ric-devflow

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

RIC DevFlow 入口

RIC DevFlow Entry

本入口负责准备和路由,不是第五个业务角色,不代替 Planner、Reviewer、Tester 或 Implementer,也不授予业务批准。保留用户原任务、约束、已有授权及进度,按以下顺序执行。
  1. 任务分类:范围内完整开发进入完整流程;显式
    ric-devflow-<role>
    或精确角色委派只准备该角色的必要闭包。解释、一般建议、独立只读审核/测试、单纯文档编辑、极小无风险编辑及无关 Skill 请求不隐式启动完整 DevFlow;遵从用户明确选择。
  2. 宿主与身份:依据实际宿主声明、可用工具及原生启动上下文确认 Codex、Claude Code 或 ZCode 和主/子会话;配置目录存在不作身份依据。不明确、不支持或身份矛盾时保留原任务并报告精确缺口。已启动子角色缺件时只回交宿主,不安装、不再派生代理。
  3. 安装范围:记录逻辑
    discovered_path
    、可信候选
    source_root
    和
    install_scope
    ;写入跟随逻辑发现位置或已确认范围,不能由 symlink 的 realpath 反推。仅源码路径且范围未知时,先明确范围再写入。
  4. 完整性与缺件守卫:先检查本入口内
    references/bootstrap.md
    及当前平台材料是否可读,未齐备前不要打开下文缺失链接。主会话缺件时,先确认发布源
    https://github.com/lichong-a/ric-dev-workflow-skills
    ,优先完整可信本地 checkout 或安装记录完整 SHA;无记录只向已确认源取一次候选并固定 SHA,不猜历史、不执行下载代码。按该 SHA 完整清单核对本次闭包全部现存包文件的类型和字节,入口本身也必须匹配;源不明、范围不明、候选不完整、定制/异版本、未知文件或断链均保持安装目标零写入并报告缺口。匹配后只从可信候选源读取 bootstrap 和当前平台材料,列出所需 Skill 闭包、必需原生定义/配置、父目录及权限,全部目标的冲突、明确禁用和权限预检通过前不复制任何缺件;通过后才排他创建并普通复制缺项,写前复核仍缺失、写后逐字读回,中断后重新比对再补缺。已启动子角色只回交宿主,不自行恢复。文件齐备后读取准备与恢复继续;完整开发需要五 Skill 与当前宿主四原生角色,独立角色仅检查入口、该角色及必要传递引用/配置,不要求无关角色。
  5. 补装与配置:仅在存在缺件或需核验配置时,按 bootstrap 读取当前平台一份参考:Codex、Claude Code 或 ZCode。固定可信候选 SHA、核对整个待修复闭包的现存包文件、预检全部目标后只补缺;冲突、禁用和权限拒绝按原边界停止,不执行下载安装代码。
  6. 加载核验:分别报告文件安装、宿主加载和实际调用。文件齐备不等于角色已加载;能继续就沿原任务前进。需要重载时给出最小恢复信息,不重复安装,不伪造调用成功或业务 Gate。
  7. 角色路由:准备就绪后才读取目标角色入口。Codex 的 Root Planner 保持直接调度;Claude/ZCode 主会话按原生平级调用转发,唯一 Planner 作决定并写状态,四子角色均不派生代理。无角色工具或目标未加载时报告缺口,不用通用代理冒充。
任务准备完成后读取Codex 原生 IDClaude/ZCode 原生 ID
完整开发或显式规划Planner
ric_devflow_planner
ric-devflow-planner
显式或委派审核Reviewer
ric_devflow_reviewer
ric-devflow-reviewer
显式或委派测试Tester
ric_devflow_tester
ric-devflow-tester
单个已批准 Task/实现类 DefectImplementer
ric_devflow_implementer
ric-devflow-implementer
表中未选择角色的链接仅是导航,不是独立角色的安装依赖。只按当前动作读取共享契约和领域材料;同一会话中已读且未变的内容无需重复读取。包准备完成不改变 G0–G10、职责分离、Spec/测试批准、真实 SHA 绑定、冻结 DAG 或 VERIFIED 到 DONE 条件。
This entry is responsible for preparation and routing. It is not a fifth business role, does not replace Planner, Reviewer, Tester, or Implementer, and does not grant business approval. Preserve the user's original task, constraints, existing authorization, and progress, and execute in the following order.
  1. Task classification: Complete development within scope enters the full process; explicit
    ric-devflow-<role>
    or precise role delegation only prepares the necessary closure for that role. Explanations, general advice, standalone read-only review/testing, documentation-only edits, minimal risk-free edits, and unrelated Skill requests do not implicitly start the full DevFlow; follow the user's explicit choice.
  2. Host and identity: Confirm Codex, Claude Code, or ZCode and the main/sub session based on actual host declarations, available tools, and native launch context; the existence of a configuration directory is not identity evidence. If unclear, unsupported, or conflicting identity, keep the original task and report the precise gaps. If an already-started sub-role is missing components, only hand back to the host; do not install or derive additional agents.
  3. Install scope: Record the logical
    discovered_path
    , trusted candidate
    source_root
    , and
    install_scope
    ; write following the logical discovery path or confirmed scope, not inferred back from symlink realpath. If only a source path exists and the scope is unknown, first determine the scope before writing.
  4. Completeness and missing-component guard: First check whether
    references/bootstrap.md
