LeemanCheung/dsh-agent-arena

dsh-agent-arena

English | 中文

A DSH coding arena for comparing 2–4 configured models in isolated Git worktrees, validating their changes deterministically, reviewing every diff, and explicitly applying one winner.

Screenshot

Generated with GPT Image from the implemented Client layout and feature set; runtime appearance follows the active DSH theme and viewport.

Execution and persistence

• Creates detached worktrees under the operating system temporary directory, outside the compared repository.
• Restricts recursive cleanup to the exact two-level match/contestant path under that dedicated Arena directory, including resolved junction and symlink checks.
• Starts each contestant with the public ctx.agents.create, SessionId, createUserMessage, and Agent.followup APIs.
• Executes Git and validation argv through ctx.subprocess.spawn; a shell is never used and shell operators are rejected.
• Persists match reports through storageDomain. An interrupted nonterminal match is marked failed after Host restart.
• Polls the generated agentArena Remote namespace in Settings and supports Start, Cancel, diff review, and Apply winner without browser globals.

Scoring and application

Each validation has a positive weight. A contestant score is the percentage of total validation weight that exits successfully; score ties use contestant id as a stable deterministic tie-breaker. No LLM judge is used. The form starts without a generic validation because git status --porcelain=v1 normally exits successfully regardless of whether it prints changes; add the repository's own test, typecheck, or build command before starting.

A match requires clean git status --porcelain=v1 before worktrees are created. The winner worktree remains available until explicit application. Apply repeats the cleanliness check, verifies that HEAD still equals the recorded base revision, runs git apply --check, then git apply --index --whitespace=error. Arena never auto-applies, commits, pushes, or rewrites history.