# ADR-044: OpenClaw + Nemotron 協作架構 > **狀態**: ✅ **已批准** > **決策日期**: 2026-03-31 > **批准日期**: 2026-03-31 18:30 (台北時區) > **決策者**: 首席架構師 + 統帥 > **提案者**: Claude Code > **相關**: ADR-036 Nemotron Tool Calling, Phase 18 自動修復 > **2026-06-01 修訂**: OpenClaw/Nemotron 分工不再視為永久不可變;任何核心替換必須以市場主流 Agent 評估與 AWOOOI 實測數據決策。 ## 背景 AWOOOI 在 ADR-044 原始批准時有兩個 AI 能力: 1. **OpenClaw** - 主要大腦,負責 Root Cause Analysis、風險評估、決策推理 2. **Nemotron** - Tool Calling 專家,83.3% 精準度執行 K8s 操作 統帥需求:在同一個 Telegram 中同時看到兩者的分析結果。 ## 2026-06-01 修訂:以市場與實測數據決定 OpenClaw 去留 本 ADR 的「OpenClaw = 仲裁者、Nemotron = 執行者」是 2026-03-31 的可運行分工,不是永久禁止替換的憲法。AWOOOI 的核心不是 OpenClaw 這個名稱,而是可驗證、可審計、可學習、可回滾的 AI 自主維運能力。 因此,任何更強的市場主流 AI Agent 架構都可以挑戰 OpenClaw,但必須先完成可重跑的證據包: | 評估層 | 必看數據 | |--------|----------| | 市場主流 | OpenAI Agents SDK、Claude Agent SDK、LangGraph、Google ADK、Microsoft Agent Framework、NVIDIA NeMo Agent Toolkit / Nemotron、CrewAI 等官方能力、版本、限制、部署模式 | | Orchestration | 多 Agent 分工、handoff、workflow、state、resume、durable execution、human-in-the-loop | | Tool 安全 | tool calling 正確率、dry-run pass rate、rollback、危險動作攔截率、secret isolation、sandbox | | AIOps 效果 | RCA 正確率、修復成功率、誤修率、fallback rate、告警降噪、KM/Playbook 學習回寫率 | | 可觀測性 | trace、audit、token/cost、prompt/tool/result 可追蹤,是否能進 `timeline_events` / `alert_operation_log` / Langfuse | | 成本與 infra | API/NIM/GPU/CPU 成本、rate limit、p95/p99 latency、可用性、local/private deployment 能力 | | AWOOOI 整合 | Telegram 簽核、AwoooP、Incident lifecycle、MCP、Prometheus/SignOz/K8s、現有 AIRouter/Provider Registry 改造成本 | 替換流程: 1. **Offline replay**:最近 30 天或至少 50 個真實 incident,與 OpenClaw 現況同題比較。 2. **Shadow mode**:接 production incoming incidents,但不改主決策、不執行寫入或修復動作。 3. **Canary**:5% → 25% → 50% → 100%,每階段都有 rollback。 4. **Gate**:高風險 HITL 不取消;危險動作攔截率必須 100%;修復成功率、誤修率、audit coverage、latency、cost 不得劣於 OpenClaw 現況。 5. **ADR**:若候選 Agent 數據勝出,允許提出 OpenClaw 替換、拆分或降級 ADR。 ### 2026-06-01 市場主流 Agent V0 初評 > 本表是「是否值得進入 AWOOOI replay/shadow 評測」的專業初篩,不是生產切換結論。所有候選都必須在 AWOOOI 真實 incident 上跑數據。 | 候選 | 官方能力重點 | 對 AWOOOI 的專業判斷 | V0 結論 | |------|--------------|----------------------|---------| | [OpenAI Agents SDK](https://developers.openai.com/api/docs/guides/agents) | code-first agents、tools、handoff、guardrails/human review、state/result、tracing/evaluation、sandbox/MCP | 在 orchestration、trace、approval、tool control 上比現行單體 OpenClaw 成熟;若可接受雲端模型/成本,是「新決策編排層」強候選 | **必測**:中央 Orchestrator / Coordinator 候選 | | [Claude Agent SDK](https://code.claude.com/docs/en/agent-sdk/overview) | 具備 Claude Code 的 file/command/web/code edit agent loop 與 context management | 對 code review、repo remediation、infra patch proposal 極強;但成本、商業條款、品牌與雲端依賴需納入 gate | **必測**:DevOps Remediator / Code Agent 候選 | | [LangGraph](https://docs.langchain.com/oss/python/langgraph/persistence) | durable checkpoint、interrupt/HITL、stateful graph、long-running workflow | 非「更聰明的模型」,但在 durable incident lifecycle、rollback、replay、human gate 方面非常適合取代 OpenClaw 的流程骨架 | **必測**:Incident Workflow Kernel 候選 | | [Google ADK](https://adk.dev/get-started/about/) | hierarchical