research
多 Agent 並行研究框架 - 多視角同時研究,智能匯總成完整報告
Source facts
- Repository
- miles990/multi-agent-workflow
- Last source activity
- May 12, 2026 at 10:52
- Detected SKILL.md language
- Chinese
- Stars
- 4
- Forks
- 0
Install options
The review-first prompt is selected by default. You can switch to a direct command or download a local copy.
Review the source files
Read SKILL.md and any companion files shown by SkillsMP before deciding whether to install.
File Explorer
100 filesShowing SKILL.md
SKILL.md
Source instructions · Read-only preview- name
- research
- version
- 3.2.0
- description
- 多 Agent 並行研究框架 - 多視角同時研究,智能匯總成完整報告
- triggers
- ["multi-research","parallel-research","多角度研究"]
- context
- fork
- allowed-tools
- ["Read","Grep","Glob","WebFetch","Write","Bash"]
- model
- sonnet
# Multi-Agent Research v3.2.0
> CT 模式偵測 → 多視角並行研究 → 交叉驗證 → CT 合規檢查 → 智能匯總 → Memory 存檔(自動 commit)
## 自動化機制
> ⚡ **本 skill 已整合 Claude Code Hooks**
>
> - Action logging、state tracking、git commit 均由 hooks 自動處理
> - 只需執行 CP1 初始化,其餘檢查點自動執行
## 使用方式
```bash
/multi-research [研究主題]
/multi-research AI Agent 架構設計模式 --deep
/multi-research --ct experiment "Multi-CT Pipeline 是否能降低 agent drift?"
/multi-research --no-ct "只要快速草稿"
```
**Flags**: `--perspectives N` | `--quick` | `--deep` | `--ct off|lite|strict|experiment` | `--no-ct` | `--no-memory`
## CT 模式(內建)
`/multi-research` 會在開始前執行 CT Escalation Router,自主選擇 CT 模式。
| Mode | 觸發情境 | 額外產物 |
|------|----------|----------|
| `off` | 使用者明確 `--no-ct` 或 `--ct off` | 無 |
| `lite` | 預設;一般研究、比較、方向整理 | synthesis 內的 evidence / uncertainty / drift guard notes |
| `strict` | 架構、風險、產品、技術決策、agent/memory/tool policy | `ct-stack.yaml`、`ct-compliance.md`、`risk-policy.yaml` |
| `experiment` | 論文、實驗、benchmark、evaluation、hypothesis、可重複驗證 | strict 產物 + `hypotheses.yaml`、`experiment-plan.md`、`eval-rubric.yaml`、`failure-modes.md`、`experiments/{topic-id}/` harness |
CT mode 會控制實際流程重量:
- `lite` 只做 evidence / uncertainty / drift guard,不產生 `ct-stack.yaml`,不跑 experiment harness,也不跑 autonomous upgrade。
- `strict` 才產生 CT stack、合規檢查與 retrospective;若發現系統性問題,只產生 self-upgrade proposal,不自動 patch。
- `experiment` 才產生可測假設、實驗設計、eval rubric、failure mining 與 experiment harness;autonomous patch 預設關閉,只收斂 proposal 與 closed-loop summary。
自動規則:
- 預設為 `lite`。
- `--deep` 會把最低模式升到 `strict`。
- 偵測到論文、實驗、benchmark、evaluation、hypothesis、變因、對照組等需求時升到 `experiment`。
- 偵測到「快速」「先粗略」「不用太詳細」會降級,除非使用者明確要求 experiment。
- 使用者 override 永遠優先。
→ Router:[shared/ct/escalation-router.md](_shared/ct/escalation-router.md)
→ Rules:[shared/ct/escalation-rules.yaml](_shared/ct/escalation-rules.yaml)
→ Runtime:[shared/ct/mode-runtime.yaml](_shared/ct/mode-runtime.yaml)
## 預設 4 視角
| ID | 名稱 | 模型 | 聚焦 |