    in this entry and the current platform materials are readable; do not open missing links below before they are ready. When the main session is missing components, first confirm the release source
    https://github.com/lichong-a/ric-dev-workflow-skills
    , preferably a complete trusted local checkout or an installation record with a full SHA; if no record exists, fetch a candidate from the confirmed source once and pin the SHA, without guessing history or executing downloaded code. Verify the type and bytes of every existing package file in this closure against the complete manifest for that SHA; the entry itself must also match. If the source is unknown, the scope is unknown, the candidate is incomplete, customized/divergent, contains unknown files, or has broken links, keep all install targets zero-write and report gaps. After matching, read bootstrap and current platform materials only from the trusted candidate source, list the required Skill closure, required native definitions/config, parent directories, and permissions; do not copy any missing components until conflict, explicit-disable, and permission prechecks for all targets pass; after passing, exclusively create and normal-copy the missing items, re-verify they are still missing before writing, read back verbatim after writing, and re-compare before filling gaps if interrupted. Already-started sub-roles only hand back to the host and do not self-recover. After the files are complete, continue with Preparation and Recovery; full development requires all five Skills and the four native roles of the current host; a standalone role only checks the entry, that role, and necessary transitive references/config, not unrelated roles.
  5. Supplement and configuration: Only when components are missing or configuration verification is required, read one current-platform reference per bootstrap: Codex, Claude Code, or ZCode. Pin the trusted candidate SHA, verify the existing package files of the whole closure to be fixed, precheck all targets, then fill only missing items; stop at the original boundaries on conflicts, disablement, and permission denials, and do not execute download/install code.
  6. Load verification: Report file installation, host loading, and actual invocation separately. File completeness does not mean the role is loaded; if able to continue, proceed along the original task. When reload is needed, provide minimal recovery information, do not re-install, and do not fake successful invocation or business Gate.
  7. Role routing: Read the target role entry only after readiness. Codex's Root Planner remains directly scheduled; Claude/ZCode main sessions forward via native peer invocation, the sole Planner makes decisions and writes state, and the four sub-roles do not derive agents. If the role tool is unavailable or the target is not loaded, report gaps; do not impersonate with a generic agent.
TaskRead after preparationCodex native IDClaude/ZCode native ID
Full development or explicit planningPlanner
ric_devflow_planner
ric-devflow-planner
Explicit or delegated reviewReviewer
ric_devflow_reviewer
ric-devflow-reviewer
Explicit or delegated testingTester
ric_devflow_tester
ric-devflow-tester
Single approved Task/implementation DefectImplementer
ric_devflow_implementer
ric-devflow-implementer
Links for roles not selected in the table are just navigation, not installation dependencies for standalone roles. Read only the shared contracts and domain materials required by the current action; content already read and unchanged in the same session does not need to be read again. Package preparation completed does not change G0–G10, separation of duties, Spec/test approval, real SHA binding, frozen DAG, or VERIFIED to DONE conditions.