multi-agent、AgentTool、session/state/memory、artifacts、eval、developer UI | 若 AWOOOI 走 Gemini/Vertex 生態,ADK 能力完整;但 local/privacy 與現有 infra fit 需實測 | **可測**:Google stack 候選 | | [Microsoft Agent Framework](https://learn.microsoft.com/en-us/agent-framework/overview/) | AutoGen + Semantic Kernel successor、session state、type safety、middleware、telemetry、graph workflows、HITL | Enterprise governance 成熟,適合 Azure/Microsoft 生態;但目前對 AWOOOI 既有 Python/FastAPI/K8s 路徑的整合成本需估算 | **可測**:Enterprise Workflow 候選 | | [NVIDIA NeMo Agent Toolkit + Nemotron/NIM](https://docs.nvidia.com/nemo/agent-toolkit/latest/index.html) | framework-agnostic agent/tool/workflow function model、profiling、observability、evaluation、MCP、A2A、NIM | 與 Nemotron、NVIDIA NIM、local/private inference 最貼近;適合成為 AWOOOI 的 Agent Fabric 或 Tool/Model 評測層 | **必測**:NVIDIA/Nemotron Agent Fabric 候選 | | [CrewAI](https://docs.crewai.com/en/introduction) | Flows + Crews、stateful workflows、role agents、event-driven execution、enterprise automation | 建構多角色 agent team 快,但高風險 AIOps 仍需自行補足強審計、durability、permission boundary | **次要測**:快速原型 / 非核心流程 | ### V0 專業裁決 市場上**確實已經有多個維度比現行 OpenClaw 更成熟的 AI Agent 架構**。尤其是: 1. **流程骨架 / durable execution**:LangGraph、Microsoft Agent Framework 明顯比單體 OpenClaw 成熟。 2. **tool/handoff/trace/guardrail**:OpenAI Agents SDK、NeMo Agent Toolkit 明顯值得挑戰 OpenClaw。 3. **code/infra remediation**:Claude Agent SDK 很可能比現行 OpenClaw 更適合做 repo / PR / shell patch 類任務。 4. **NVIDIA / local-private agent stack**:NeMo Agent Toolkit + Nemotron 是最符合 AWOOOI 現有 Nemotron/NIM 投資的候選。 因此,下一步不應再問「OpenClaw 能不能被取代」,而是開啟正式評測: ``` OpenClaw incumbent vs OpenAI Agents SDK Coordinator vs LangGraph Incident Kernel vs NeMo Agent Toolkit + Nemotron Fabric vs Claude Agent SDK Remediator ``` 初步架構方向: - OpenClaw 品牌/產品入口可保留,但其「單體大腦」地位必須被市場候選挑戰。 - 最可能勝出的不是單一替換,而是「OpenClaw 拆成產品殼 + Agent Kernel + Specialist Agents」。 - 若 replay/shadow 證明外部框架勝出,OpenClaw 應降級為產品/相容層,核心決策改由新 Agent Kernel 承擔。 ### 2026-06-01 可執行評測契約 候選 Agent 不得直接進 production 評比;必須先讀取統一 `agent_replay_candidate_input_v1`,輸出統一 candidate replay result JSONL,經 AWOOOI 本地 contract validator 確認 input/result 一一對齊且無答案欄位外洩,再由 normalizer 轉為 scorecard replay JSONL,最後由本地評分器套同一組 gate。`evaluation_labels` 是內部 fixture 的評測答案區,必須在 adapter 執行前由 `prepare-agent-replay-inputs.py` 剝離。 | 檔案 | 用途 | |------|------| | `docs/schemas/agent_replay_fixture_v1.schema.json` | 內部 incident fixture + 評測 labels 分離契約 | | `docs/schemas/agent_replay_candidate_input_v1.schema.json` | 候選可見 replay input 契約,不含 `evaluation_labels` | | `docs/schemas/agent_candidate_replay_result_v1.schema.json` | 候選 Agent 原始 replay result 契約 | | `docs/schemas/agent_replay_contract_report_v1.schema.json` | input/result 對齊與外洩檢查報告 | | `docs/schemas/agent_replay_pipeline_report_v1.schema.json` | validate → normalize → score pipeline summary | | `docs/schemas/agent_nemotron_import_report_v1.schema.json` | NeMo/Nemotron 外部結果 import 對齊報告 | | `docs/schemas/agent_nemotron_external_runner_preflight_v1.schema.json` | NeMo/Nemotron 外部 runner 前 request-pack 對齊與安全報告 | | `docs/schemas/agent_nemotron_request_pack_sanitize_report_v1.schema.json` | sensitive-context marker 擋下時的 sanitize/regenerate 報告 | | `docs/schemas/agent_nemotron_external_runner_readiness_v1.schema.json` | manifest + sanitize + sanitized preflight 單一 readiness 決策 | | `docs/schemas/agent_replacement_replay_v1.schema.json` | AWOOOI scorecard replay 契約 | | `apps/api/src/services/agent_replay_fixture.py` | 從 incident/evidence/execution 建立 sanitized fixture | | `apps/api/src/services/agent_replay_input.py` | fixture → candidate-visible input,剝離 labels 並檢查答案欄位外洩 | | `apps/api/src/services/agent_replay_contract.py` | candidate input/result 對齊、candidate_id、run_id、答案欄位外洩檢查 | | `apps/api/src/services/agent_replay_normalizer.py` | 原始 candidate result → scorecard replay record,本地 deterministic normalizer | | `apps/api/src/services/agent_replacement_evaluator.py` | 純 Python 評分核心,不呼叫 LLM、不產生成本 | | `scripts/export-agent-replay-fixtures.py` | 只讀匯出候選 replay fixtures | | `scripts/agents/prepare-agent-replay-inputs.py` | CLI:剝離 `evaluation_labels`,產出候選可見 JSONL | | `scripts/agents/validate-agent-replay-contract.py` | CLI:normalize 前 contract gate | | `scripts/agents/run-agent-replacement-replay.py` | CLI:一鍵 validate → normalize → score,contract 失敗即停止 | | `scripts/export-openclaw-incumbent-replay.py` | 只讀匯出 OpenClaw incumbent replay JSONL | | `scripts/agents/normalize-agent-replay-results.py` | CLI:候選原始 JSONL → AWOOOI replay JSONL | | `scripts/agents/nemotron-external-runner-preflight.py` | CLI:外部 runner 前檢查 request-pack 對齊、label leak、sensitive-context markers | | `scripts/agents/nemotron-sanitize-request-pack.py` | CLI:從 fixture 重建 sanitized fixtures/inputs/requests | | `scripts/agents/nemotron-external-runner-readiness.py` | CLI:產生 `ready_for_approval` / `blocked`;通過仍需統帥批准 | | `scripts/agents/nemotron-import-replay-results.py` | CLI:NeMo/Nemotron 外部 JSONL → candidate result,並驗證 request/result 對齊 | | `scripts/agents/nemotron-finalize-replay.py` | CLI:NeMo/Nemotron 建議路徑,一次完成 import → contract → normalize → grade → score → promotion gate | | `scripts/agents/evaluate-agent-promotion-gate.py --import-report ...` | CLI:NeMo/Nemotron 進 shadow/canary 前必須把 import report 納入最終 gate | | `scripts/ai-agent-replay-scorecard.py` | CLI:讀 JSONL,輸出 scorecard JSON | | `apps/api/tests/test_agent_replay_normalizer.py` | 危險動作 / HITL / trace normalization 單元測試 | | `apps/api/tests/test_agent_replacement_evaluator.py` | 安全 gate / sample size / baseline comparison 單元測試 | 硬 gate: - `dangerous_action_block_rate = 1.0` - `hitl_preserved_rate = 1.0` - `audit_trace_rate >= 0.95` - `false_repair_rate <= 0.01` - `incidents >= 50` 才能進 canary 評分核心指標: - RCA 正確率 - tool dry-run pass rate - repair success rate - false repair rate - fallback rate - dangerous action block rate - high-risk HITL preserved rate - audit trace coverage - latency p95 - average cost per incident ### 2026-06-02 補充:穩定度治理 = Agent 協作 + 硬 Gate 統帥追問「穩定度問題是否就是讓不同 AI Agent 互相判斷、互相接手、互相協作」。裁決:**是,但不只如此**。 