|----|------|------|------|
| `architecture` | 架構分析師 | sonnet | 系統結構、設計模式 |
| `cognitive` | 認知研究員 | sonnet | 方法論、思維框架 |
| `workflow` | 工作流設計 | haiku | 執行流程、整合策略 |
| `industry` | 業界實踐 | haiku | 現有框架、最佳實踐 |
→ 模型路由配置:[shared/config/model-routing.yaml](_shared/config/model-routing.yaml)
## 執行流程
```
CP1: 工作流初始化 ⚡ 手動執行
python scripts/hooks/init_workflow.py --topic "{topic}" --stage RESEARCH
↓
Phase 0: CT Escalation Router → 偵測 ct_mode、理由、信心度
├── 檢查 user override: --ct / --no-ct
├── 依規則計分: lite / strict / experiment
├── 套用 downgrade guard: 快速 / 粗略 / 不用太詳細
└── 寫入 meta.yaml: ct_detection
↓
Phase 1: 北極星錨定 → 定義研究目標、成功標準
↓
Phase 2: Memory 搜尋 → 避免重複研究
↓
Phase 3: 視角分解 → 為每視角生成專屬 prompt + CT envelope
↓
Phase 4: MAP(並行研究)✅ 自動追蹤
┌──────────┬──────────┬──────────┬──────────┐
│架構分析師│認知研究員│工作流設計│業界實踐 │
└──────────┴──────────┴──────────┴──────────┘
[CP2/CP3 由 hooks 自動處理 Agent 狀態追蹤]
⚠️ **並行執行關鍵**:
在單一訊息中發送 4 個 Task 工具呼叫:
- Task({description: "架構視角", ...})
- Task({description: "認知視角", ...})
- Task({description: "工作流視角", ...})
- Task({description: "業界視角", ...})
這樣才能真正並行執行!
⚠️ **強制**:每個 Agent 必須在完成前執行:
1. mkdir -p .claude/memory/research/{topic-id}/perspectives/
2. Write → .claude/memory/research/{topic-id}/perspectives/{perspective_id}.md
未執行 Write = 任務失敗,工作流中止
↓
Phase 5: REDUCE(交叉驗證 + mode-scoped CT checks + 匯總)
↓
Phase 6: Memory 存檔 → 品質閘門檢查 → 存儲報告
↓
Phase 7: CT Retrospective(strict / experiment)
├── 檢查 selected_mode 是否正確
├── 比對 expected artifacts vs actual outputs
├── 挖掘 workflow failures / false positives / false negatives
└── 寫入 ct-retrospective.md
↓
Phase 8: Self-Upgrade Proposal(strict / experiment,必要時)
├── 若 mode 選錯、artifact 缺漏、gate 誤判、工具解析失敗
├── 產生 self-upgrade-proposal.md
└── runtime 變更需附驗證命令與 rollback plan
↓
Phase 9: Experiment Harness(CT-experiment 必須)
├── 產生 experiments/{topic-id}/cases.yaml
├── 產生 run-config.yaml / rubric.yaml
├── 至少對本次 condition 跑 score-run.py
├── 產生 results.jsonl
└── 產生 analysis.md
↓
Phase 10: Autonomous Upgrade Decision(experiment,proposal-only by default)
├── L1/L2: 只記錄 / 只提案,不修改
├── L3: 可修改 docs、templates、CT examples
├── L4a: 可修改非關鍵 runtime rules / validators,必須跑 focused smoke test
├── L4b/L4c: gates / workflow scripts 只提案,需人工批准
├── L5: 架構或降弱 gate 的變更停止並要求人工批准
└── 寫入 upgrade-decision.yaml;有 patch 時寫入 upgrade-report.md
↓
Phase 11: Closed-Loop Summary(成果收斂)
├── 彙整 research conclusion、CT mode review、upgrade decision
├── 彙整 quality gates / DAG / status / action log 驗證結果
└── 寫入 closed-loop-summary.md
↓
CP4: Task Commit ✅ 自動執行
[寫入 .claude/memory/ 時自動 git commit]
```
## CP4: Task Commit ✅ 自動
> **由 `post_write.py` hook 自動處理**
當 Write 工具寫入 `.claude/memory/` 目錄時,hook 會自動:
1. `git add .claude/memory/research/{topic-id}/`
