dsh-review
为 DeepSeek Harness 打包的多智能体对抗式代码审查：针对每个审查视角并行运行一个查找器，然后为每个发现运行一个任务是反驳该发现的验证器，因此只报告已确认的缺陷。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add dsh-review@0.1.1Review 以 dsh-review 发布，当前版本为 0.1.1。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。
点击下方 Badge 复制 Markdown，粘贴到 README 即可。
验证 package.json 声明的 GitHub 仓库，即可管理这个公开页面。认领后，Hub 会优先安排当前版本的安全扫描，并在通过后公开展示结果。
认领这个 Plugin →Packaged multi-agent adversarial code review for DeepSeek Harness.
English | 中文
Status: M0 — the review tool works end to end.
dsh ships the primitives for multi-agent work — subagents, workflows, ralph loops — but nothing that packages them into a review you can trust. Run naive parallel reviewers over a diff and you get a wall of plausible-sounding findings, most of them wrong; triaging them costs more than the review saved. "AI code review produces too many false positives" is the reason people stop using it.
Two stages, and the second is the one that matters:
Measured on this repo's two sibling plugins (dsh-preview, dsh-pilot): 73 agents, 49 confirmed findings, 14 refuted. Two of the confirmed ones were only established because a verifier wrote a script and reproduced the failure — including a defect where the agent silently clicked the wrong same-named button on shadow-DOM pages.
dsh plugin --profile web add dsh-review
Requires a composed subagent provider (the stock spawn provider in dsh-base is the default) and Node ^22.19 || >=24.
One tool, review. It is the most expensive thing in a session — every lens is an agent and every finding costs another — so it is a pre-release audit, not a per-commit check. A full run on a real change took minutes and a double-digit number of agents in our own use; budget for that, and use depth: quick when you want a cheap look at a small change.
Describe the target the way you would brief a colleague who has the repository but not the context:
Review the uncommitted changes in src/policy.ts and src/index.ts (run git diff).
They add an origin gate that must follow the session's approval stance: a session
with approval policy 'never' passes silently, an 'ask' session is prompted once
per origin, and a grant must never leak to another session.
The tool returns confirmed findings — file, line, what is wrong, the failure scenario, and a suggested fix — plus the titles of findings that were refuted, so you can see what the verification stage filtered out rather than wondering what it missed.
review { target: "...", depth: "quick" }
quick caps the run at two lenses, four verified findings, and one verifier — roughly a third of the cost. Verification runs at both depths: a cheaper review looks at less rather than trusting more, because reporting an unverified claim is the failure this plugin exists to avoid.
The bundled adversarial-review skill teaches the agent when a review is worth its cost and how to act on the two categories differently.
The evidence here is that the method finds real defects — 49 confirmed across the sibling plugins, several reproduced by a verifier writing a script. What has not been measured is whether it beats simply asking the model to review the same diff: no A/B, no false-positive rate against a baseline. Treat the refutation stage as a design argument backed by observed refusals (14 findings dropped), not as a proven improvement over the obvious alternative.
- id: review
name: dsh-review
config:
subagentProvider: spawn # which composed provider runs the children
lenses: [] # [] runs every built-in lens
verifiersPerFinding: 1 # raise for a stricter panel; all must confirm
maxFindings: 12 # verification budget, worst severities first
maxConcurrentChildren: 8 # cap on children running at once
maxDepth: 2 # delegation-depth cap for review children
registerSkill: true
Lenses: correctness, lifecycle, contract, security. Each is one child agent, and each finding costs verifiersPerFinding more — a review is the most expensive tool in a session, which is why the skill tells the model to use it deliberately.
verifiersPerFinding > 1, one refutation is enough to drop a finding — the asymmetry is deliberate.dsh-review reviewed its own source and found three defects, which are fixed and pinned by tests: a zero dedupeThreshold merged every finding in a file (distinct defects silently discarded as duplicates), the verifier fan-out had no concurrency bound, and maxDepth was the one numeric config never validated at load. It also refuted two findings, one of which was a genuine false positive about a value the entry point already validates.
| Plugin | What it gives your agent |
|---|---|
| dsh-preview | 👁 Eyes — verify what it builds: open, read, screenshot, self-check |
| dsh-pilot | ✋ Hands — operate any page by accessibility refs, with a native permission model |
| dsh-review (this repo) | 🔍 Judgement — find defects, then try to refute each one before reporting it |
| dsh-design | 🎨 Taste — constrain the choices, then measure whether the result kept them |
MIT © Viger1