Skip to content

desktop: the chat page in a window that starts colibri by itself - #2534

Merged
JustVugg merged 3 commits into
devfrom
feat/gui
Nov 1, 2026
Merged

JustVugg merged 3 commits into
devfrom
feat/gui

Conversation

@JustVugg

@JustVugg JustVugg commented Nov 1, 2026 •

Copy link
Copy Markdown
Owner

The Tauri shell in desktop/ becomes a real program: it opens the chat page (ui/, the one coli ui serves) and brings up what the page needs, with nothing to type.

the colibri window on Windows: the chat page, a question, the reasoning folded, the answer and its speed

What it does

  • Finds coli next to itself (a release archive) or above it (a checkout, from desktop/src-tauri/target/release); COLIBRI_COLI names it by hand.
  • Finds Python 3: py -3, python, python3, then the Windows install folders (python.org, Anaconda, Miniconda), asking each interpreter for sys.executable, so the Microsoft Store alias does not count as a Python; COLI_PYTHON names it by hand.
  • Uses a colibri server that already runs (the one coli setup configured, or 127.0.0.1:8000): a second engine next to it would only run the machine out of memory. Otherwise it runs coli start and waits for the model, showing how long it has been loading; when the server dies while loading it shows the tail of coli logs.
  • Serves the page with ui/server.py on 127.0.0.1:18790, a fixed port because the page keeps its conversations in the browser storage of its address (the next free port up to 18799 when another program holds it).
  • Closing the window stops the page's server, and runs coli stop for a server the window started. A server that was already running keeps running.
  • --backend URL or COLIBRI_BACKEND: another server (an ssh tunnel works); the window then starts and stops nothing.
  • Links that leave colibri open in the system's browser.
  • While it starts, a loader page (Italian or English, from the system language) says the step it is at; when something fails it says what to do (coli setup, install Python) with a Retry button.

The React dashboard in web/ is no longer wrapped (beforeDevCommand and the Vite dev URL are gone): the window shows the same page as the browser, no second frontend.

The chat page's first-run dialog now waits for the models service to finish measuring the machine. On a fresh state it opened at once and said "- GB di RAM" with no model that fits; the window's first start made that visible.

Tested on Windows 11 (i7-1355U, 34 GB, WebView2)

Built with the GNU toolchain, cargo build --release (13 MB with strip), cargo fmt --check and cargo clippy -- -D warnings clean.

  • --backend through an ssh tunnel to a CPU-only server with Qwen3.8-Flash-Next vq2: the page loads, a chat streams (9.7 tok/s), closing the window leaves no Python process and port 18790 free.
  • No --backend, a setup whose server is stopped (a stand-in server through COLI_SETUP_HOME, so no model loaded on the laptop): "Looking for the colibri server", "Starting", the page once /health answers; closing the window ran coli stop, which the server took (pidfile gone, its port closed).
  • Nothing set up and nothing on 8000: the message with coli.cmd setup and Retry.
  • Fresh models state: the first-run dialog waited and showed the machine (34.0 GB, Iris Xe, 336 GB free) with the models that fit.

In the release archives

release.yml builds desktop/ with cargo on the three platforms (WebKitGTK on Linux; pwsh on Windows, so the MSVC linker is the one found), copies colibri-desktop / colibri-desktop.exe next to coli, and the verify step checks it is in the unpacked archive. It needs a display to run, so CI proves the build and the packaging, not a window. coli.cmd's help names it.

Run here on Windows only: the Linux and macOS builds are proved by this PR's Release run.

JustVugg added 3 commits November 1, 2026 13:30
The Tauri shell now shows the chat page (ui/, the one `coli ui` serves)
instead of the React dashboard, and brings up what the page needs:

- finds coli next to itself or above it (COLIBRI_COLI), and a Python 3
  (py -3, python, python3, then the Windows install folders; COLI_PYTHON)
- uses a colibri server that already runs (the configured one or
  127.0.0.1:8000), otherwise runs `coli start` and waits for the model,
  showing how long it has been loading
- serves the page with ui/server.py on 127.0.0.1:18790, a fixed port
  because the page keeps its conversations per origin
- closing the window stops the page's server, and `coli stop` for a
  server the window started; one that was running keeps running
- --backend URL / COLIBRI_BACKEND for another server: starts nothing
- links that leave colibri open in the system's browser
- a loader page (Italian or English) with the step it is at, the log
  tail and a Retry button when something fails

The chat page's first-run dialog now waits for the models service to
finish measuring the machine: on a fresh state it opened at once and
showed "- GB di RAM" with no model that fits.
The release workflow builds desktop/ with cargo on each platform
(WebKitGTK on Linux, pwsh on Windows for the MSVC linker), copies
colibri-desktop next to coli, and the verify step checks it is in the
archive. coli.cmd's help says the window is there; the README and the
changelog say where it ships.
@JustVugg
JustVugg merged commit 1733f4e into dev Nov 1, 2026
121 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant