forge

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

FORGE

Use this skill when you want agents to stop behaving like unpredictable pair-programmers and start working like disciplined teammates. FORGE gives AI work a real execution model: clear task boundaries, hard stops when things get fuzzy, safer commits, review gates, and a workflow teams can actually trust.

FORGE works best when you want to:

bootstrap or refresh
```
docs/forge/
```
without bloating the repo
execute the next task with bounded scope and explicit stop conditions
keep solo work checkpointed and committed one task at a time
coordinate multiple developers or IDE agents without losing track of ownership
add CI-backed enforcement when a project needs stronger guarantees

当你希望Agent不再像不可预测的结对程序员，而是像训练有素的团队成员那样工作时，请使用此skill。FORGE为AI工作提供了一套切实可行的执行模型：清晰的任务边界、遇到模糊问题时的强制终止、更安全的提交、评审关卡，以及团队真正可以信赖的工作流。

FORGE最适用于以下场景：

在不膨胀仓库的情况下搭建或更新
```
docs/forge/
```
文档
在限定范围和明确终止条件下执行下一项任务
让单人工作按任务逐个进行检查点记录和提交
在不混淆职责归属的前提下协调多名开发者或IDE Agent
当项目需要更强保障时，添加基于CI的执行约束

Quick Routing

快速导航

For creating or refreshing FORGE project docs, use
```
forge-bootstrap
```
.
For selecting and implementing the next bounded task, use
```
forge-execute-task
```
.
For scope and quality review before completion, use
```
forge-critique
```
.
For checklist-based security review, use
```
forge-security-review
```
.
For definition-of-done and evidence checks, use
```
forge-evaluation
```
.
For reading and updating reusable lessons, use
```
forge-memory
```
.
For private/public tool release workflows, use
```
forge-tool-workflow
```
.
For multi-developer coordination rules, branch discipline, and task claiming, read
```
references/team-mode.md
```
.
For repo-shape routing such as contract-first or tooling projects, read
```
references/repo-flavors.md
```
.
For agent-specific files such as
```
AGENTS.md
```
,
```
CLAUDE.md
```
, Cursor rules, Copilot instructions, Codex hooks, or Windsurf rules, read
```
references/agent-flavors.md
```
.
For DevSecOps gate profiles, repository hardening, CI/CD security, SCA, or SBOM controls, read
```
references/devsecops-gates.md
```
.
For human-facing application documentation (overview, architecture, threat model, developer guide, interfaces, deployment, runbooks, ADRs), read
```
references/application-docs.md
```
.

若要创建或更新FORGE项目文档，请使用
```
forge-bootstrap
```
。
若要选择并执行下一项限定范围的任务，请使用
```
forge-execute-task
```
。
若要在完成前进行范围和质量评审，请使用
```
forge-critique
```
。
若要基于清单进行安全评审，请使用
```
forge-security-review
```
。
若要进行完成定义和证据检查，请使用
```
forge-evaluation
```
。
若要读取和更新可复用经验，请使用
```
forge-memory
```
。
若要进行私有/公有工具发布工作流，请使用
```
forge-tool-workflow
```
。
关于多开发者协作规则、分支规范和任务认领，请阅读
```
references/team-mode.md
```
。
关于仓库形态相关指引（如契约优先或工具类项目），请阅读
```
references/repo-flavors.md
```
。
关于Agent专属文件（如
```
AGENTS.md
```
、
```
CLAUDE.md
```
、Cursor规则、Copilot指令、Codex钩子或Windsurf规则），请阅读
```
references/agent-flavors.md
```
。
关于DevSecOps关卡配置、仓库加固、CI/CD安全、SCA或SBOM管控，请阅读
```
references/devsecops-gates.md
```
。
关于面向人类的应用文档（概述、架构、威胁模型、开发者指南、接口、部署、运行手册、ADR），请阅读
```
references/application-docs.md
```
。

Operating Model

运行模型

Keep the runtime contract lean:

Read only the minimum project-local files needed for the current step.
Preserve bounded-task checkpoints at all times. If the project explicitly permits batch or auto behavior, that only means the agent may continue to the next task after finishing the current one, updating task state, and creating the required Conventional Commit.
Stop on ambiguity, missing prerequisites, architecture conflict, or unresolved security concerns.
Record evidence and memory updates in deterministic project-local files.

