法典甲板流程

工程开发 社区

技能解读

当前原文的自动摘译,尚未人工审校。

强制激活合同

如果用户明确提及$codex-deck-flow或codex-deck-flow,请将其视为使用此工作流程系统的严格指令。当该请求处于活动状态时,请勿切换到直接临时编码作为主要路径。在激活后的第一个响应中,说明 codex-deck-flow 编排处于活动状态并说明您正在运行的第一个命令。如果用户为当前会话尚未创建、加载或已解析的工作流程提供 project root 加 workflow ID,则首先运行 resolve-workflow,然后继续运行 validate、reconcile 和 status。

严格的任务创建模式

当用户明确包含 Strict task creating mode 时,请将其视为位于正常 codex-deck-flow 创建流程之上的工作流程创作协议。规则:仍然像往常一样运行create,然后运行validate,然后运行show或status。草稿工作流 JSON 在create之后已经存在;批准仍然是trigger,而不是草稿创建。如果您必须在工作流程准备就绪之前向用户询问一个澄清问题,则助理消息必须以以下内容结尾:

核心运行时模型

一个工作流程 = ./.codex-deck/ 下的一个 JSON 文件。守护进程是编排所有者,可以为多个项目根的工作流提供服务。如果需要,该技能会启动守护进程,发送一条控制消息,然后停止。工作任务在隔离的git worktree目录中运行。调度程序连续性应使用 codex-deck-flow 非交互式应用程序服务器支持的恢复行为(scripts/codex_resume_noninteractive.py 行为模型),而不是守护进程模式下的原始 codex resume SESSION_ID 语义。

按意图路由

使用此决策路线,然后按照references/commands.md中的详细信息进行操作: 新工作流程请求:create -> validate -> show/status -> 用户批准 -> trigger 按已知路径的现有工作流程:validate -> reconcile -> status当前会话中路径未知时,按项目根 + 工作流 ID 表示的现有工作流: resolve-workflow -> validate -> reconcile -> status 显式守护进程控制消息: daemon-send 或 stop-workflow-processes

参考载荷边界

仅调度程序任务:references/scheduler.md调度程序/工作流任务指导文件规则和路由:references/guidance-routing.md非调度程序角色不应加载特定于调度程序的突变指导。

流狐档案 运行与权限由流狐静态检查推断,不代表作者声明或实际测试结果。
领域
工程开发
作者声明的智能体
未找到明确声明;不据此推断已兼容或已测试
静态检查
94 / 100 · 启发式扫描,不代表运行安全
运行环境证据
未执行;当前没有隔离运行测试证据
作者 / 版本 / 许可证
@asfsdsf · MIT
上下文用量评级
低消耗
接入复杂程度
即装即用
是否需要外部接口密钥
不需要
兼容的系统
未声明(默认跨平台)
运行环境要求
未检出明确要求
检测到的文件与系统行为
  • 只读
  • 允许写入 / 修改
网络行为
仅限本地
安装命令数
无(仅作为资料)

需要注意: 未声明工具权限范围;实际权限由运行环境决定。

作者说明文件(原文) SKILL.md

# Orchestrate Persistent Multi-Agent Workflows

Use this skill to create workflows, inspect workflows, and send control messages to the background codex-deck-flow daemon.

This skill is an **operator control client**, not the long-running orchestrator. After the daemon has started successfully and a control message has been sent successfully, the skill should stop.

Detailed operational guidance is routed through `references/*.md`:

- Command sequences and syntax: `references/commands.md`
- Workflow/task authoring rules: `references/authoring.md`
- Failure recovery runbooks: `references/runbooks.md`
- Scheduler vs workflow-task guidance placement: `references/guidance-routing.md`
- Architecture, contracts, and role prompts: `references/agent-flow-design.md`
- Scheduler mutation heuristics (scheduler task only): `references/scheduler.md`

