dsh-review-kit

A code review plugin for DeepSeek Harness that refuses to report what it cannot prove.

_{by DeepSeek · built on DeepSeek Harness}

English

What it is

dsh-review-kit is a read-only, evidence-gated code review plugin for DeepSeek Harness. It registers one model-facing tool (review) and one on-demand skill (review-kit).

Most "multi-agent code review" plugins ask several agents to list problems and then print whatever they said. This one is built around the opposite idea: a finding only reaches the report if it can be proven against the actual change snapshot and survives an independent attempt to refute it.

Why this one

#	Mechanism	What it does
1	Verbatim anchors	Every finding must quote one real line from the change snapshot. The plugin locates the line itself; a quote that does not exist is rejected as invalid output, never counted as a false positive.
2	Adversarial verification	Each surviving finding gets an independent verifier whose job is to refute it. The verifier never sees the finder's lens or reasoning — only the code and the claimed failure. Unverifiable means not confirmed.
3	Origin attribution	Verifiers classify `introduced` / `preexisting` / `unknown`. A baseline line match is only a hint, never a verdict.
4	Transparent coverage	The report always states which lenses ran, which failed, which files were skipped, which findings the budget dropped, and how many children ran. A clean run reports its scope instead of just saying "no issues".

Also included: a kind whitelist (correctness | state | breaking | security), fingerprint deduplication, a severity-ordered budget with automatic quick degradation, one shared concurrency limiter, and full cancellation propagation.

Requirements

DeepSeek Harness 0.2.2-rc.1 (desktop profile tested), Node >= 22
The subagents service with a provider that supports structured output (spawn qualifies) — the plugin refuses to run rather than silently degrade
Git for the uncommitted / base / commit scopes (files works without it)

Install

Install the directory as a bundle in a Harness profile:

plugin_manager(action: "install_bundle", target: "<abs path>/dsh-review-kit")

The bundle row is declared in cordis.patch.yml. There is no build step and no @deepseek-ai/* import: the package is plain ESM plus Node builtins, so it works as a linked workspace bundle.

To pin configuration, override the row in the profile:

- id: review-kit
  name: '@local/dsh-review-kit'
  config:
    depth: quick
    maxConcurrentChildren: 4

Usage

Ask in plain language — "review this change" — or call the tool:

{ "target": "uncommitted" }                                  // default: worktree vs HEAD
{ "target": "base", "base": "origin/main" }                  // vs a baseline ref
{ "target": "commit", "commit": "<sha>" }                    // the change one commit introduced
{ "target": "files", "files": ["src/a.ts", "src/b.ts"] }     // explicit files
{ "depth": "quick" }                                         // <= 2 lenses, <= 4 verified
{ "lenses": ["security"], "maxFindings": 4 }                 // narrow scope

The model receives a markdown report; the canonical value keeps structured confirmed / refuted / unverified arrays plus coverage and stats for programmatic callers.

A live run against a one-file fixture with lenses: ["correctness"], depth: "quick" took 1m19s and 2 children: one finder reported a missing-key TypeError, the verifier reproduced it with a throwaway script outside the repo (confidence 0.96, origin introduced).

Read the report

Confirmed — file:line, the verbatim quoted line, the failure scenario, the verifier's conclusion and confidence, and the origin. These are the ones to fix.
Refuted — collapsed; the verifier's counter-evidence. Do not report these as problems.
Unverified — the verifier was unavailable. Not confirmed.
Coverage / skipped / stats — the honest part. ✅ ran, nothing found is not the same as ⚠️ failed. A review with coverage gaps must be reported as such.