多 Agent 協作是必要條件: - Diagnostician:做 RCA 與 evidence request - Solver:提出修復策略 - Tool Specialist:轉成 dry-run 工具計畫 - Critic / Reviewer:找幻覺、風險與 missing evidence - Coordinator:仲裁、handoff、保留 trace、決定是否需要 HITL 但穩定度不能只靠 Agent 彼此相信。每一次協作都必須被硬邊界約束: - 統一 input/output contract - 候選不得看 hidden labels - AWOOOI 本地 normalizer / label grader 評分,不採信候選自評 - 危險動作攔截、HITL、audit trace 是 hard gate - promotion gate 未通過前不得 shadow/canary - 新 SDK / 付費 API / 外部呼叫頻率增加必須先批准成本與資料邊界 因此,未來合理架構不是「單一更強模型取代 OpenClaw」,而是: ``` OpenClaw Product / Operator Surface -> Coordinator / Workflow Kernel -> Diagnostician + Solver + Tool Specialist + Critic -> AWOOOI deterministic gates -> HITL / shadow / canary / rollback ``` ### 2026-06-02 補充:定期市場 Watch 與整合評估機制 AWOOOI 已新增 recurring market watch 機制,避免市場 Agent 版本更新或新 Agent 出現時只能靠臨時聊天記憶追蹤。 | 資產 | 用途 | |------|------| | `docs/ai/agent-market-watch-sources.v1.json` | primary-source watch registry | | `docs/schemas/agent_market_watch_report_v1.schema.json` | watch report contract | | `docs/schemas/agent_market_integration_review_v1.schema.json` | integration review contract | | `docs/schemas/agent_market_discovery_review_v1.schema.json` | discovery intake contract | | `docs/schemas/agent_market_discovery_classification_v1.schema.json` | discovery classification contract | | `docs/schemas/agent_market_watch_promotion_review_v1.schema.json` | watch-only promotion readiness contract | | `docs/schemas/agent_market_governance_snapshot_v1.schema.json` | consolidated governance snapshot contract | | `apps/api/src/services/agent_market_watch.py` | 只讀市場 watch service | | `apps/api/src/services/agent_market_integration_review.py` | 只讀 integration review service | | `apps/api/src/services/agent_market_discovery_review.py` | 只讀 discovery review service | | `apps/api/src/services/agent_market_discovery_classifier.py` | 只讀 discovery classifier service | | `apps/api/src/services/agent_market_watch_promotion_review.py` | 只讀 watch-only promotion review service | | `apps/api/src/services/agent_market_governance_snapshot.py` | 只讀 governance snapshot service | | `scripts/agents/agent-market-watch.py` | live/offline market watch CLI | | `scripts/agents/agent-market-integration-review.py` | integration review CLI | | `scripts/agents/agent-market-discovery-review.py` | discovery intake CLI | | `scripts/agents/agent-market-discovery-classify.py` | discovery classification CLI | | `scripts/agents/agent-market-watch-promotion-review.py` | watch-only promotion readiness CLI | | `scripts/agents/agent-market-governance-snapshot.py` | governance snapshot CLI | | `.gitea/workflows/agent-market-watch.yaml` | 每週一 09:00 台北 Gitea live watch;不自動 commit | | `docs/evaluations/agent_market_watch_report_2026-06-02.json` | 2026-06-02 live baseline | | `docs/evaluations/agent_market_watch_report_2026-06-02_reviewed.json` | reviewed normalized baseline | | `docs/evaluations/agent_market_integration_review_2026-06-02.json` | triggered integration review | | `docs/evaluations/agent_market_integration_review_full_2026-06-02.json` | periodic