Chi (χ)

Chi ("kai") is an open-source autoresearch harness: point a fleet of LLM coding agents (any vendor) at any problem with a programmatic evaluator — a build, a correctness check, and a score — and let them iterate unattended.

Status: v1 (Phases 0–1) — single coder agent, two adapters (headless vendor CLIs and a LiteLLM tool loop), enforced SQLite+JSONL run store, hard budget caps, two-layer steering, deterministic watchdog.

Try it in 10 seconds (no API key)

Chi ships an offline demo that needs no provider, no key, and no network. The scripted adapter replays three canned candidates for the bundled optimize_function problem and really evaluates each one:

uv tool install getchi
chi run examples/offline.yaml   # run from a checkout of this repo

You'll see the champion beat the O(n²) baseline as the run replaces it with the O(n) itertools.accumulate rewrite — a real "★ new best", zero setup:

"baseline_score": 25.18,   # O(n²) prefix sums
"champion_score": 0.055,   # O(n) itertools.accumulate
"status": "done"

Run it inside the interactive session (chi, then /run examples/offline.yaml) to watch the ★ new best lines stream in live. This is the whole harness — evaluator, run store, champion selection, watchdog — exercised with no key.

Quick start

uv tool install getchi         # real install; then just run `chi`
chi                            # opens the full-terminal session (Textual UI)

For local development instead: uv tool install --editable . (or uv venv --python 3.12 && uv pip install -e ".[dev]").

The session has a Claude Code-style interface: scrolling transcript, bottom input with a slash-command dropdown, live status bar, and modal fuzzy pickers. chi --plain gives a minimal line-based REPL (also used automatically when stdout is not a terminal).

Inside the session:

/vendors            pick providers (fuzzy); store keys with `chi providers --set-key X`
/models             pick coder models — saved as your global defaults
/run fleet.yaml     start a run; iteration results stream in live
just type           plain text while a run is active becomes a steering directive
/stop  /status  /ledger  /champion  /quit

Everything also works non-interactively for scripts and CI:

chi providers --enable anthropic,deepseek
chi models --pick anthropic/claude-sonnet-5,claude
chi validate examples/fleet.yaml
chi ping --fleet examples/fleet.yaml
chi run examples/fleet.yaml
chi steer runs/<run_id> "stop micro-tuning; try itertools"

Inspect results:

uv run chi status runs/<run_id>
uv run chi ledger runs/<run_id> --negative
uv run chi champion runs/<run_id> --export best.py

Define your own problem

A problem is a directory with a problem.yaml (see problems/optimize_function/): entrypoint commands for correctness and benchmark, held-out seeds, and a score metric/direction. Correctness is a hard gate; candidates never see reference outputs.

Coder adapters

• litellm_loop — tools-in-a-loop over any LiteLLM-routable model (Anthropic, OpenAI, DeepSeek, GLM, MiniMax, ...).

• cli_subprocess — drives a headless vendor CLI per iteration, e.g.:

  coders:
    - id: c1
      model: claude
      adapter: cli_subprocess
      command: "claude -p --permission-mode acceptEdits {prompt_file}"
  

• scripted — deterministic playback of canned candidates; powers the test suite and the no-key examples/offline.yaml demo.

Design

See docs/superpowers/specs/2026-07-25-chi-v1-design.md (v1 design, grounded in a multi-agent fleet postmortem) and docs/product-spec-v1.md (full product spec). License: Apache-2.0.