Configuration

Field	Default	Meaning
`subagentProvider`	`spawn`	Registered `ctx.subagents` provider
`depth`	`full`	`full` or `quick`
`lenses`	`[]`	`[]` = all four (`correctness`, `lifecycle`, `contract`, `security`)
`verifiersPerFinding`	`1`	1–3; all must confirm
`maxFindings`	`12`	Findings that reach verification
`maxCandidatesPerLens`	`8`	Truncation only, not a failure
`maxChildren`	`24`	Estimate = lenses + findings × verifiers
`maxConcurrentChildren`	`6`	Shared limiter
`childTimeoutMs`	`300000`	Per finder/verifier child
`onVerifierFailure`	`refute`	`refute` (fail closed) or `observation`
`childMaxDepth`	`1`	`1` = children exist but cannot delegate further
`maxFiles`	`40`	Changed files read per review
`maxDiffChars`	`400000`	Prompt text budget
`gitMode`	`auto`	`auto` / `tool` / `process`

gitMode: auto runs git through the calling agent's shell tool (ctx.tools.execute), so the deployment's permissions, sandbox and audit path stay in charge. process runs git directly, read-only plumbing only.

The plugin is read-only: no workspace writes, no report files, no git add/commit/stash/restore/worktree. The only writes anywhere happen inside a finder/verifier child's own temporary directory when it reproduces a bug.

Architecture

lib/index.js        apply(): register `review` + the `review-kit` skill
lib/orchestrator.js target -> snapshot -> budget -> find -> gate -> dedupe -> cap -> verify -> render
lib/config.js       defaults, lens/kind/severity vocabularies, doc-file policy
lib/target.js       scope parsing + path-escape rejection
lib/git.js          read-only git through the caller's shell tool (or a process fallback)
lib/snapshot.js     before/after/diff per file, binary/large/doc skips, changed-line ranges
lib/lenses.js       four orthogonal lenses (constraints, not personas)
lib/prompt.js       finder + verifier prompts, prompt-budgeted file blocks
lib/schema.js       parameter schema, finder/verifier output schemas, tool output schema
lib/gate.js         kind whitelist + anchor existence + line resolution
lib/fingerprint.js  sha256(kind, path, normalised anchor)[0:16]
lib/dedupe.js       fingerprint merge + conservative near-duplicate merge
lib/budget.js       child-count estimate, quick degradation, refusal
lib/concurrency.js  shared FIFO limiter
lib/dispatch.js     one finder child per lens; a lens failure loses only that lens
lib/verify.js       independent verifiers; policy-driven failure handling
lib/render.js       markdown report
lib/subagents.js    managed activations via ctx.subagents (`delivery: 'caller'`)
lib/skill.js        on-demand skill provider (rank 550)

Tests

# 45 unit tests (pure modules + a real temporary git repository)
node --test test/unit/*.test.mjs

# scripted end-to-end: real snapshot/gate/dedupe/verify/render path, fake children
node test/e2e/run.mjs

The e2e harness runs the fixture project at test/fixtures/buggy-project (3 seeded bugs, 2 decoys) and proves:

a fabricated anchor and a style kind are rejected as invalid output (2 of 5 candidates),
the two real bugs are confirmed and the <= pagination finding is refuted,
every confirmed line number points at the exact quoted source line,
a clean run still reports coverage, an aborted run returns a schema-valid partial result,
a documentation-only scope is refused.

Runtime notes

The plugin follows the installed Harness runtime rather than older plugin sources:

Subagent API — ctx.subagents.startActivation({ provider, label, request, signal, delivery: 'caller' }), awaiting activation.result and disposing on timeout. delivery: 'caller' keeps one completion notice per child out of the parent conversation.
maxDepth semantics — a request's maxDepth is the absolute delegation depth cap for the child, so 0 refuses to create the child at all. The default is 1.
Schema dialect — parameters and output.schema use ordinary JSON Schema with required: [...] arrays; subagent outputSchema is standard JSON Schema too.

Not included yet

Settings page, /review* slash commands, persisted reports, CI blocking, heterogeneous child models, and a measured false-positive baseline. The tool never auto-fixes code.