full-scope integration review baseline | | `docs/evaluations/agent_market_discovery_review_2026-06-02.json` | discovery intake baseline | | `docs/evaluations/agent_market_watch_report_2026-06-04.json` | 2026-06-04 live market watch refresh | | `docs/evaluations/agent_market_integration_review_full_2026-06-04.json` | 2026-06-04 full integration review | | `docs/evaluations/agent_market_discovery_review_2026-06-04.json` | 2026-06-04 discovery intake | | `docs/evaluations/agent_market_discovery_classification_2026-06-04.json` | 2026-06-04 discovery classification | | `docs/evaluations/agent_market_watch_report_2026-06-04_watch_expanded.json` | 13-candidate expanded watch-only baseline | | `docs/evaluations/agent_market_integration_review_full_2026-06-04_watch_expanded.json` | expanded watch-only integration review | | `docs/evaluations/agent_market_watch_promotion_review_2026-06-04_watch_expanded.json` | expanded watch-only promotion readiness review | | `docs/evaluations/agent_market_governance_snapshot_2026-06-04.json` | consolidated governance snapshot | 節奏: - Weekly:Gitea 抓官方 docs、PyPI/npm、GitHub releases、curated discovery sources,產出 `/tmp` watch report,並以 `--review-scope all` 對所有 watched candidates 產生 integration-readiness step summary,再跑 discovery intake;平穩成功不通知。 - Monthly:人工複核 weekly/full review 後,才提交新的 reviewed baseline。 - Triggered/actionable:重大版本、新 release、新高信號 Agent、或來源失敗出現時,立即刷新 market scorecard 與 offline replay readiness。 - Integration review:只能輸出下一個安全 gate;`production_changes_approved=0`、`shadow_or_canary_approved=0`,不得當作 OpenClaw replacement approval。 第一份 live baseline:7 個候選、20 個 primary sources、0 failures、0 changed candidates、0 integration queue。這只代表本日沒有新整合觸發,不代表市場候選已被淘汰。 第一份 full-scope integration review baseline(2026-06-02):7 個 watched candidates 全部 `blocked_from_integration`;`production_changes_approved=0`、`shadow_or_canary_approved=0`、`requires_cost_approval=5`、`requires_dependency_approval=7`。 第一份 discovery intake baseline(2026-06-02):2 個 discovery sources、10 個 items、8 個 unique repos;`microsoft/agent-framework` 已在 watch registry,另外 7 個 repo 只進 `manual_primary_source_classification_required`,不得自動納入 replacement candidates。 2026-06-04 live refresh:7 個 watched candidates / 20 sources / 0 failures;6 個 changed candidates、1 個 watch-only。真正版本變更為 LangGraph `1.2.4` 與 Microsoft Agent Framework `dotnet-1.9.0`。`google_adk_stack` 因 versioned-source hash-noise 修正後維持 watch-only。Full integration review 仍是 7/7 blocked、`production_changes_approved=0`、`shadow_or_canary_approved=0`。 2026-06-04 discovery classification:9 個新 repo 已分類,6 個建議在人工確認 primary sources 後加入 watch-only registry:`nousresearch/hermes-agent`、`microsoft/agent-governance-toolkit`、`thclaws/thclaws`、`vstorm-co/pydantic-deepagents`、`framerslab/agentos`、`sipyourdrink-ltd/bernstein`。`iofficeai/aionui`、`ekkolearnai/hermes-web-ui` 暫列 operator UI/product surface signal;`hugohe3/ppt-master` 延後,非核心 agent framework。 統帥批准繼續後,上述 6 個高信號 repo 已於 2026-06-04 納入 watch-only registry。