wishgraph

SKILL开发
开发 未标注权限 许可证不明确
安装这个技能

WishGraph 是一套面向 AI 辅助软件项目的"文件化治理"方法:它在人与模型这两个黑盒之间插入一个可编译、可审计的项目接口,用意图编译管线(Wish → Spec → Task → Worker → Validation → Report)把模糊需求转成有边界、可回滚的执行单元,并通过"讨论只决策、工人只执行"的角色分离、四维度状态机与"未激活即忽略"的激活门控,把 AI 行为关进可控、可审查的笼子里,每次执行都需回答推断意图、改动范围、验证方式与残留风险等反黑盒问题。落地后,它会把治理过程沉淀为持久的 PRD、架构文档、代码地图(CODEMAP.md)、带验证面的 Task Spec、独立可控且过程可见的 Worker 执行、不可篡改的 Run Report,以及讨论局域集成写回的项目当前状态快照——从而支持跨会话 / 跨模型 / 跨宿主的安全交接、沿"错误→状态→代码→规格"的因果调试,让项目不再因大量 AI 生成代码而变得不可解释,重新变得可理解、可审计、可回滚、可治理(支持中 / 英 / 双语项目)。

一句话选型结论

按资料字段整理

wishgraph是开发类的 Agent Skill,安装方式未标注,权限级别未标注,许可证不明确,有可直接复制的安装命令。技能本质是给模型读的说明书,效果取决于模型与客户端;装完先用一个小任务验证,再放进正式流程。

分类开发 安装方式未标注 权限级别未标注权限 许可证许可证不明确
累计安装 未标注 本周 未标注
GitHub 星标 未标注 仓库 未标注
下载量 未标注 技能市场下载数
权限级别 未标注权限 安装前请确认风险范围
许可证 许可证不明确 未说明

安装

未标注
在支持 Skills 的客户端里执行,装完问一句「你有哪些技能」验证
skillhub install user_2b88f683/wishgraph
  1. 1
    确认客户端支持

    确认你的客户端支持 Skills(Claude、Codex、Cursor 等),并升级到较新的版本。

  2. 2
    拿到技能文件

    执行上面的命令安装;也可以把仓库里的 SKILL.md 手动放进客户端的 skills 目录。

  3. 3
    刷新技能列表

    重新打开客户端,或刷新技能列表,让新放进来的技能被加载。

  4. 4
    验证是否生效

    问一句「你有哪些技能」,确认列表里出现它,再拿一个小任务试跑。

技能档案

开发
字段来自技能仓库与公开索引,中文说明由本站标注
作者 / 维护方user_2b88f683仓库 未标注 · 版本 未标注
宿主客户端未标注技能可被这些客户端加载
安装方式未标注安装包 community
权限级别未标注权限安装前请确认风险范围
许可证未说明许可证不明确,商用前请自行确认条款
仓库可用性available技能仓库当前是否可访问
最近核验未标注上游最后核验时间 · 数据更新 1791625257
正文体量10.2 KBSKILL.md 原文大小,越大说明说明书写得越细

典型使用场景

3 类
按分类整理的常见用法,具体以 SKILL.md 为准
写代码与改 bug

按需求补代码、定位并修掉报错。

读仓库与生成改动

让模型读懂项目结构,再给出可落地的改动。

跑测试与排查故障

按现象排查问题,补测试用例并复跑。

选择建议

装之前先看这几条
按条目字段给出的判断,不替你决定用不用

值得选的地方

  • 有可直接复制的安装命令,不用自己翻文档找入口。

先想清楚的地方

  • 没有标注权限级别,实际能做什么要装好后确认。
  • 许可证不明确,商用或二次分发前先到仓库确认条款。
  • 没有做站内核验,内容以作者仓库为准。
  • 技能只是说明书,实际效果取决于模型与客户端,装完先用小任务试一遍。

更适合

  • 需要在开发场景里给模型补一套做法的人。
  • 愿意先读一遍 SKILL.md、确认它会让模型做什么的人。

不太适合

  • 对许可证要求很严、必须明确可商用的人。
  • 只想装完就自动变强、不打算看说明书的人。

同类对比

6 个同类
同分类里挑最接近的几条,字段以各自详情页为准
技能权限级别安装方式累计安装宿主客户端安装命令
wishgraph当前 未标注权限 未标注 未标注 未标注 有
技能审查员 浏览器控制 ClawHub 市场 877.9 万 OpenClaw 有
Vercel React 最佳实践 浏览器控制 GitHub 仓库 849.1 万 Codex、Claude、Cursor 有
GitHub 代码协作 联网访问 ClawHub 市场 763.3 万 OpenClaw 有
网页设计指南 联网访问 GitHub 仓库 668.6 万 Codex、Claude、Cursor 有
本体论 只读 ClawHub 市场 663.7 万 OpenClaw 有

SKILL.md 全文

10.2 KB
模型实际读到的说明书原文,来自技能仓库

WishGraph

WishGraph

Overview

Use WishGraph to compile human intent into an auditable project-state graph:

Wish -> Spec -> Task -> Worker -> Validation -> Run Report -> Integration -> Review

Keep uncertainty, authority, evidence, and current project truth in repository files so work can
continue without chat history.

