# Steak Pi [![Latest release](https://img.shields.io/github/v/release/michael-berardi/steak-pi?label=release)](https://github.com/michael-berardi/steak-pi/releases/latest) [![MIT License](https://img.shields.io/github/license/michael-berardi/steak-pi)](LICENSE) ![Pi 0.87+](https://img.shields.io/badge/Pi-0.87%2B-8A2BE2) ![Node.js 22.19+](https://img.shields.io/badge/node-22.19%2B-339933) Subagents, local compaction and edit verification for the [Pi coding agent](https://pi.dev). For developers working through multi-file changes and long terminal sessions: keep Pi's editor and tools, add bounded workers, a session plan and a status line. Provider-neutral: no model, provider preference or paid route is selected for you. Use your own Pi credentials and model selection; configure worker routes separately. Steak Pi also ships inside [UltraTerm](https://paretocybernetics.com/software/). ## Measured Historical same-model benchmark on macOS; small Python tasks, not a claim about every workload: - [40/40 correct in both harnesses; cost $0.053 vs Claude Code’s $0.089](benchmarks/ab/results-2026-11-04.jsonl). - [Input tokens: 1.96 M vs 3.71 M; wall time: 482 s vs 434 s](benchmarks/ab/results-2026-11-04.jsonl). - [Cost ratio 0.59 (95% CI 0.50–0.70): cheaper, not faster; reproduce the summary](benchmarks/ab/summarize.py). ## Install **macOS or Linux:** ```sh curl -fsSL https://github.com/michael-berardi/steak-pi/releases/latest/download/install.sh | sh ``` **Windows (PowerShell):** ```powershell irm https://github.com/michael-berardi/steak-pi/releases/latest/download/install.ps1 | iex ``` These commands execute the release installer. Inspect it first if needed. It verifies the package archive's SHA-256, installs Node.js 22.19+ and the pinned Pi version if missing, then registers the package with Pi. No sudo is needed. If the installer prints a PATH instruction, apply it before continuing. Already have Pi? Install the current package directly: ```sh pi install git:github.com/michael-berardi/steak-pi@v0.35.5 ``` Requires Node.js 22.19+ and Pi 0.87+; development checks use Pi 1.0.4. Distribution is through Git tags and [release archives](https://github.com/michael-berardi/steak-pi/releases/latest), not npm. See [installation details](docs/INSTALL.md) for paths, pins, archive installation and removal. ## First use From a project you trust: ```sh steak-pi status steak-pi run ``` With a direct Pi install, start `pi` instead. Authenticate with `/login` and choose your model with `/model` as needed. Ask: "Read this project's README and summarize how to run its tests. Do not change any files." **Shell access is unrestricted.** Workers' path ownership is coordination, not an OS sandbox; a worker with shell access can bypass it. Use trusted projects and review [the security model](SECURITY.md) before granting access. ## What you get | Feature | Use | | --- | --- | | Subagents | Delegate exact tasks with ownership, deadlines, cancellation and a message relay | | Compaction | Use an optional local binary instead of a model call; retain raw-session recall | | Verify after edit | Run a configured project check and return failures to the agent | | Session plan | Keep phased todos visible and saved with the session | | Write guard | Require a read before replacing an existing file | | Finish guard | Request another turn for unfinished work, within a configured limit | | Companion UI | Show state, model, context pressure, cache use and cost | Verification is **off until configured**. Put this in `.steak-pi/config.json` (replace the command with your project's check): ```json { "verify": { "command": "npm run -s typecheck", "failLimit": 2 } } ``` Subagents need an explicit selection or your own worker routing configuration. See [worker setup and trust boundaries](docs/REFERENCE.md#subagents-usap) and [model routing](docs/MODEL-ROUTING.md). No fallback chain ships with the package. Local compaction requires [UltraCompress](https://github.com/michael-berardi/ultracompress) on PATH; without it, Pi's built-in compaction remains available. See [compaction settings](docs/COMPACTION.md) and [snapshot retention](extensions/ultracompress/README.md). ## Local models Slot any GGUF into Pi as `local/`: `steak-pi local add ` reads the header, sizes the context to this machine (one slot, flash attention, q8_0 KV cache, 6 GiB of RAM kept free) and writes one `local` provider to `models.json`. Then `pi --model local/`: `llama-server` starts on demand, is held only while a prompt runs, and stops after 10 minutes with no run. It refuses, naming the cause, when memory is short, the disk is above 90% or another model server runs. No model ships with Steak Pi; workers use `ensure`/`release`. Details: [docs/LOCAL-MODELS.md](./docs/LOCAL-MODELS.md). ## Update ```sh steak-pi update --check steak-pi update ``` In Pi, `/steak-update` installs an update and reloads. Existing sessions keep their loaded runtime. `STEAK_PI_NO_UPDATE_CHECK=1` disables background update checks. `steak-pi update --models` refreshes provider catalogs without installing a package. ## Reference - [Session reference](docs/REFERENCE.md): workers, verification, finish guard, settings and adaptive tools - [Installation and launcher](docs/INSTALL.md): paths, environment, updates and uninstall - [Companion UI](docs/COMPANION-UI.md) and [themes](docs/THEME-PASS.md) - [Worker profiles](docs/PROFILES.md) and [subagent protocol](docs/ULTRATERM-SUBAGENT-PROTOCOL.md) - Optional [Claude Code adapter](docs/CLAUDE-CODE.md) and [local models](docs/LOCAL-MODELS.md) - [Reproducible overhead benchmark](benchmarks/overhead/README.md) and [changelog](CHANGELOG.md) ## From source ```sh git clone https://github.com/michael-berardi/steak-pi.git cd steak-pi npm ci --ignore-scripts npm run verify ``` `verify` runs TypeScript checks, commit-identity checks, the full test suite and isolated TUI tests (which require tmux and Python 3). No runtime npm dependencies beyond Pi peers. See [CONTRIBUTING.md](CONTRIBUTING.md). [MIT license](LICENSE). Part of [Pareto Cybernetics](https://paretocybernetics.com/projects/steak-pi/): free, private, open-source tools for working with AI agents.