Attribution

Built by DeepSeek with DeepSeek Harness. The four mechanisms were inspired by publicly documented, MIT-licensed review plugins; no source code was copied.

中文

这是什么

dsh-review-kit 是 DeepSeek Harness 的只读、证据门禁型代码评审插件。它注册一个模型工具（review）和一个按需加载的技能（review-kit）。

市面上多数「多智能体评审」插件是让几个 agent 各列一遍问题,然后把说过的话原样打印出来。这个插件的出发点相反:一条发现只有在真实改动快照里能逐字对上、并且扛住一次独立复核的推翻尝试, 才会进入报告。

差异点

#	机制	作用
1	引用可校验	每条发现必须逐字引用快照里真实存在的一行。插件自己反查行号;引用的行不存在时记为「无效输出」,而不是误报。
2	对抗复核	每条存活发现都派一个独立复核者,任务就是推翻它。复核者看不到 finder 的视角和推理,只能自己读代码。复核不成立 = 不算确认。
3	归属区分	复核者判定 `introduced` / `preexisting` / `unknown`。基线里有同形行只是提示,不是结论。
4	覆盖透明	报告必然写明:哪些视角跑了、哪些失败、哪些文件被跳过、哪些发现被预算裁掉、用了多少子代理。干净结果也要说清范围,而不是一句「没问题」。

另外还有:kind 白名单(correctness | state | breaking | security)、指纹去重、按严重度消耗的预算与 quick 自动降级、共享并发 limiter、全链路取消透传。

前置条件

DeepSeek Harness 0.2.2-rc.1(desktop profile 实测),Node >= 22
subagents 服务,且 provider 支持结构化输出(spawn 满足);不满足时插件直接报错,不会静默降级
uncommitted / base / commit 范围需要 git(files 不需要)

安装

把目录作为 bundle 装进 Harness profile:

plugin_manager(action: "install_bundle", target: "<绝对路径>/dsh-review-kit")

bundle 行声明在 cordis.patch.yml。没有构建步骤,也不 import @deepseek-ai/*:整个包是纯 ESM + Node 内置模块,作为链接式工作区 bundle 直接用。

要固定配置,在 profile 里覆盖该行:

- id: review-kit
  name: '@local/dsh-review-kit'
  config:
    depth: quick
    maxConcurrentChildren: 4

用法

用自然语言说「审一下这次改动」即可,或直接调工具:

{ "target": "uncommitted" }                                  // 默认:工作区 vs HEAD
{ "target": "base", "base": "origin/main" }                  // 相对某个基线
{ "target": "commit", "commit": "<sha>" }                    // 某次提交引入的改动
{ "target": "files", "files": ["src/a.ts", "src/b.ts"] }     // 指定文件
{ "depth": "quick" }                                         // <= 2 个视角,<= 4 条复核
{ "lenses": ["security"], "maxFindings": 4 }                 // 收窄范围

模型拿到 markdown 报告;规范结果里另有结构化的 confirmed / refuted / unverified 数组与 coverage、stats,方便程序化调用。

实测:对一个单文件 fixture 用 lenses: ["correctness"], depth: "quick", 1m19s、2 个子代理 —— finder 报出缺失键 TypeError,复核者在仓库外写临时脚本复现, 置信度 0.96,归属 introduced。

怎么读报告

确认项 —— 文件:行、逐字引用的源码行、失败场景、复核结论与置信度、归属。这些才是要修的。
被推翻 —— 折叠显示,附带复核者的反证。不要把这里的东西当成问题汇报。
未验证 —— 复核者不可用。不等于已确认。
覆盖 / 跳过 / 统计 —— 最诚实的一段。✅ 已跑,无发现 和 ⚠️ 失败 是两回事; 有覆盖缺口就必须如实说明。