Project Activation

  • Treat global Skill installation as availability, never project activation.
  • Before routing generic entry phrases, read only the current Git root's .wishgraph/config.json.
  • WishGraph is active only when that file exists and mode is warn or enforce; missing config or mode: off means inactive.
  • In an inactive project, do not reinterpret 开始讨论, 刷新项目状态, or 执行 NNN 任务 as WishGraph commands. Do not read References, create project files, or install hooks.
  • Only an explicit request naming WishGraph, such as 使用 WishGraph or Use WishGraph, may begin first-time project setup.
  • On first activation, distinguish current_host from the project's required_hosts. Ask: “Which Agents should use WishGraph in this project?” Offer Codex + Claude Code (recommended), Codex only, or Claude Code only. Never silently turn the current host into the project scope.
  • First-time activation persists that host choice, completes setup atomically, and keeps the window neutral. Tell the user to reopen the current Agent session. Require a later explicit 开始讨论 / Start discussion event to enter Discussion.
  • Keep Doctor out of the normal first-use path. Route an explicit 检查 WishGraph 状态 / Check WishGraph status, or a reopened session that still does not respond, to the read-only Doctor before loading workflow context.
  • Route an explicit project update to the fingerprinted safe-upgrade path; preserve unknown or locally modified runtime files and ask before any forced replacement.
  • In an active project with a current runtime, repair only the explicitly requested host's missing or outdated adapter; changing required_hosts is a separate explicit action.

Role Boundaries

  • Treat every new host window as neutral. 开始讨论 enters Discussion for planning. In Discussion, an exact Task command routes an independent Formal Worker; in an ordinary neutral window, the same exact command binds that current window as the Worker. enforce requires Claim acquisition first; warn continues from the exact approved Task if Claim automation is unavailable. Never create a second Worker from the neutral direct-execution path.
  • Use discussion for planning, Task creation, Worker routing, integration, and result presentation.
  • Use worker only in a separate user-visible and inspectable Agent thread or window bound to one Task or Revision.
  • Treat Integration as an automatically triggered, Discussion-local Flow Phase, never a role or separate window.
  • Treat Review as result presentation inside Discussion, never a fourth Agent.
  • A Worker is explicitly authorized, inspectable, bounded by one Task, and produces structured terminal evidence. In enforce, it is also Claim-bound and mechanically gated. In warn, missing Hook, receipt, or Claim automation must not prevent a visible Worker from receiving and completing the Task.
  • Agent or subagent identity never creates authority. Never give Explorer, Reviewer, Helper, or Hidden/Internal Agents a Worker Claim.
  • Never let Discussion implement business code, run implementation builds/tests, or become the Worker fallback.

Core Workflow

For the default single-Task path, keep the user-visible loop to discuss -> approve -> independent Worker -> validated result. Hide Claim, lease, runtime, session, capability, and
runtime mechanics unless failure, recovery, or an explicit user request requires them.

  1. Inspect the repository and reuse compatible project truth instead of creating parallel memory.
  2. Compile intent into observable behavior, acceptance criteria, non-goals, and material decisions.
  3. Create only the minimum durable project-state graph and bounded executable work units.
  4. Classify work, record one exact expected transition, and obtain the required human authority.
  5. Route from Discussion to a separate inspectable Worker, or bind an explicitly addressed neutral window as the Worker; never let Discussion implement the Task.
  6. Require scoped implementation, prescribed validation, and immutable evidence; require Claim closeout when strict automation is active.
  7. Send every terminal result to Discussion-local integration automatically; ask only material decisions.
  8. Present the integrated result, evidence, residual risk, and next action without erasing history.

When warn is active and the host does not run Hooks, Discussion may send the exact approved Task directly to one user-visible, inspectable Worker. The Task scope, validation plan, allocated Run Report, and returned commit remain required; Hook receipts, Claims, and leases are optional automation on this path. Do not ask the user to restart repeatedly or move to another host merely to obtain a receipt.