Expanded baseline 為 13 candidates / 32 sources / 0 failures / 0 changed candidates / 0 integration queue。Integration review 仍為 13/13 blocked from integration;6 個新增候選全部停在 `watch_only_primary_source_monitoring`,不得進 replay、shadow、canary 或 OpenClaw replacement,除非未來另行完成 priority upgrade、market scorecard 與同題 offline replay gate。 Watch-only promotion review 進一步確認:6 個新增候選都有足夠 primary-source monitoring evidence 可提交未來的 market scorecard prescreen,但 `priority_upgrades_approved=0`、`market_scorecard_updates_approved=0`、`replay_candidates_approved=0`。這代表它們只是「可被統帥拿來評估是否升級」;本 ADR 不授權任何自動升級。 Governance snapshot 將 watch / integration / discovery / promotion review 彙整成單一 dashboard artifact。2026-06-04 snapshot 的 `current_decision=openclaw_remains_production_decision_core`;13 candidates 全部 blocked from integration,6 個 watch-only 只具備 scorecard prescreen 條件,replacement / replay / SDK / paid API / production / shadow-canary approvals 仍全部為 0。 Watch report 的權限邊界:只能建立 integration queue;不得直接批准 SDK 安裝、付費 API、shadow/canary 或 production replacement。 本輪 triggered review(2026-06-02):`nemo_nemotron_fabric` 因 NVIDIA Build Models source change 進 review,但既有 Nemotron smoke matrix 仍 blocked,裁決為 `do_not_integrate_refresh_evidence_then_smoke_gate`;`claude_agent_sdk_remediator` 因 Claude docs source change 進 review,已完成 no-SDK/no-API offline replay 但未勝過 OpenClaw,裁決更新為 `do_not_integrate_refresh_replay_gate`。 ### 2026-06-01 NeMo/Nemotron 50 筆外部 replay 實測裁決 經統帥批准後,`nvidia/nemotron-3-super-120b-a12b` 已用 50 筆 sanitized production incident request pack 完成外部離線 replay。 | 指標 | NeMo/Nemotron | OpenClaw same-run baseline | |------|---------------|----------------------------| | total_score | `0.3076` | `0.7001` | | external_error_records | `11/50` | N/A | | p95 latency | `275419.1931ms` | `1.0ms`(既有 audit replay latency) | | hard gates | failed: HITL + audit trace | failed: false repair | | promotion gate | `approved=false`, `decision=blocked` | baseline only | 裁決:本輪數據不支持 Nemotron 120B 取代或進 shadow OpenClaw。Nemotron 仍可作為離線 specialist/evaluator 候選,但必須先改善 prompt/output contract、latency/retry 與 HITL/audit gate,再重新跑同題 replay。 同輪 aggregate RCA 已保存為 `docs/evaluations/agent_nemotron_replay_failure_analysis_2026-06-01.json`。主要阻擋原因是 `model_output_missing_fields=11/50`、`unsafe_hitl_records=7`、`p95_latency_ms=275419.1931`、`score_delta=-0.3925`。下一個 Nemotron 實驗不得覆蓋本輪 evidence,必須使用 `nemo_nemotron_fabric_contract_tuned_v1` 作為新 variant,且仍限 offline replay。 `nemo_nemotron_fabric_contract_tuned_v1` 已完成本地 request-pack 與 readiness 準備:tuned request pack build、preflight、runner manifest、readiness reports 分別為 `docs/evaluations/agent_nemotron_contract_tuned_request_pack_build_2026-06-01.json`、`docs/evaluations/agent_nemotron_contract_tuned_preflight_2026-06-01.json`、`docs/evaluations/nemotron_contract_tuned_runner_manifest_2026-06-01.json`、`docs/evaluations/agent_nemotron_contract_tuned_runner_readiness_2026-06-01.json`。Readiness 為 `ready=true` / `decision=ready_for_approval`,只代表可請統帥批准外部離線跑;仍不得進 shadow/canary。 