genie-orca-review

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

genie-orca:review

genie-orca:review

Runtime syntax: invoke the plugin copy through the active runtime's owner-qualified skill selector; use a bare selector only when intentionally selecting a user-tier copy (a separately installed personal copy; Genie no longer seeds this tier). Cross-skill prose below uses bare names as portable semantic routes; the orchestrator resolves the selector for the active runtime.
A reviewer is a read-only worker dispatched by the coordinator, never the engineer of the same group, preferably a different model family (brain default: Codex reviews Claude-Sonnet work; Fable/Opus for the two gate reviews).
运行时语法: 通过当前运行时的所有者限定技能选择器调用插件副本;仅当有意选择用户层级副本(单独安装的个人副本;Genie不再默认提供该层级副本)时,才使用裸选择器。以下跨技能说明使用裸名称作为可移植语义路径;编排器会为当前运行时解析选择器。
审查者是由协调器调度的只读工作者,绝不能是同一群组的工程师,最好使用不同的模型系列(默认配置:Codex审查Claude-Sonnet的工作;Fable/Opus用于两次gate审查)。

Contract

契约

  • Input: the group spec (
    WISH.md
    section), ground truth (
    SCOUT.md
    ),
    git diff <wish-branch>..HEAD
    , the validation command.
  • The reviewer re-runs validation and quotes the summary line. A review that did not run the gate is not a review.
  • Body starts with
    VERDICT: SHIP | FIX-FIRST | BLOCKED
    , then numbered findings
    [critical|major|minor]
    with
    file:line
    and a concrete fix. FIX-FIRST only on critical/major.
  • One
    worker_done
    ,
    --outcome succeeded
    = review delivered (the verdict is in the body). The reviewer's
    worker_done
    never authorizes coordinator edits; fixes are re-dispatched.
  • Name the environments in the adversarial question: the dev box with the product installed, the compiled binary from cwd
    /
    with no env, a DSN/detached service, the vault literally named like the product. On brain, the same precedence bug survived four gates because each ran in one environment.
  • Ask the adversarial question explicitly in the brief ("how does this still fail in the compiled binary / under DSN / on a box with brain installed?"). On brain, 3 of 5 groups went FIX-FIRST from exactly that prompt.
  • 输入:群组规范(
    WISH.md
    章节)、基准事实(
    SCOUT.md
    )、
    git diff <wish-branch>..HEAD
    、验证命令。
  • 审查者必须重新运行验证并引用摘要行。未运行gate的审查无效。
  • 正文以
    VERDICT: SHIP | FIX-FIRST | BLOCKED
    开头,随后是编号的问题项,标记为
    [critical|major|minor]
    ,包含
    file:line
    和具体修复方案。仅当问题为critical或major时,结论才为FIX-FIRST。
  • 发送一条
    worker_done
    消息,
    --outcome succeeded
    表示审查已完成(结论在正文中)。审查者的
    worker_done
    消息绝不授权协调器进行编辑;修复任务会被重新调度。
  • 在对抗性问题中明确指定环境:安装了产品的开发机、当前目录
    /
    下无环境变量的编译二进制文件、DSN/独立服务、与产品同名的vault。在Brain项目中,同一优先级bug通过了四次gate审查,因为每次都只在一个环境中运行。
  • 在简介中明确提出对抗性问题(“为什么在编译二进制文件/DSN环境/安装了Brain的机器上仍然失败?”)。在Brain项目中,5个群组中有3个正是因为这个提示得出了FIX-FIRST的结论。

Tiers

层级

WhenReviewersMerge rule
per group1 capable model (≠ engineer)verdict as-is
wish-approval, PR3 in parallel (claude / codex / third), same read-only worktreeseverity-max; any BLOCKED → BLOCKED; SHIP only if all SHIP; one merged Linear comment
councillenses (questioner / architecture / simplifier / perf …) on a decisionsynthesis + unresolved tensions, persisted next to the wish
retrothe run's
RETRO.md
from
skills/genie-orca-work/scripts/retro-collect.ts
findings → skill edits, not prose
场景审查者合并规则
每个群组1个合格模型(≠ 工程师)结论直接生效
愿望清单审批、PR3个模型并行(claude / codex / 第三个),使用相同的只读工作树取最高严重程度;任意一个BLOCKED则整体为BLOCKED;仅当所有模型均给出SHIP时才为SHIP;合并为一条Linear评论
Council针对决策的多视角审查者(提问者/架构师/简化师/性能专家…)综合结论 + 未解决的分歧,与愿望清单一同保存
Retro来自
skills/genie-orca-work/scripts/retro-collect.ts
的运行时
RETRO.md
问题项 → 技能编辑,而非文字描述

Fix loop

修复循环

Coordinator re-dispatches a fast worker into the same worktree with the findings quoted verbatim and "apply exactly this, nothing else". Cap 2 loops per group; the coordinator may verify a trivial delta itself instead of a second review. After the cap → human gate.
协调器将快速工作者重新调度到同一工作树,直接引用问题项并要求“严格按照此方案修复,不得修改其他内容”。每个群组最多允许2次循环;协调器可自行验证微小变更,无需二次审查。超过循环上限后 → 人工gate审查。

What the integrated gate catches that group review does not

集成gate审查能发现群组审查遗漏的问题

Run the full suite on the integrated branch before declaring SHIP — on brain, G4's artifact fallback passed its group gate and its review, and only the integrated gate (a box with
~/.brain
installed) exposed the cwd hijack. Per-group validation is necessary, not sufficient.
在宣布SHIP前,需在集成分支上运行完整测试套件——在Brain项目中,G4的工件回退通过了群组gate审查和AI审查,但只有集成gate审查(安装了
~/.brain
的机器)暴露了当前目录劫持问题。每个群组的验证是必要的,但并不充分。