Send. - The text field only takes the keyboard when it is clicked, so a question never steals what is being typed in the terminal. - A question that can't be shown (not 1-4 questions, fewer than two options, a card already up, another agent, paused) is handed straight back to the terminal. PermissionRequest for this tool is declined the same way, for older Claude Code. Answering in the terminal takes the card away.
As the Mac app does since Louis-CFM#156 (docs/INTEGRATIONS.md, "Modèles locaux"), plus any OpenAI-compatible server: - Settings -> Local models connects Ollama or LM Studio (empty address = the usual one on this machine) or a server of the user's own that speaks the OpenAI API, with an optional key kept in the keychain. Connect checks the server answers and lists its chat models (embedding and rerank models are left out); "Chat with" picks who the chat talks to. Disconnect forgets the address, and the key. - The answer is streamed (POST /v1/chat/completions) and shown as it is written, at most 15 times a second. <think> blocks of reasoning models stay hidden while open and are dropped from the finished answer. A dropped text file goes along inline, 24 000 characters at most; images and PDFs by name only. The history is kept apart from Claude's, which carries tool blocks a local server would not understand. - The chat asks for light Markdown now (the "no markdown" rule is gone from the system prompt, as on the Mac) and renders it for every provider: headings, lists, bold, italic, inline code, links, quotes, rules and code blocks with a copy button. It is built as DOM nodes, never HTML, and only http(s) links open. - A local model is not told it has web search. Nothing leaves the machine unless the address points elsewhere.
AinzDerErste
marked this pull request as ready for review
October 3, 2026 13:57
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Part 3 of 3. Draft until #158 is merged: it sits on top of #158 and of the two PRs before it (#171, #172), so until they are merged the diff also shows their commits. Only the last commit, "Chat with local models, and Markdown in the answers", is new here.
#156 for the Tauri build, plus any OpenAI-compatible server:
POST /v1/chat/completions) and shown as it is written, at most 15 times a second;<think>blocks stay hidden while open and are dropped from the finished answer. A dropped text file goes along inline (24 000 characters at most), images and PDFs by name only. The history is kept apart from Claude's, and a local model is not told it has web search.The picture comes from the slowed-down
tests/fake_local_llm.pyof this repo, so the answer is the fake one.Checked:
tsc,vite build,cargo test(URL cleaning, event stream parsing, think filtering, model list and connect against a one-shot local HTTP server, the key header, file inlining), no new warnings; a throwaway test of the Markdown parser and the link rule. The streaming against the fake server, and real chats with my own Unsloth server on KDE/Wayland, with a code block copied out of an answer.Not checked: real Ollama and LM Studio servers, Windows, macOS, GNOME.
Not ported: the Gemini and OpenAI cloud providers of the Mac chat. The French docs still say local models are macOS only; I left them alone and will change them if you want.