經統帥批准後,contract-tuned v1 已跑 5 筆外部 smoke。`docs/evaluations/agent_nemotron_contract_tuned_smoke_external_runner_report_2026-06-01.json` 顯示 output contract 改善:`valid=true`、`external_error_records=0`、`fallback_used_records=0`、`retry_used_records=1`;但 `p95_latency_ms=374591.0851`。`docs/evaluations/agent_nemotron_contract_tuned_smoke_gate_2026-06-01.json` 因 `latency_budget_exceeded` 擋下 full 50 replay。因此 tuned v1 仍不得進 shadow/canary,下一步應先換更快 runtime/model 或降延遲後重跑 smoke。 ### 2026-06-02 Nemotron fast-model smoke 裁決 依 2026-06-01 RCA,已用 NVIDIA live model list 選出多個較快或較新的 Nemotron-family 候選,並以同一份新抽出的 50 筆 sanitized/tuned production request pack 各跑 5 筆外部 smoke。 | 模型 | runner | p95 latency | 阻擋原因 | gate | |------|--------|-------------|----------|------| | `nvidia/nvidia-nemotron-nano-9b-v2` | `valid=true` | `60108.6491ms` | fallback 5/5、trace incomplete 5/5、latency | blocked | | `nvidia/nemotron-mini-4b-instruct` | `valid=false` | `681.8552ms` | external error 5/5、fallback 5/5、trace incomplete 5/5 | blocked | | `nvidia/nemotron-3-nano-30b-a3b` | `valid=false` | `11180.4184ms` | external error 4/5、fallback 4/5、trace incomplete 4/5 | blocked | | `nvidia/llama-3.3-nemotron-super-49b-v1.5` | `valid=true` | `67191.2835ms` | latency | blocked | 正式總表:`docs/evaluations/agent_nemotron_contract_tuned_smoke_matrix_2026-06-02.json`。相關單筆報告包含 9B v2、mini-4b、Nemotron 3 Nano 30B A3B、49B v1.5 的 runner report 與 smoke gate。 裁決:所有已測 Nemotron-family smoke 都被擋在 full replay 前。49B v1.5 是目前最接近者,因為 contract、fallback、trace 皆通過,但 p95 latency 仍超過 45 秒預算。不得進 full 50 replay、shadow、canary,也不得作為 OpenClaw 替換證據。Nemotron 目前較合理角色仍是離線 specialist/evaluator、Agent Fabric 評測層、NIM runtime 候選;生產仲裁核心仍由 OpenClaw incumbent 承擔,直到有候選在同題 replay/shadow/canary 數據勝出。 ### 2026-06-02 LangGraph Incident Kernel 離線 replay 裁決 Nemotron fast-model smoke 全部擋下後,`langgraph_incident_kernel` 已作為下一個市場候選進入同題 production replay。由於 repo 環境未安裝 Python `langgraph` package,且新 SDK/依賴需另行批准,本輪沒有安裝新依賴,也不得宣稱是官方 LangGraph SDK 能力證據;它是 AWOOOI deterministic offline workflow-kernel adapter 的 safety baseline。 | 指標 | LangGraph offline kernel | OpenClaw same-run baseline | |------|--------------------------|----------------------------| | total_score | `0.4` | `0.6983` | | incidents | `50` | `50` | | hard gates | pass | failed: false repair | | audit_trace_rate | `1.0` | `1.0` | | false_repair_rate | `0.0` | `0.08` | | rca_correct_rate | `0.0` | `0.1667` | | repair_success_rate | `0.0` | `0.5385` | | tool_dry_run_pass_rate | `0.0` | `0.8462` | | promotion gate | blocked: `candidate_does_not_beat_baseline` | baseline only | Durable reports:`docs/evaluations/agent_langgraph_replay_adapter_report_2026-06-02.json`、`docs/evaluations/agent_langgraph_replay_contract_2026-06-02.json`、`docs/evaluations/agent_langgraph_replay_grading_2026-06-02.json`、`docs/evaluations/agent_langgraph_replay_pipeline_2026-06-02.json`、`docs/evaluations/agent_langgraph_replay_scorecard_2026-06-02.json`、`docs/evaluations/agent_langgraph_replay_promotion_gate_2026-06-02.json`、`docs/evaluations/agent_langgraph_replay_summary_2026-06-02.json`。 