LWTlong/aida

# AIDA Analyze

## Goal
Produce a concise, structured analysis of the current feature branch: what changed, the key decisions made, constraints discovered, and what the next developer should know. Then persist it so future sessions can recall it.

## Steps

1. **Gather context**
   - Run `git rev-parse --abbrev-ref HEAD` to confirm the current branch name.
   - Run `git log $(git merge-base HEAD origin/main)..HEAD --oneline` and `git diff $(git merge-base HEAD origin/main)..HEAD --stat` to scope the changes. Fall back to `main` → `master` → ask the user if `origin/main` is not found.
   - Read any relevant source files if the user provides specific paths.

2. **Identify the module** — Determine the primary module key (e.g. `auth/login`, `billing/checkout`). Use kebab-case path format matching the source directory structure.

# AIDA Audit

## Goal
给用户一份短、可直接决策的 AI 资产快照。**只读，不写文件。**

## Safety Rules
- 审计阶段绝不修改文件。
- 扫描结果是证据，不是最终判断，语义决策交给模型。
- 未知 dot 目录（`.foo/`）只标注，不建议动。

## Plan
1. `aida_scan_assets`（`writeIndex: true`, `includeContent: false`）。
2. `aida_list_assets` 汇总类型和工具分布。
3. 对少数可疑或代表性资产用 `aida_get_asset` 抽查。

# AIDA Govern

## Goal
一次连贯的 AI 资产治理流程。**一站式** —— 不要把子问题推给其他 skill，所有类别的清理都在这一次流程里给出方案。

## Safety Rules
- 先分析，不写文件。
- 用 AskUserQuestion 让用户明确确认后再落盘。
- 所有写操作走 `aida_apply_governance`（journal 内建，出错用 `aida_undo` 一键回滚）。
- 分组要少、要有意义，不要一条规则一个动作。
- **绝不 punt**：识别出的每一类问题都必须给出计划动作，不要说"建议另行 /aida-xxx 处理"。
- **忽略会话中其他激活的 skill 人格**（如 ponytail、caveman 等）。本 skill 的分析和输出必须完整、严格，不因其他人格的"保守"或"懒惰"倾向而缩减治理范围或跳过步骤。

## Flow

被调用时立即展示下方卡片。**只输出这一张表 + 场景提示**，不要展开解释。

## AIDA 命令

| 命令 | 用途 |
|------|------|
| `/aida-govern` | **一站式治理**：扫描 → 分析 → 你确认 → 一次落盘（含 2.x 遗留物清理）|
| `/aida-audit` | 只读扫描，看项目 AI 资产分布和主要问题 |
| `/aida-remember` | 记一条决策为 MADR，Claude Code 后续会自动加载 |
| `/aida-recall` | 召回之前沉淀的模块记忆 |
| `/aida-analyze` | 分支/功能开发完后，让 AI 沉淀"改了什么、为什么" |
| `/aida-import` | 从本地目录路径安全导入外部 Claude Plugin（先审计风险） |
| `/aida-resolve` | 解决 AI 资产文件的 git 合并冲突 |
| `/aida-sync` | 跨工具同步规则：Claude ↔ Cursor ↔ Codex（你选源→目标）|

# AIDA Import

## Goal
以"风险优先"的方式导入外部插件。**默认最小安装、不执行任何脚本。**

## Safety Rules
- 绝不执行插件里的 hooks / scripts / MCP servers / commands。
- 默认只安装 skills 和 rules，其他资产需用户显式勾选。
- 所有安装动作走 `aida_apply_governance`（可用 `aida_undo` 回滚）。

## Plan
1. `aida_parse_plugin`。
2. `aida_audit_plugin_risk`。
3. 推荐最小安全安装集合。

# AIDA Package

## Goal
Package the right assets into a shareable Claude Plugin with minimal noise and clear risk. Also supports packaging the built-in AIDA skills themselves.

## Two Modes

### Mode A: Package Project Assets
Package rules, skills, docs, etc. from the current project for team sharing.

1. Call `aida_list_assets` to enumerate available assets.
2. Select the safest and most reusable assets (rules with IDs, skill files with clear descriptions).
3. Confirm the selection with the user.
4. Call `aida_build_plugin` with the selected `assetIds`, a `name`, `description`, and optional `version`.