## Mandatory activation contract

- If the user explicitly mentions `$codex-deck-flow` or `codex-deck-flow`, treat that as a strict instruction to use this workflow system.
- Do not switch to direct ad-hoc coding as the primary path while that request is active.
- In the first response after activation, say that codex-deck-flow orchestration is active and state the first command you are running.
- If the user gives `project root` plus `workflow ID` for a workflow that the current session has not created, loaded, or already resolved, first run `resolve-workflow`, then continue with `validate`, `reconcile`, and `status`.
- If the current session already knows the workflow path from `create`, a previous `resolve-workflow`, or earlier inspection in the same chat, reuse that path instead of resolving again.
- If the user already gave a workflow path, start with `validate`, then `reconcile`, then `status`.
- If no workflow exists yet, start with `create`, then immediately `validate`, then `show` or `status`, and present the authored tasks/prompts for user approval before `trigger`.
- If the user explicitly includes `Strict task creating mode`, keep the normal create/validate/show flow, but use the strict authoring protocol described below.
- Keep merge explicit. Never run `merge --apply` unless the user explicitly asked to apply the integration branch into the target branch.
- Once `trigger` or `daemon-send` succeeds, stop. Do not continue the same skill turn as if it were the orchestrator.

## Strict task creating mode

When the user explicitly includes `Strict task creating mode`, treat that as a workflow-authoring protocol layered on top of the normal codex-deck-flow create flow.

Rules:

- Still run `create`, then `validate`, then `show` or `status` as usual.
- The draft workflow JSON already exists after `create`; approval still gates `trigger`, not draft creation.
- If you must ask the user a clarifying question before the workflow is ready, the assistant message must end with exactly:
  - `<codex-deck-flow>Question<codex-deck-flow/>`
- When asking the user to approve the authored workflow, the assistant message must end with exactly:
  - `<codex-deck-flow>Request approve:workflow-file-name.json<codex-deck-flow/>`
- Replace `workflow-file-name.json` with the JSON file name only, for example `feature-delivery.json`.
- Use only the file name token after `Request approve:`. Do not add brackets or extra wrappers.
- Do not put any additional text after the strict tag suffix.
- Once the user approves, exit strict task creating mode and continue with the normal codex-deck-flow flow.
- If the approval message includes workflow data for an already-created draft, reuse that workflow and do **not** run `create` again.
- The approval workflow data may include fields such as `workflowPath`, `workflowFileName`, `workflowId`, `projectRoot`, and `boundSessionId`; treat that as the canonical handoff for reusing the already-authored draft.

## Core runtime model

- One workflow = one JSON file under `./.codex-deck/`.
- The daemon is the orchestration owner and can serve workflows from multiple project roots.
- The skill starts daemon if needed, sends one control message, then stops.
- Worker tasks run in isolated `git worktree` directories.
- Scheduler continuity should use codex-deck-flow non-interactive app-server-backed resume behavior (`scripts/codex_resume_noninteractive.py` behavior model), not raw `codex resume SESSION_ID` semantics in daemon mode.

Use `references/agent-flow-design.md` for full runtime contracts and role data.

## Route by intent

Use this decision route, then follow details in `references/commands.md`:

1. New workflow request: `create` -> `validate` -> `show/status` -> user approval -> `trigger`
2. Existing workflow by known path: `validate` -> `reconcile` -> `status`
3. Existing workflow by project root + workflow ID when path unknown in current session: `resolve-workflow` -> `validate` -> `reconcile` -> `status`
4. Explicit daemon control message: `daemon-send` or `stop-workflow-processes`
5. Merge request: `merge` preview, and `merge --apply` only if explicitly requested

## Reference loading boundaries

- Scheduler task only: `references/scheduler.md`
- Scheduler/workflow-task guidance file rules and routing: `references/guidance-routing.md`
- Non-scheduler roles should not load scheduler-specific mutation guidance.

## References

Canonical references:

- `references/commands.md`
- `references/authoring.md`
- `references/runbooks.md`
- `references/guidance-routing.md`
- `references/agent-flow-design.md`
- `references/scheduler.md` (scheduler task only)

Implementation scripts:

- `scripts/workflow.py`
- `scripts/task_runner.py`
- `scripts/daemon.py`
- `scripts/daemon_client.py`
- `scripts/codex_resume_noninteractive.py`
- `scripts/session_state.py`
- `scripts/session_summary.py`

讨论

基于 GitHub Discussions。登录 GitHub 即可参与讨论、点赞、订阅更新。