配置

字段	默认	含义
`subagentProvider`	`spawn`	已注册的 `ctx.subagents` provider
`depth`	`full`	`full` 或 `quick`
`lenses`	`[]`	`[]` = 全部四个(`correctness`、`lifecycle`、`contract`、`security`)
`verifiersPerFinding`	`1`	1–3;全部确认才算通过
`maxFindings`	`12`	进入复核的条数上限
`maxCandidatesPerLens`	`8`	只做截断,不算失败
`maxChildren`	`24`	估算 = 视角数 + 发现数 × 复核数
`maxConcurrentChildren`	`6`	共享并发上限
`childTimeoutMs`	`300000`	单个 finder/verifier
`onVerifierFailure`	`refute`	`refute`(宁漏不错)或 `observation`
`childMaxDepth`	`1`	`1` = 子代理存在,但不能继续委派
`maxFiles`	`40`	单次读取的改动文件上限
`maxDiffChars`	`400000`	提示词文本预算
`gitMode`	`auto`	`auto` / `tool` / `process`

gitMode: auto 会让 git 走调用方 agent 的 shell 工具(ctx.tools.execute), 从而保留部署的权限、沙箱与审计链路;process 直接跑 git,但只跑只读命令。

插件是只读的:不写工作区、不落盘报告、不执行 git add/commit/stash/restore/worktree。唯一的写操作发生在 finder/verifier 子代理为了复现问题、在它们自己的临时目录里建脚本时。

结构

lib/index.js        apply():注册 `review` 工具 + `review-kit` 技能
lib/orchestrator.js 目标 -> 快照 -> 预算 -> 找 -> 门禁 -> 去重 -> 截断 -> 复核 -> 渲染
lib/config.js       默认值、视角/种类/严重度词表、文档文件策略
lib/target.js       范围解析 + 路径逃逸拒绝
lib/git.js          经调用方 shell 工具的只读 git(或进程回退)
lib/snapshot.js     逐文件 before/after/diff,二进制/过大/文档跳过,改动行区间
lib/lenses.js       四个正交视角(约束,不是人设)
lib/prompt.js       finder + verifier 提示词,带预算的文件块
lib/schema.js       参数 schema、finder/verifier 输出 schema、工具输出 schema
lib/gate.js         kind 白名单 + anchor 存在性 + 行号反查
lib/fingerprint.js  sha256(kind, path, 归一化 anchor)[0:16]
lib/dedupe.js       指纹合并 + 保守的近似重复合并
lib/budget.js       子代理数估算、quick 降级、超限拒绝
lib/concurrency.js  共享 FIFO limiter
lib/dispatch.js     每个视角一个 finder;单个视角失败只损失该视角
lib/verify.js       独立复核者;失败处理按配置策略
lib/render.js       markdown 报告
lib/subagents.js    经 ctx.subagents 的受管 activation(`delivery: 'caller'`)
lib/skill.js        按需加载的技能 provider(rank 550)

测试

# 45 个单测(纯模块 + 真实临时 git 仓库)
node --test test/unit/*.test.mjs

# 脚本化端到端:真实 snapshot/gate/dedupe/verify/render 链路,子代理用脚本替代
node test/e2e/run.mjs

端到端用 test/fixtures/buggy-project(3 个预埋缺陷 + 2 个诱饵)证明:

编造的引用和 style kind 都被判为无效输出(5 个候选中 2 个);
两个真实缺陷被确认,<= 分页那条被复核推翻;
每条确认项的行号与引用行逐字一致;
干净跑仍然报覆盖;中断跑返回 schema 合法的部分结果;
纯文档范围会被拒绝。

运行时说明

实现跟的是当前安装的 Harness,而不是旧插件源码:

子代理 API —— ctx.subagents.startActivation({ provider, label, request, signal, delivery: 'caller' }),等待 activation.result,超时即 dispose。 delivery: 'caller' 避免每个子代理的完成通知都灌进父会话。
maxDepth 语义 —— 请求里的 maxDepth 是子代理的绝对深度上限,传 0 会直接拒绝创建子代理,所以默认 1。
schema 方言 —— parameters 与 output.schema 用标准 JSON Schema 的 required 数组;子代理 outputSchema 同样是标准 JSON Schema。

