dsh-agent-arena
隔离的多模型编程对赛，配备确定性验证、评分和报告
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:LeemanCheung/dsh-agent-arena#9e3a4bdab028b3cb1c0db1999cfdd19e161eeabeAgent Arena 以 dsh-agent-arena 发布，当前版本为 1.0.1。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
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.
Agent Arena results and diff review
Generated with GPT Image from the implemented Client layout and feature set; runtime appearance follows the active DSH theme and viewport.
ctx.agents.create, SessionId, createUserMessage, and Agent.followup APIs.ctx.subprocess.spawn; a shell is never used and shell operators are rejected.storageDomain. An interrupted nonterminal match is marked failed after Host restart.agentArena Remote namespace in Settings and supports Start, Cancel, diff review, and Apply winner without browser globals.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.
The Settings field accepts conservative whitespace-separated argv such as corepack pnpm test or corepack pnpm typecheck. Quotes, shell variables, pipes, redirections, command separators, backticks, and newlines are rejected. Configure commands whose arguments do not require shell quoting.
dsh plugin --profile web add github:LeemanCheung/dsh-agent-arena
Restart the existing DSH Web process and refresh its page. The selected profile must provide agents, subprocess, storageDomain, Typert Remotes, Settings, and at least two usable provider/model routes..
Every contestant receives the user-entered Arena objective as one user message in its own session and worktree. Token and KV-cache usage therefore scales with 2–4 independent contestant sessions. Validation, scoring, comparison, cancellation, persistence, and application do not call another model.
Validation argv uses a deliberately conservative tokenizer and cannot represent arguments containing whitespace. Cancellation depends on the selected provider honoring the supplied abort signal. The Host captures binary-capable git diff HEAD, so staged, unstaged, and intent-to-add files are reviewed and applied consistently. Patches over the 1 MB review/apply bound are rejected before mutation rather than truncated.
From the repository root run corepack pnpm typecheck, corepack pnpm test, corepack pnpm build, and corepack pnpm pack:check.
Version 1.0.1 is marked compatible with the DSH 0.1.2-rc.1 web profile after Windows QA loaded its Host service, browser Client, generated Remote namespace, and Settings section in both the isolated QA instance on port 3081 and the existing local profile on port 3080. The form showed no misleading prefilled validation and kept Start disabled until required fields were supplied. Automated coverage contains 15 passing tests. QA did not start a real model match, so provider execution and cancellation remain outside this compatibility claim.
CI is configured to rebuild the committed lib artifacts on both Windows and Linux and reject any tracked or untracked difference.
MIT. See LICENSE.