Skip to content
PanoplosPublic
forked from earendil-works/pi

About

AI agent toolkit: unified LLM API, agent loop, TUI, coding agent CLI

Resources

Contributing

Security policy

Stars

0 stars

Watchers

1 watching

Forks

 
 

Repository files navigation

Pixi logo

upstream synced with upstream pi

New issues and PRs from new contributors are auto-closed by default. Maintainers review auto-closed issues daily. See CONTRIBUTING.md.

Pixi

Pi eXperience Improved.

Pixi is Pi with extra UX and provider work merged in. Nothing is taken away: providers, config, extensions, skills, prompt templates, themes, packages, RPC, and the SDK behave exactly as Pi documents them.

This README covers what Pixi adds and how to install it. Pi itself is documented elsewhere, and those docs apply here as-is:

Installing Pixi

curl -fsSL https://raw.githubusercontent.com/Panoplos/Pixi/main/scripts/install.sh | sh

The installer downloads the latest release binary for your platform from the releases page and installs it as pixi.

Until the first release is published, run Pixi from source:

git clone https://github.com/Panoplos/Pixi && cd Pixi
npm install --ignore-scripts
npm run build
mkdir -p ~/.local/bin && ln -s "$PWD/packages/coding-agent/dist/bundle/cli.js" ~/.local/bin/pixi

Pixi needs Node.js 22.19 or newer. Start it in the directory you want to work in, connect a provider with /login, and give it a task:

cd /path/to/project
pixi

Enhancements in this fork

  • Extra providers — Bitdeer, DeepInfra, and Inco are wired in as first-class providers, with model catalogs seeded from live API data and refreshed by the model generator.
  • Next-message suggestions — after each assistant reply, a small model drafts the suggested next prompt, shown as ghost text that fills in word by word; Tab accepts the whole line, or the right arrow accepts it a word at a time. The suggestion model is picked separately from your main model — choose it in /settings → Suggestion model, or set suggestions.model (a "provider/model" string) in ~/.pi/agent/settings.json — so a fast, cheap model can draft suggestions while your main model does the heavy lifting.

Ghost-text suggestion filling in as you pause typing

  • Model picker grouped by lab — /model groups models by the company that makes them, and /thinking and /context let you tune reasoning effort and context display without digging through settings.
  • Editor UX — click to position the cursor, drag to select with delete/replace, and pasted images become atomic [Image N] markers that move and delete as one unit. Selections copy straight to the system clipboard.

Drag-selecting and deleting text in the editor

  • Queued messages — type follow-ups while Pixi is still working; they send in order, images intact.
  • Write/edit previews — tool previews render proper diffs for file writes and edits, so you can review changes before they land.

Write/edit tool preview rendering a diff

  • Collapsible activity — batches of tool calls collapse into one-line summaries ("ran 1 shell command, read 1 file") that expand in place with ctrl+o, and a thinking spinner shows what Pixi is doing between actions.

Collapsed action details expand in place with ctrl+o

Thinking activity while Pixi works

  • Configurable prompt prefix — the input editor prefix is a promptPrefix setting with a matching optional theme color (falls back to accent).

Changing the input prefix with a promptPrefix setting

  • RPC on a running session — an interactive session serves the RPC protocol on ~/.pi/agent/rpc/<session id>.sock, so a script or editor can prompt or inspect a session you are using in the terminal; the footer shows rpc <session id> while it is on, and --no-rpc turns it off. See RPC mode and CLI Integration for RpcClient and connectRpcClient().

Prompting a terminal session from another shell with pixi-rpc send

  • Secret masking — key-guard.ts masks API keys, tokens, and PEM blocks before they enter model context. It shells out to pixi-mask, a 1.5 MB Rust CLI that runs the betterleaks rule set over a tool result in a few milliseconds, and falls back to in-process patterns when the binary is not on PATH.

A credential masked before the model sees it

Sandboxing

Like Pi, Pixi has no built-in permission system: it runs with the permissions of the user and process that launched it. For stronger boundaries, containerize it — see containerization for a host-plus-micro-VM extension, plain Docker, and policy-controlled sandboxes.

Working on Pixi

npm install --ignore-scripts  # Install all dependencies without running lifecycle scripts
npm run build         # Refresh model data, then build all packages
npm run build:offline # Rebuild using existing model data without network access
npm run check         # Lint, format, and type check
./test.sh            # Run tests (skips LLM-dependent tests without API keys)
./pi-test.sh         # Run pixi from sources (can be run from any directory)

Pixi follows Pi's monorepo layout and build scripts. For Nix builds, local package artifacts, standalone binaries, and the supply-chain rules (pinned dependencies, the lockfile gate), use the matching sections of Pi's README.

Contributing

See CONTRIBUTING.md for contribution guidelines and AGENTS.md for project-specific rules (for both humans and agents). Longer term plans for Pi live in the RFCs.

License

MIT

About

AI agent toolkit: unified LLM API, agent loop, TUI, coding agent CLI

Resources

Contributing

Security policy

Stars

0 stars

Watchers

1 watching

Forks

Releases

Packages

Contributors

Languages