暂未包含

设置页、/review* 斜杠命令、报告落盘、CI 阻断、异构子代理模型,以及实测的误报率基线。插件不会自动修代码。

署名

本项目由 DeepSeek 构建,运行于 DeepSeek Harness。四个核心机制参考了公开文档化的 MIT 协议评审插件,但未复制任何源码。

dsh-review-kit

A code review plugin for DeepSeek Harness that refuses to report what it cannot prove.

_{by DeepSeek · built on DeepSeek Harness}

English · 中文

English

What it is

dsh-review-kit is a read-only, evidence-gated code review plugin for DeepSeek Harness. It registers one model-facing tool (review) and one on-demand skill (review-kit).

Why this one

#	Mechanism	What it does
1	Verbatim anchors	Every finding must quote one real line from the change snapshot. The plugin locates the line itself; a quote that does not exist is rejected as invalid output, never counted as a false positive.
2	Adversarial verification	Each surviving finding gets an independent verifier whose job is to refute it. The verifier never sees the finder's lens or reasoning — only the code and the claimed failure. Unverifiable means not confirmed.
3	Origin attribution	Verifiers classify `introduced` / `preexisting` / `unknown`. A baseline line match is only a hint, never a verdict.
4	Transparent coverage	The report always states which lenses ran, which failed, which files were skipped, which findings the budget dropped, and how many children ran. A clean run reports its scope instead of just saying "no issues".

Requirements

DeepSeek Harness 0.2.2-rc.1 (desktop profile tested), Node >= 22
The subagents service with a provider that supports structured output (spawn qualifies) — the plugin refuses to run rather than silently degrade
Git for the uncommitted / base / commit scopes (files works without it)

Install

Install the directory as a bundle in a Harness profile:

plugin_manager(action: "install_bundle", target: "<abs path>/dsh-review-kit")

The bundle row is declared in cordis.patch.yml. There is no build step and no @deepseek-ai/* import: the package is plain ESM plus Node builtins, so it works as a linked workspace bundle.

To pin configuration, override the row in the profile:

- id: review-kit
  name: '@local/dsh-review-kit'
  config:
    depth: quick
    maxConcurrentChildren: 4

Usage

Ask in plain language — "review this change" — or call the tool:

{ "target": "uncommitted" }                                  // default: worktree vs HEAD
{ "target": "base", "base": "origin/main" }                  // vs a baseline ref
{ "target": "commit", "commit": "<sha>" }                    // the change one commit introduced
{ "target": "files", "files": ["src/a.ts", "src/b.ts"] }     // explicit files
{ "depth": "quick" }                                         // <= 2 lenses, <= 4 verified
{ "lenses": ["security"], "maxFindings": 4 }                 // narrow scope

The model receives a markdown report; the canonical value keeps structured confirmed / refuted / unverified arrays plus coverage and stats for programmatic callers.

Read the report

Confirmed — file:line, the verbatim quoted line, the failure scenario, the verifier's conclusion and confidence, and the origin. These are the ones to fix.
Refuted — collapsed; the verifier's counter-evidence. Do not report these as problems.
Unverified — the verifier was unavailable. Not confirmed.
Coverage / skipped / stats — the honest part. ✅ ran, nothing found is not the same as ⚠️ failed. A review with coverage gaps must be reported as such.

