---
name: yunxiao-wedrive-agentic-collaboration
description: Use when the user asks to adapt AI multi-agent development collaboration into zhctprompt, especially code review agents, project-management/task-scheduling agents, Yunxiao kanban automation, or enterprise WeDrive-backed collaboration.
metadata:
  version: 0.1.0
---

# Yunxiao WeDrive Agentic Collaboration

## Purpose

Use this skill to run or design the two-agent collaboration model for `zhctprompt`:

- Code review agent: quality gate, `review_guardian`, review report, fix suggestions.
- Project management and task scheduling agent: Yunxiao task tree, kanban status, dependencies, owners, evidence, and blocker alerts.

This skill coordinates existing project skills rather than replacing them.

## Default Decision

The 2026-04-27 AI collaboration meeting fixed the baseline: do not migrate the organization from Yunxiao to Linear / Multica by default. Use Yunxiao as the collaboration base, use Codex / AI Agent as the execution layer, and keep human responsibility for scope tradeoff, acceptance criteria, launch judgment, permissions, tokens, sensitive data, and private-deployment boundaries.

The stable project knowledge page is `standards-stack/llm-wiki/ai-yunxiao-codex-collaboration/README.md`. Load it when the user asks why Yunxiao is the base, how Yunxiao and Codex should collaborate, or how to run a smart-canteen pilot from requirement to acceptance.

The user does not need to provide a formal mode, input, or target template. Treat ordinary Chinese requests, screenshots, chat excerpts, Yunxiao links, Codeup links, MR notes, WeDrive paths, and short task descriptions as raw intake. The agent must infer intent, normalize the request into internal modules, and then select the right workflow.

## Trigger Criteria

Use this skill when the user mentions:

- Shubham Saboo multi-agent development collaboration.
- AI 多智能体开发协作系统.
- 代码审查智能体.
- 项目管理智能体 or 任务调度智能体.
- 云效 + 微盘协作.
- 云效看板自动推进.
- AI 自动 Code Review.

Also use this skill when the user uses natural-language equivalents, such as:

- "帮我看看这个 MR 能不能合".
- "这个需求帮我拆一下并推进到云效".
- "根据微盘资料生成任务并安排人处理".
- "这些工作项哪些阻塞、哪些能并行".
- "帮我检查这个任务做完没、风险在哪里".
- "把这个聊天记录整理成云效任务".
- "完善开会流程 / AI协作流程".
- "会议纪要沉淀成流程".
- "Codex 沟通过程要留痕到项目文件".

## Required Reading

1. `standards-stack/prompt/governance/YUNXIAO_WEDRIVE_AGENTIC_COLLABORATION.md`
2. `standards-stack/prompt/governance/CODEX_AUTO_REVIEW_GUARDIAN.md`
3. `standards-stack/prompt/governance/YUNXIAO_COLLABORATION.md`
4. `standards-stack/prompt/governance/YUNXIAO_AGENT_REQUIREMENT_PLANNING.md`
5. `standards-stack/agent-skills/skills/company-wedrive-knowledge-base/SKILL.md`
6. `standards-stack/agent-skills/skills/yunxiao-cli-interaction/SKILL.md`
7. `standards-stack/agent-skills/skills/code-review-and-quality/SKILL.md`

## Natural Language Intake And Routing

Default behavior:

- Do not require the user to write "mode/input/target".
- Do not ask the user to choose between code review agent and project-management agent when the intent can be inferred.
- Build an internal `ROUTING_NOTE` from the user's words before acting.
- Ask a clarifying question only when a required external write, owner, credential, target repository, destructive action, or high-risk business decision is missing.

Internal `ROUTING_NOTE` fields:

```text
raw_input:
inferred_intent:
source_refs:
target_project_or_repo:
selected_mode:
selected_downstream_skills:
missing_required_inputs:
safe_next_action:
```

Routing rules:

| User intent signal | Selected mode | Downstream workflows |
| --- | --- | --- |
| MR, diff, branch, Code Review, "能不能合", "有没有风险" | Mode A: Code Review Agent | `CODEX_AUTO_REVIEW_GUARDIAN.md`, `code-review-and-quality`, `yunxiao-codeup-pr-association` when Codeup/MR is involved |
| 需求、任务拆解、排期、负责人、看板、阻塞、并行 | Mode B: Project Management And Task Scheduling Agent | `yunxiao-cli-interaction`, `yunxiao-requirement-task-lifecycle`, `yunxiao-batch-subagent-execution` when real Yunxiao work items are involved |
| 微盘资料、公司知识库、项目交付资料、方案/手册/接口/验收材料 | Use WeDrive knowledge lookup first, then Mode A or Mode B | `company-wedrive-knowledge-base`, then review or scheduling workflow |
| "实现这个云效需求", "拉取并完成任务", "修复这个工作项" | Lifecycle orchestration, then review | `yunxiao-requirement-task-lifecycle`, `karpathy-guidelines`, target repo rules, then Mode A review |
| 截图、聊天记录、简短中文需求要进云效 | Scheduling intake | `yunxiao-smart-canteen-requirements` when it targets `【自研】智慧营养健康餐厅` |
| 会议纪要、开会流程、AI 协作流程、Codex 留痕、自测、用户手册、操作视频 | Meeting and AI collaboration governance | `standards-stack/prompt/governance/MEETING_AND_AI_COLLABORATION_WORKFLOW.md`, then update durable rules when the workflow is reusable |

Modular prompt expansion:

1. Restate the inferred task contract briefly for yourself.
2. Split the task into source module, target module, action module, evidence module, and risk module.
3. Load only the skills needed for those modules.
4. Execute the smallest safe workflow that satisfies the inferred goal.
5. Preserve the generated `ROUTING_NOTE` or an equivalent summary in the evidence path when the task is material.

