Agentarena — Testing skill for Claude Code

Agent Arena

Make your coding agents fight for the merge.

Agent Arena is a local Node.js library and CLI that gives two coding agents the same immutable run specification, validates both patches, and runs three attack–repair rounds. Attacks are executable test patches, not critiques. The harness reproduces them twice, verifies that their expected behavior is supported by frozen task text, resolves damage and recoil simultaneously, gives both contestants bounded repair opportunities, and exports the evidence and final patches.

Surviving the arena is additional evidence, not a correctness guarantee.

Review in chat

The primary workflow is a chat or IDE host calling the exported typed operations:

1. reviewRun presents the arena champion, correctness-first recommendation, evidence, choices, and exact patch digests.
2. inspectPatch reads a summary, diff, checks, or quality evidence.
3. recordReviewDecision accepts or rejects only an authenticated user action bound to the current prompt, base commit, contestant, and digest.
4. applyAcceptedPatch applies only that accepted patch.
5. planDelivery, recordDeliveryDecision, executeDelivery, and getDeliveryStatus provide a separate, gated GitHub delivery workflow.

The core never parses conversational language. A chat host maps an authenticated user action to the strict schemas exported from `agent-arena`. Tokens and raw chat text are not stored.

Requirements

• Node.js 22 or newer
• Git
• A clean local Git repository with a passing test command
• Two installed and authenticated CLIs from Codex, Claude Code, and Gemini CLI

Agent-generated code and tests execute with your current OS account's permissions. Git worktrees prevent accidental patch cross-contamination; they are not a hostile-code sandbox. Use a sanitized account or an external container when the host has sensitive credentials.

Install and run

npm install
npm run build
npm link

agent-arena fight \
  "fix issue #241: