$ cat ~/map/pi-arena.md
pi-arena
Fair, replayable tournaments for pi coding agents: same task, isolated workspaces, objective scoring, and a browser replay console.
tools✓ Nov 1
Facts
- repository
- github.com/Waystom/pi-arena ↗
- GitHub Stars
- ★ 0
- Forks
- 0
- license
- MIT
- npm / mo
- —
- last push
- Oct 30, 2026
install
Clone the repository, then run npm run demo for the deterministic three-contestant local tournament (no API key) and open arena-demo/index.html for the replay console; per SKILL_USAGE.md, pi install <path to your local clone> adds the package to pi, and /skill:arena loads the tournament protocol from skills/arena/SKILL.md inside a session.## TL;DR
A tournament harness for pi coding agents: contestants run the same task in isolated workspaces, required tests and measurable signals score the results, and a shareable browser report replays every run — built for fair comparisons between models, providers, prompts, and agent strategies, with a deterministic local demo that needs no API key.
## Key points
- $Fairness is enforced by code, not trusted to the model: the tournament contract in skills/arena/SKILL.md fixes equal starting state, isolated workspaces, one shared prompt, and model-independent tool policy.
- $Not an LLM judge disguised as a benchmark: required tests are objective, model-based judging is optional, and a judge score never overrides a failing required test.
- $The demo scoring contract: 70% required evaluation checks, 20% efficiency from recorded actions, 10% scope discipline including avoiding test edits; real projects define their own weights in a versioned evaluation file.
- $The replay console turns each run into a small race with action timelines and evaluation checks, and reports separate objective results from optional judging.
- $Scores are only comparable when the task, starting revision, evaluator, and budget are identical; a score is evidence for the declared evaluator, not proof of overall software quality.
- $Safety boundaries: network access, credentials, and destructive commands require explicit user policy, credentials never enter a worktree or report, and the runner fails closed when the requested isolation cannot be established.
- $The first release targets local Node.js fixtures and Linux/macOS development environments; the demo runs three deterministic local contestants without API spending.
- $Roadmap: a pi JSONL RPC contestant adapter, git worktree isolation, pluggable evaluators for npm test, pytest, and type checking, blind human review with signed report manifests, and a GitHub Action publishing static reports.
## FAQ
Q Does a high score prove one model is best?
No. pi-arena reports what happened on a stated task under a stated budget and does not claim any model is universally best.
Q Does trying it require an API key?
No. npm run demo plays a deterministic local tournament with three contestants and no API spending.
Q Can a model judge pass a failing test?
No. Model-based judging is optional, and a judge score never overrides a failing required test.
Q Can contestants run as real pi processes?
Not yet — the JSONL RPC contestant adapter is first on the roadmap; the current release targets local Node.js fixtures.
Q When are two scores comparable?
Only when the task, starting revision, evaluator, and budget are identical.
## References
- README (scoring contract, design boundaries, roadmap) ↗
- Tournament protocol (skills/arena/SKILL.md) ↗
## Related entries
$ pi-dsh-minimal
Pi extension that anchors DeepSeek V4 Pro and V4 Flash to the official DSH minimal surface (one system prompt, two tools) on the first request, then restores pi's normal tools.
$ aft
Sensorimotor-cortex layer that ships thin adapters for Pi and OpenCode, replacing raw read/write/edit/bash with semantic operations.
$ pi-coding-agent-action
GitHub Action that runs the pi coding agent in CI for comment-driven reviews, audits, and PR triage.
## Guides
Last updated: Oct 30, 2026