## Operating Modes

### Mode A: Code Review Agent

Use when a task needs review of code or MR readiness.

Steps:

1. Confirm task contract, target repo, branch, diff, tests, and evidence path.
2. Read related Yunxiao work item and Codeup MR through CLI/OpenAPI-backed paths when real external state is required.
3. Use WeDrive indexes only when the review depends on project docs, interface specs, manuals, delivery evidence, or historical requirements.
4. Run the review in this order:
   - scope match
   - tests and validation evidence
   - correctness and logic
   - security and privacy
   - performance
   - compatibility and rollback
   - manual/video gate if user-visible
5. Output `auto_pass_candidate`, `needs_review`, or `block`.
6. If fixes are needed, write suggested minimal fixes. Do not silently expand scope or auto-merge.

Required output:

- review decision
- findings grouped by severity
- exact files or evidence paths
- suggested fixes
- validation gaps
- cloud/Yunxiao/Codeup回填文本 when relevant

### Mode B: Project Management And Task Scheduling Agent

Use when a task needs decomposition, task assignment, kanban tracking, or blocker monitoring.

Steps:

1. Preserve source references: Yunxiao URL/ID, WeDrive source path/index row, chat note, meeting note, or user text.
2. Convert the input into a task contract with in-scope, out-of-scope, acceptance criteria, target repo, and evidence path.
3. Build a work item tree:
   - parent requirement
   - child requirements when needed
   - development tasks
   - test tasks
   - documentation/manual tasks
4. Mark dependencies and parallelism:
   - same API/DB/permission/menu/config/common component means serial until the shared contract is locked
   - independent repos or independent file sets may be parallelized with separate branches/worktrees
5. Prepare cloud write preview before any Yunxiao write.
6. If confirmed and credentials are available, use `yunxiao-cli-interaction` for real Yunxiao writes.
7. Update local recovery surfaces:
   - `tasks/QUEUE.md`
   - `HEARTBEAT.md`
   - `control/task-index/tasks.tsv`
   - task evidence summary under `work*/`

Required output:

- task tree or scheduling board
- owner/assignee status
- dependency map
- kanban state
- blocker list
- next safe actions

### Mode C: Meeting And AI Collaboration Governance

Use when a task needs meeting SOP, AI collaboration SOP, Codex traceability, or bug/new-demand routing rules.

Steps:

1. Preserve the meeting note or chat source in a project document when material.
2. Read `standards-stack/prompt/governance/MEETING_AND_AI_COLLABORATION_WORKFLOW.md`.
3. Classify work items:
   - bugs from customers, delivery, product, development, or other users do not need product review; developers create Yunxiao work items and fix them.
   - new requirements need product review; delivery records them in the online sheet first, product approves them, then Yunxiao receives the requirement.
4. Make meeting flow explicit:
   - pre-meeting goal,
   - recording,
   - host topic intro,
   - topic-based discussion,
   - host summary,
   - post-meeting task and evidence deposition.
5. Make AI collaboration flow explicit:
   - create or pull Yunxiao task,
   - discuss the solution with Codex and write the thinking into local project files,
   - self-test,
   - user manual,
   - operation video,
   - Yunxiao and local evidence backfill.
6. If the user taught a reusable rule, update `AGENTS.md`, `standards-stack/prompt/AGENTS.md`, a governance doc, or an existing skill so future threads can recover the behavior.

Required output:

- changed durable files,
- exact workflow or rule added,
- validation command result,
- any external automation still needing explicit confirmation.

## WeDrive Rules

- Start from `modules/project/README.md`, `work_company_knowledge/README.md`, and `work_company_knowledge/knowledge-map.md`.
- Use derived indexes to locate source files.
- Open only the selected source files required for the task.
- Do not bulk-copy WeDrive binaries into Git.
- Do not expose credentials, raw production payloads, certificate private keys, complete SQL dumps, or unmasked personal data.

## Yunxiao Rules

- All Yunxiao reads/writes use `yunxiao-cli-interaction`.
- If `config/local/yunxiao.env` is missing or incomplete, ask for the current colleague's personal Yunxiao token and name before writing.
- Browser UI is only for visual confirmation unless the user explicitly overrides the CLI-only rule.
- Code work for a Yunxiao requirement uses `yunxiao-requirement-task-lifecycle`.
- Batch work uses `yunxiao-batch-subagent-execution`.

## Done Criteria

The skill was applied correctly when:

- the selected mode is explicit,
- evidence and source paths are preserved,
- code review does not bypass `review_guardian`,
- task scheduling maps to Yunxiao and local queue/index surfaces,
- WeDrive is used as indexed source knowledge, not as a copied raw dump,
- blockers are explicit when automation cannot safely write external systems.