Configuration

Field	Default	Meaning
`subagentProvider`	`spawn`	Registered `ctx.subagents` provider
`depth`	`full`	`full` or `quick`
`lenses`	`[]`	`[]` = all four (`correctness`, `lifecycle`, `contract`, `security`)
`verifiersPerFinding`	`1`	1–3; all must confirm
`maxFindings`	`12`	Findings that reach verification
`maxCandidatesPerLens`	`8`	Truncation only, not a failure
`maxChildren`	`24`	Estimate = lenses + findings × verifiers
`maxConcurrentChildren`	`6`	Shared limiter
`childTimeoutMs`	`300000`	Per finder/verifier child
`onVerifierFailure`	`refute`	`refute` (fail closed) or `observation`
`childMaxDepth`	`1`	`1` = children exist but cannot delegate further
`maxFiles`	`40`	Changed files read per review
`maxDiffChars`	`400000`	Prompt text budget
`gitMode`	`auto`	`auto` / `tool` / `process`

Architecture

lib/index.js        apply(): register `review` + the `review-kit` skill
lib/orchestrator.js target -> snapshot -> budget -> find -> gate -> dedupe -> cap -> verify -> render
lib/config.js       defaults, lens/kind/severity vocabularies, doc-file policy
lib/target.js       scope parsing + path-escape rejection
lib/git.js          read-only git through the caller's shell tool (or a process fallback)
lib/snapshot.js     before/after/diff per file, binary/large/doc skips, changed-line ranges
lib/lenses.js       four orthogonal lenses (constraints, not personas)
lib/prompt.js       finder + verifier prompts, prompt-budgeted file blocks
lib/schema.js       parameter schema, finder/verifier output schemas, tool output schema
lib/gate.js         kind whitelist + anchor existence + line resolution
lib/fingerprint.js  sha256(kind, path, normalised anchor)[0:16]
lib/dedupe.js       fingerprint merge + conservative near-duplicate merge
lib/budget.js       child-count estimate, quick degradation, refusal
lib/concurrency.js  shared FIFO limiter
lib/dispatch.js     one finder child per lens; a lens failure loses only that lens
lib/verify.js       independent verifiers; policy-driven failure handling
lib/render.js       markdown report
lib/subagents.js    managed activations via ctx.subagents (`delivery: 'caller'`)
lib/skill.js        on-demand skill provider (rank 550)

Tests

# 45 unit tests (pure modules + a real temporary git repository)
node --test test/unit/*.test.mjs

# scripted end-to-end: real snapshot/gate/dedupe/verify/render path, fake children
node test/e2e/run.mjs

The e2e harness runs the fixture project at test/fixtures/buggy-project (3 seeded bugs, 2 decoys) and proves:

a fabricated anchor and a style kind are rejected as invalid output (2 of 5 candidates),
the two real bugs are confirmed and the <= pagination finding is refuted,
every confirmed line number points at the exact quoted source line,
a clean run still reports coverage, an aborted run returns a schema-valid partial result,
a documentation-only scope is refused.

Runtime notes

The plugin follows the installed Harness runtime rather than older plugin sources:

Subagent API — ctx.subagents.startActivation({ provider, label, request, signal, delivery: 'caller' }), awaiting activation.result and disposing on timeout. delivery: 'caller' keeps one completion notice per child out of the parent conversation.
maxDepth semantics — a request's maxDepth is the absolute delegation depth cap for the child, so 0 refuses to create the child at all. The default is 1.
Schema dialect — parameters and output.schema use ordinary JSON Schema with required: [...] arrays; subagent outputSchema is standard JSON Schema too.

Not included yet

Settings page, /review* slash commands, persisted reports, CI blocking, heterogeneous child models, and a measured false-positive baseline. The tool never auto-fixes code.

Attribution

Built by DeepSeek with DeepSeek Harness. The four mechanisms were inspired by publicly documented, MIT-licensed review plugins; no source code was copied.