2. `git commit -m "docs(research): complete {topic} research"`
3. 記錄 action 到 `actions.jsonl`
→ Hook 設定:[.claude/settings.local.json.template](../../.claude/settings.local.json.template)
→ 協議:[shared/checkpoints/mandatory-checkpoints.md](_shared/checkpoints/mandatory-checkpoints.md)
## 品質閘門
通過條件(RESEARCH 階段):
- ✅ 至少 2 視角達成共識
- ✅ 無未解決的關鍵矛盾
- ✅ 品質分數 ≥ 70
- ✅ CT-lite 以上需標記 evidence / uncertainty / drift guard
- ✅ CT-strict 以上不得有 HIGH CT 違規
- ✅ CT-experiment 需產出可測假設與實驗設計
- ✅ CT-experiment 需產出 experiment harness 並至少跑一次 scorer
CT gates 依 mode 分級:
- `CT_LITE`:只檢查 evidence / uncertainty / drift guard
- `CT_STRICT`:檢查 ct-stack、ct-compliance、HIGH 違規與 evidence coverage
- `CT_EXPERIMENT`:檢查 hypotheses、experiment plan、readiness 與 scorer result
→ 閘門配置:[shared/quality/gates.yaml](_shared/quality/gates.yaml)
## 早期終止
當 `consensus_rate >= 0.9` 時,可跳過衝突解決。
→ 配置:[shared/config/early-termination.yaml](_shared/config/early-termination.yaml)
## Context7 整合
自動偵測技術棧關鍵字(react, vue, fastapi 等)時,查詢最新文檔。
→ 配置:[shared/integration/context7.yaml](_shared/integration/context7.yaml)
## 輸出結構
```
.claude/memory/research/[topic-id]/
├── meta.yaml # 元數據,包含 ct_detection
├── ct-stack.yaml # CT-strict / experiment 產出
├── ct-compliance.md # CT-strict / experiment 產出
├── risk-policy.yaml # CT-strict / experiment 產出
├── hypotheses.yaml # CT-experiment 產出
├── experiment-plan.md # CT-experiment 產出
├── eval-rubric.yaml # CT-experiment 產出
├── failure-modes.md # CT-experiment 產出
├── claims/ # CT-strict / experiment structured claims
│ ├── architecture.claims.yaml
│ ├── cognitive.claims.yaml
│ ├── workflow.claims.yaml
│ └── industry.claims.yaml
├── ct-retrospective.md # CT 閉環產出
├── self-upgrade-proposal.md # 有改善建議時產出
├── upgrade-decision.yaml # 自主升級決策
├── upgrade-report.md # 有實際 patch 時產出
├── closed-loop-summary.md # 最終收斂成果報告
├── experiments/
│ └── {topic-id}/
│ ├── cases.yaml
│ ├── run-config.yaml
│ ├── rubric.yaml
│ ├── results.jsonl
│ ├── condition-comparison.md
│ └── analysis.md
├── perspectives/ # 完整視角報告(MAP 產出,保留)
│ ├── architecture.md
│ ├── cognitive.md
│ ├── workflow.md
│ └── industry.md
├── summaries/ # 結構化摘要(REDUCE 產出,供快速查閱)
│ ├── architecture.yaml
│ ├── cognitive.yaml
│ ├── workflow.yaml
│ └── industry.yaml
├── synthesis.md # 匯總報告(主輸出)
└── metrics.yaml # 階段指標
```
Mode-specific artifact rules:
- `lite` must not require `ct-stack.yaml`, `ct-compliance.md`, experiment harness, or self-upgrade artifacts.
- `strict` may produce `ct-retrospective.md` and `self-upgrade-proposal.md`, but must not apply autonomous patches.
- `experiment` must produce experiment artifacts and `closed-loop-summary.md`; autonomous patching remains proposal-only unless explicitly approved.