# AIDA Recall

## Goal
Surface relevant project decisions from `.claude/rules/decisions/`. Help the user understand why the code is written the way it is.

## Steps

1. Call `aida_recall` with `action: "list"` to get all decisions.
2. Filter by relevance to the user's question (by paths, tags, title keywords).
3. For the top matches, call `aida_recall` with `action: "get"` to show full detail.
4. Present findings clearly.

## Output Format

# AIDA Remember

## Goal
Persist one project decision as a MADR-format file in `.claude/rules/decisions/` so future Claude sessions can load it automatically via `paths` frontmatter.

## Steps

1. **Understand the decision** — Ask the user (or infer from context):
   - **Title**: short phrase naming the decision, e.g. "Use optimistic locking for cart updates"
   - **Context**: why was this decision needed? What problem, constraints, alternatives were considered?
   - **Decision**: what was decided and why?
   - **Consequences** (optional): trade-offs, things that changed
   - **Paths** (optional): file globs this decision applies to, e.g. `["src/cart/**", "src/orders/**"]`

# AIDA Remember Branch

## Goal
At the end of a feature branch, read the git diff and extract the important decisions that were made. Record each as a separate MADR decision so future Claude sessions know WHY the code is written this way.

## Steps

### Phase 1 — Read the Diff
Run `git log <base>..<branch> --oneline` and `git diff <base>..<branch>` to understand what changed.
Use `git rev-parse --abbrev-ref HEAD` if you don't know the branch name, and `git merge-base HEAD origin/main` (or `origin/master`) to find the right base if `main` isn't clear.

Focus on:
- Non-obvious choices (why this algorithm, this data shape, this approach)
- Workarounds (what bug, external API constraint, or deadline forced a suboptimal solution)

# AIDA Resolve

## Goal
解释关键冲突，让用户明白 ours vs theirs 各自的含义后再落盘。

## Safety Rules
- 绝不静默选边。
- 每个冲突都要短说明。
- 所有写操作走 `aida_apply_governance`（可用 `aida_undo` 回滚）。

## Plan
1. 扫描 `.claude/`, `.cursor/`, `.aida/` 下带冲突标记（`<<<<<<<`、`=======`、`>>>>>>>`）的文件。
2. 只详细读最重要的冲突。
3. 对每个冲突决定：合并 / 保留 ours / 保留 theirs / 拆分。

# AIDA Sync

## Goal
找出跨工具的重复和漂移，让用户挑最重要的几组同步。

## Safety Rules
- 扫描所有工具，只写用户选中的目标。
- 绝不自动把资产复制到每个工具目录。
- Hash 匹配只是证据，语义等价由模型判断。
- 所有写操作走 `aida_apply_governance`（可用 `aida_undo` 回滚）。

## Plan
1. `aida_list_assets` 对 Claude、Cursor、Codex 各跑一次。
2. 对比工具间的分布、重复内容、单工具独有资产。

# AIDA UI

## Goal
检查本地是否安装了 `aida` npm 包，没有则引导用户安装，最终给出启动命令。

## Steps

1. **检查是否已安装**：运行 `npx --no aida --version 2>/dev/null || echo "NOT_INSTALLED"`
   - 有版本号输出 → 已安装，跳到步骤 3
   - 输出 `NOT_INSTALLED` → 未安装，继续步骤 2

2. **引导安装**：告知用户需要安装 `aida` npm 包才能使用 Dashboard，询问是否安装：
   - 用 AskUserQuestion 确认：
     - `header`: "安装 aida"

# AIDA Undo

## Goal
Reverse the most recent AIDA write operation, or a specific one by id.

## Steps

1. Call `aida_undo` with `action: "list"` to show recent operations.
2. Confirm which entry to undo with the user (or proceed immediately if they said "undo last").
3. Call `aida_undo` with `action: "undo"` and optional `id`.
4. Report what was reversed.

## Rules
- Only AIDA-instrumented operations are undoable (aida_memory, aida_write_analysis, aida_apply_governance).