中文

这是什么

dsh-review-kit 是 DeepSeek Harness 的只读、证据门禁型代码评审插件。它注册一个模型工具（review）和一个按需加载的技能（review-kit）。

差异点

#	机制	作用
1	引用可校验	每条发现必须逐字引用快照里真实存在的一行。插件自己反查行号;引用的行不存在时记为「无效输出」,而不是误报。
2	对抗复核	每条存活发现都派一个独立复核者,任务就是推翻它。复核者看不到 finder 的视角和推理,只能自己读代码。复核不成立 = 不算确认。
3	归属区分	复核者判定 `introduced` / `preexisting` / `unknown`。基线里有同形行只是提示,不是结论。
4	覆盖透明	报告必然写明:哪些视角跑了、哪些失败、哪些文件被跳过、哪些发现被预算裁掉、用了多少子代理。干净结果也要说清范围,而不是一句「没问题」。

另外还有:kind 白名单(correctness | state | breaking | security)、指纹去重、按严重度消耗的预算与 quick 自动降级、共享并发 limiter、全链路取消透传。

前置条件

DeepSeek Harness 0.2.2-rc.1(desktop profile 实测),Node >= 22
subagents 服务,且 provider 支持结构化输出(spawn 满足);不满足时插件直接报错,不会静默降级
uncommitted / base / commit 范围需要 git(files 不需要)

安装

把目录作为 bundle 装进 Harness profile:

plugin_manager(action: "install_bundle", target: "<绝对路径>/dsh-review-kit")

bundle 行声明在 cordis.patch.yml。没有构建步骤,也不 import @deepseek-ai/*:整个包是纯 ESM + Node 内置模块,作为链接式工作区 bundle 直接用。

要固定配置,在 profile 里覆盖该行:

- id: review-kit
  name: '@local/dsh-review-kit'
  config:
    depth: quick
    maxConcurrentChildren: 4

用法

用自然语言说「审一下这次改动」即可,或直接调工具:

{ "target": "uncommitted" }                                  // 默认:工作区 vs HEAD
{ "target": "base", "base": "origin/main" }                  // 相对某个基线
{ "target": "commit", "commit": "<sha>" }                    // 某次提交引入的改动
{ "target": "files", "files": ["src/a.ts", "src/b.ts"] }     // 指定文件
{ "depth": "quick" }                                         // <= 2 个视角,<= 4 条复核
{ "lenses": ["security"], "maxFindings": 4 }                 // 收窄范围

模型拿到 markdown 报告;规范结果里另有结构化的 confirmed / refuted / unverified 数组与 coverage、stats,方便程序化调用。

怎么读报告

确认项 —— 文件:行、逐字引用的源码行、失败场景、复核结论与置信度、归属。这些才是要修的。
被推翻 —— 折叠显示,附带复核者的反证。不要把这里的东西当成问题汇报。
未验证 —— 复核者不可用。不等于已确认。
覆盖 / 跳过 / 统计 —— 最诚实的一段。✅ 已跑,无发现 和 ⚠️ 失败 是两回事; 有覆盖缺口就必须如实说明。

配置

字段	默认	含义
`subagentProvider`	`spawn`	已注册的 `ctx.subagents` provider
`depth`	`full`	`full` 或 `quick`
`lenses`	`[]`	`[]` = 全部四个(`correctness`、`lifecycle`、`contract`、`security`)
`verifiersPerFinding`	`1`	1–3;全部确认才算通过
`maxFindings`	`12`	进入复核的条数上限
`maxCandidatesPerLens`	`8`	只做截断,不算失败
`maxChildren`	`24`	估算 = 视角数 + 发现数 × 复核数
`maxConcurrentChildren`	`6`	共享并发上限
`childTimeoutMs`	`300000`	单个 finder/verifier
`onVerifierFailure`	`refute`	`refute`(宁漏不错)或 `observation`
`childMaxDepth`	`1`	`1` = 子代理存在,但不能继续委派
`maxFiles`	`40`	单次读取的改动文件上限
`maxDiffChars`	`400000`	提示词文本预算
`gitMode`	`auto`	`auto` / `tool` / `process`

gitMode: auto 会让 git 走调用方 agent 的 shell 工具(ctx.tools.execute), 从而保留部署的权限、沙箱与审计链路;process 直接跑 git,但只跑只读命令。