保持运行时契约简洁：

仅读取当前步骤所需的最少项目本地文件。
始终保留限定任务的检查点。如果项目明确允许批量或自动行为，仅意味着Agent在完成当前任务、更新任务状态并创建符合约定的提交（Conventional Commit）后，可以继续执行下一项任务。
遇到歧义、缺失前置条件、架构冲突或未解决的安全问题时终止工作。
在确定性的项目本地文件中记录证据和记忆更新。

Token Discipline

Token规范

Default to persistent low-token behavior across working responses.

Drop by default:

pleasantries and conversational filler
repeated context recap
narration of routine inspection or planning steps
file-purpose explanations when the file path already says enough
reasoning already captured in project docs or changed files
file content echo after writing files
changelog-style narration unless the user asks for it

Prefer by default:

short action updates
direct statements
fragments when they remain clear
file references over prose recap
outcome first, next step second

Response shape:

working update:

Status: <done/doing/blocker>. Next: <next step>.

task closeout:

Done: <result>. Changed: <files or areas>. Next: <next step or none>.

blocker:

Blocked: <fact>. Need: <decision or missing prerequisite>.

Soft caps:

working updates: 2 short lines max
normal task closeout: 4 short lines max
bootstrap closeout: 6 short lines max

Exceptions:

security warnings
destructive action confirmations
places where extra clarity prevents misread risk

在工作响应中默认采用持久化的低Token消耗行为。

默认省略：

客套话和对话填充内容
重复的上下文回顾
常规检查或规划步骤的叙述
文件路径已足够说明时的文件用途解释
已记录在项目文档或变更文件中的推理过程
写入文件后的文件内容回显
除非用户要求，否则不采用变更日志式的叙述

默认优先：

简短的动作更新
直接的陈述
保持清晰前提下的片段式内容
文件引用而非文字回顾
先说明结果，再说明下一步

响应格式：

工作更新：

Status: <已完成/进行中/阻塞>. Next: <下一步>.

任务收尾：

Done: <结果>. Changed: <文件或区域>. Next: <下一步或无>.

阻塞状态：

Blocked: <事实>. Need: <决策或缺失的前置条件>.

软限制：

工作更新：最多2行短内容
常规任务收尾：最多4行短内容
搭建任务收尾：最多6行短内容

例外情况：

安全警告
破坏性操作确认
额外清晰度可避免误读风险的场景

Team Mode

团队模式

When a repository is worked on by multiple developers or multiple IDE agents at the same time:

require feature branches rather than direct work on shared branches
require task claiming before implementation begins
require explicit
```
file_scope
```
for executable tasks
prefer append-only evidence and memory records over rewriting shared summaries
require CI enforcement and protected branch merging for completed task work

Read these shared references only when needed:

```
references/skill-pack-overview.md
```
```
references/doc-minimums.md
```
```
references/forge-to-skills-mapping.md
```
```
references/team-mode.md
```
```
references/repo-flavors.md
```
```
references/agent-flavors.md
```
```
references/devsecops-gates.md
```
```
references/application-docs.md
```
```
references/token-efficiency.md
```

Use templates from

assets/templates/

only when scaffolding or migrating a project into this skill-based flow.

当仓库由多名开发者或多个IDE Agent同时协作时：

要求使用功能分支，而非直接在共享分支上工作
要求在开始实现前认领任务
要求为可执行任务明确指定
```
file_scope
```
优先采用追加式的证据和记忆记录，而非重写共享摘要
要求已完成的任务工作需经过CI验证并通过受保护分支合并

仅在需要时阅读以下共享参考文档：

```
references/skill-pack-overview.md
```
```
references/doc-minimums.md
```
```
references/forge-to-skills-mapping.md
```
```
references/team-mode.md
```
```
references/repo-flavors.md
```
```
references/agent-flavors.md
```
```
references/devsecops-gates.md
```
```
references/application-docs.md
```
```
references/token-efficiency.md
```

仅在搭建项目或迁移至此基于skill的工作流时，使用

assets/templates/

中的模板。