Mandatory Gates

  • Resolve IDs exactly from structured state. Never prefix-match 012, 012b, and 012ba.
  • Interpret short approvals only when one unique structured expected_transition exists; explicit commands take priority.
  • Require explicit human authority before creating a Formal Worker. Persist waiting_for_worker only after a real stable thread/session ID is saved. Host limitations never expand Discussion authority.
  • In enforce, require a live Worker Claim bound to work unit, session, branch, absolute worktree, scope, and validation before implementation. In warn, keep Claim acquisition best-effort and continue from the approved Task when host automation is unavailable.
  • In enforce, treat the Git-common-dir execution Run as the canonical transient fact for dispatching, running, terminal evidence, and Integration. In warn, use it when available; otherwise the approved Task, immutable report, bounded result commit, and validation are the direct handoff. Session state, notifications, Task lifecycle labels, and host observations never override either evidence path.
  • When a warn Worker returns valid report and commit evidence without a usable Claim or Run callback, let Discussion invoke the existing Host Adapter terminal-result adoption path. It verifies the evidence, fills the same canonical Run when possible, and deduplicates the existing notification; never launch a replacement Worker or ask for authorization again.
  • In enforce, require the selected current host, current Adapter, and recent matching Hook receipt before Worker creation or Claim acquisition. In warn, report those facts only in explicit diagnostics and continue distribution.
  • Release or revoke any still-active acquired Claim before Worker rebind. Preserve stale Claim records as evidence, but never let them block a replacement Worker or carry old scope forward.
  • In enforce, require a reducer-issued one-time transition grant and bound Discussion-local Integration lease before merge resolution, combined validation, shared-state writes, or integration commit. In warn, try that path when available; otherwise Discussion verifies the visible Worker report, result commit, and validation, then integrates directly.
  • Keep shared project memory single-writer: Workers propose updates in Run Reports; Integration applies them.
  • Preserve immutable Task, attempt, Claim, Revision, Run Report, and losing competitive evidence.
  • Use write/build gate: required only for enforce. In warn, Hooks are advisory and must never deny ordinary tools. Describe source-read enforcement as read gate: host capability dependent unless the host can intercept every read path.
  • Treat hooks as mechanical gates and state readers, never as product decision-makers or hidden executors.
  • Run prescribed validation and the WishGraph checker before claiming completion; record skipped checks and residual risk.
  • Keep routine non-commit PreToolUse bounded to configured state and the requested operation; never enumerate the business source tree.

Reference Routing

Read only the references required for the current request:

  • Installation, prerequisites, safe/strict mode, upgrade, or recovery: references/installation.md.
  • Project concepts, naming, or public explanation: references/core-concepts.md.
  • New or existing project adoption, governance layout, templates, or bilingual memory: references/project-bootstrap.md.
  • Roles, phases, Task lifecycle, exact commands, approvals, and pure transitions: references/orchestration-state-machine.md.
  • Ordinary Task execution: read only the fast-path and closeout sections of references/worker-execution.md; use its recovery sections only after a Claim, session, takeover, or rebind failure.
  • Formal Task structure, quality criteria, and worked example: references/task-spec.md.
  • Clear low-risk correction: read only references/task-revisions.md; expand only through its exception table.
  • Multiple candidates for the same goal, scorecards, winner selection, and loser closeout: references/competitive-execution.md.
  • Automatic integration, leases, combined validation, shared-memory writeback, material decisions, or result presentation: references/integration-flow.md.
  • Bug triage or regression diagnosis: references/debug-causality.md.
  • Hook events, checker commands, host limits, performance baseline, or runtime troubleshooting: references/memory-sync-hooks.md.

常见问题

5 问
答案由本站依据条目字段整理
wishgraph怎么安装?

在支持 Skills 的客户端里执行:skillhub install user_2b88f683/wishgraph。安装方式为未标注。装完刷新技能列表或重开客户端。

wishgraph支持哪些客户端?

资料没有标注宿主客户端,但只要客户端支持 Skills(Claude、Codex、Cursor 等),一般都能加载。

wishgraph权限高吗,安全吗?

权限级别为未标注权限。安装前先看它会读取或写入哪些内容。

wishgraph收费吗?

本站不代收费用,也没有标注具体价格。是否收费取决于技能本身和所用的模型或客户端,正式使用前按作者说明确认。

wishgraph用什么许可证?

许可证标注为 未说明(许可证不明确)。商用或二次分发前请到仓库确认条款原文。

热门专题

按场景挑技能
全部专题 →

技能内容以作者仓库为准;分类、权限级别与许可证状态的中文说明为中文整理,最近核验 未标注。