> ⚠️ perspectives/ 保存完整報告,summaries/ 保存結構化摘要,兩者都必須保留。
## Agent 能力限制
**視角 Agent 不應該開啟 Task**:
| 允許的操作 | 說明 |
|-----------|------|
| ✅ Read | 讀取檔案 |
| ✅ Glob/Grep | 搜尋檔案和內容 |
| ✅ Explore agent | 輕量級探索 |
| ✅ Bash | 執行命令 |
| ✅ WebFetch | 抓取網頁 |
| ✅ Write | 寫入報告 |
| ❌ Task | 開子 Agent |
## 網頁抓取策略
當需要抓取網頁時,使用以下順序:
1. **優先使用 WebFetch** - 快速、輕量
2. **如果 WebFetch 失敗**,使用 Chrome:
```
a. mcp__claude-in-chrome__tabs_create_mcp → 建立新分頁
b. mcp__claude-in-chrome__navigate → 導航到 URL
c. mcp__claude-in-chrome__get_page_text → 讀取內容
```
3. **如果仍然失敗**,記錄 URL 供人工處理
## 行動日誌 ✅ 自動
> **由 Claude Code Hooks 自動處理**
工具調用自動記錄到 `.claude/workflow/{workflow-id}/logs/actions.jsonl`。
**自動記錄的工具**:
| 工具 | 觸發 Hook | 記錄內容 |
|------|-----------|----------|
| Task | pre_task.py / post_task.py | Agent 啟動/完成狀態 |
| Write | post_write.py | 檔案路徑、Memory commit |
**排查問題**:
```bash
# 查看 RESEARCH 階段所有失敗行動
jq 'select(.stage == "RESEARCH" and .status == "failed")' \
.claude/workflow/{workflow-id}/logs/actions.jsonl
# 查看特定視角 Agent 的行動
jq 'select(.agent_id == "architecture")' \
.claude/workflow/{workflow-id}/logs/actions.jsonl
# 查看即時狀態
cat .claude/workflow/{workflow-id}/current.json | jq .
```
→ Hook 腳本:[scripts/hooks/](_scripts/hooks/)
→ 日誌規範:[shared/communication/execution-logs.md](_shared/communication/execution-logs.md)
## 共用模組
| 模組 | 用途 |
|------|------|
| [ct/escalation-router.md](_shared/ct/escalation-router.md) | CT 模式自動偵測 |
| [ct/escalation-rules.yaml](_shared/ct/escalation-rules.yaml) | CT 升級/降級規則 |
| [02-ct-mode/lite.md](02-ct-mode/lite.md) | CT-lite envelope |
| [02-ct-mode/strict.md](02-ct-mode/strict.md) | CT-strict envelope |
| [02-ct-mode/experiment.md](02-ct-mode/experiment.md) | CT-experiment envelope |
| [02-ct-mode/compliance.md](02-ct-mode/compliance.md) | CT compliance rules |
| [02-ct-mode/experiment-design.md](02-ct-mode/experiment-design.md) | 可驗證實驗設計 |
| [ct/experiment-harness/README.md](_shared/ct/experiment-harness/README.md) | 可重跑 experiment harness |
| [02-ct-mode/retrospective.md](02-ct-mode/retrospective.md) | CT 閉環檢查 |
| [02-ct-mode/self-upgrade-proposal.md](02-ct-mode/self-upgrade-proposal.md) | 自我改善提案 |
| [02-ct-mode/autonomous-upgrade.md](02-ct-mode/autonomous-upgrade.md) | 自主升級決策與執行 |
| [02-ct-mode/closed-loop-summary.md](02-ct-mode/closed-loop-summary.md) | 閉環成果收斂報告 |
| [ct/retrospective.md](_shared/ct/retrospective.md) | 共用 CT retrospective 規範 |
| [ct/autonomous-upgrade.md](_shared/ct/autonomous-upgrade.md) | 共用自主升級規範 |
| [ct/self-upgrade-policy.yaml](_shared/ct/self-upgrade-policy.yaml) | 自動化等級與安全規則 |
| [coordination/map-phase.md](_shared/coordination/map-phase.md) | 並行協調 |
| [coordination/reduce-phase.md](_shared/coordination/reduce-phase.md) | 匯總整合、大檔案處理 |
| [synthesis/cross-validation.md](_shared/synthesis/cross-validation.md) | 交叉驗證 |
| [quality/gates.yaml](_shared/quality/gates.yaml) | 品質閘門 |
| [config/model-routing.yaml](_shared/config/model-routing.yaml) | 模型路由 |
## 工作流位置
```
RESEARCH → PLAN → TASKS → IMPLEMENT → REVIEW → VERIFY
↑
你在這裡
```
研究結果可被 `plan` skill 引用,作為規劃的輸入。
View on GitHub