结构

lib/index.js        apply():注册 `review` 工具 + `review-kit` 技能
lib/orchestrator.js 目标 -> 快照 -> 预算 -> 找 -> 门禁 -> 去重 -> 截断 -> 复核 -> 渲染
lib/config.js       默认值、视角/种类/严重度词表、文档文件策略
lib/target.js       范围解析 + 路径逃逸拒绝
lib/git.js          经调用方 shell 工具的只读 git(或进程回退)
lib/snapshot.js     逐文件 before/after/diff,二进制/过大/文档跳过,改动行区间
lib/lenses.js       四个正交视角(约束,不是人设)
lib/prompt.js       finder + verifier 提示词,带预算的文件块
lib/schema.js       参数 schema、finder/verifier 输出 schema、工具输出 schema
lib/gate.js         kind 白名单 + anchor 存在性 + 行号反查
lib/fingerprint.js  sha256(kind, path, 归一化 anchor)[0:16]
lib/dedupe.js       指纹合并 + 保守的近似重复合并
lib/budget.js       子代理数估算、quick 降级、超限拒绝
lib/concurrency.js  共享 FIFO limiter
lib/dispatch.js     每个视角一个 finder;单个视角失败只损失该视角
lib/verify.js       独立复核者;失败处理按配置策略
lib/render.js       markdown 报告
lib/subagents.js    经 ctx.subagents 的受管 activation(`delivery: 'caller'`)
lib/skill.js        按需加载的技能 provider(rank 550)

测试

# 45 个单测(纯模块 + 真实临时 git 仓库)
node --test test/unit/*.test.mjs

# 脚本化端到端:真实 snapshot/gate/dedupe/verify/render 链路,子代理用脚本替代
node test/e2e/run.mjs

端到端用 test/fixtures/buggy-project(3 个预埋缺陷 + 2 个诱饵)证明:

编造的引用和 style kind 都被判为无效输出(5 个候选中 2 个);
两个真实缺陷被确认,<= 分页那条被复核推翻;
每条确认项的行号与引用行逐字一致;
干净跑仍然报覆盖;中断跑返回 schema 合法的部分结果;
纯文档范围会被拒绝。

运行时说明

实现跟的是当前安装的 Harness,而不是旧插件源码:

子代理 API —— ctx.subagents.startActivation({ provider, label, request, signal, delivery: 'caller' }),等待 activation.result,超时即 dispose。 delivery: 'caller' 避免每个子代理的完成通知都灌进父会话。
maxDepth 语义 —— 请求里的 maxDepth 是子代理的绝对深度上限,传 0 会直接拒绝创建子代理,所以默认 1。
schema 方言 —— parameters 与 output.schema 用标准 JSON Schema 的 required 数组;子代理 outputSchema 同样是标准 JSON Schema。

暂未包含

设置页、/review* 斜杠命令、报告落盘、CI 阻断、异构子代理模型,以及实测的误报率基线。插件不会自动修代码。

署名

本项目由 DeepSeek 构建,运行于 DeepSeek Harness。四个核心机制参考了公开文档化的 MIT 协议评审插件,但未复制任何源码。

XWH-0v0/dsh-review-kit

dsh-review-kit

English

What it is

Why this one

Requirements

Install

Usage

Read the report

Configuration

Architecture

Tests

Runtime notes

Not included yet

Attribution

中文

这是什么

差异点

前置条件

安装

用法

怎么读报告

配置

结构

测试

运行时说明

暂未包含

署名

Community comments

XWH-0v0/dsh-review-kit

dsh-review-kit

English

What it is

Why this one

Requirements

Install

Usage

Read the report

Configuration

Architecture

Tests

Runtime notes

Not included yet

Attribution

中文

这是什么

差异点

前置条件

安装

用法

怎么读报告

配置

结构

测试

运行时说明

暂未包含

署名

Community comments