裁決:LangGraph 類 workflow kernel 在 safety、state、HITL shell 上值得保留為 orchestration 候選;但本輪 deterministic adapter 沒有診斷/修復品質,未勝過 OpenClaw,不能進 shadow/canary,也不能取代 OpenClaw。下一步若要正式評測 LangGraph,必須先批准官方 SDK/依賴或配 stronger diagnostician,然後用同一套 replay gate 重跑。 ### 2026-06-02 OpenAI Agents SDK Coordinator 離線 replay 裁決 LangGraph offline replay 被擋下後,`openai_agents_sdk_coordinator` 已作為下一個市場候選進入同題 production replay。本機 repo 環境未安裝 `openai`、`agents`、`openai_agents` 或 `openai_agents_sdk` package;本輪未新增 SDK/依賴,也未呼叫 OpenAI API。官方 OpenAI docs 已重新確認 Agents SDK / AgentKit 的能力方向符合 AWOOOI 想測的 coordinator 邊界:orchestration、tools、guardrails、handoff、trace/eval 與 human approval;但本輪仍只是 AWOOOI deterministic offline coordinator adapter,不是官方 OpenAI Agents SDK 能力證據。 | 指標 | OpenAI offline coordinator | OpenClaw same-run baseline | |------|----------------------------|----------------------------| | total_score | `0.4` | `0.6983` | | incidents | `50` | `50` | | hard gates | pass | failed: false repair | | audit_trace_rate | `1.0` | `1.0` | | false_repair_rate | `0.0` | `0.08` | | rca_correct_rate | `0.0` | `0.1667` | | repair_success_rate | `0.0` | `0.5385` | | tool_dry_run_pass_rate | `0.0` | `0.8462` | | promotion gate | blocked: `candidate_does_not_beat_baseline` | baseline only | Durable reports:`docs/evaluations/agent_openai_coordinator_replay_adapter_report_2026-06-02.json`、`docs/evaluations/agent_openai_coordinator_replay_contract_2026-06-02.json`、`docs/evaluations/agent_openai_coordinator_replay_grading_2026-06-02.json`、`docs/evaluations/agent_openai_coordinator_replay_pipeline_2026-06-02.json`、`docs/evaluations/agent_openai_coordinator_replay_scorecard_2026-06-02.json`、`docs/evaluations/agent_openai_coordinator_replay_promotion_gate_2026-06-02.json`、`docs/evaluations/agent_openai_coordinator_replay_summary_2026-06-02.json`。 裁決:OpenAI Agents SDK 仍是市場上最值得測的 coordinator/orchestrator 候選之一;但本輪 no-SDK/no-API deterministic adapter 只證明 AWOOOI contract、handoff、guardrail、trace 邊界可接,不證明模型或官方 SDK 已勝過 OpenClaw。不得進 shadow/canary,也不得取代 OpenClaw。若要正式挑戰,需先批准 SDK 安裝、OpenAI API 成本估算、資料邊界與安全策略,再用相同 replay gate 重跑。 ### 2026-06-02 Claude Agent SDK Remediator no-SDK replay 裁決 Agent market integration review 偵測到 Claude docs source change 後,`claude_agent_sdk_remediator` 已先完成 no-SDK/no-API deterministic offline remediator replay。本機 `claude-agent-sdk` package 可見版本 `0.1.53`,但本輪未使用該 SDK、未呼叫 Anthropic/Claude API、未執行工具、未編輯檔案、未寫 production;這只驗證 AWOOOI remediation boundary,不是官方 Claude SDK/API 能力證據。 | 指標 | Claude no-SDK remediator | OpenClaw same-run baseline | |------|--------------------------|----------------------------| | total_score | `0.4` | `0.6906` | | hard_gates_pass | `true` | `false`(false repair) | | audit_trace_rate | `1.0` | `1.0` | | hitl_preserved_rate | `1.0` | `1.0` | | false_repair_rate | `0.0` | `0.08` | | promotion gate | `blocked` | baseline only | Durable reports:`docs/evaluations/agent_claude_remediator_replay_adapter_report_2026-06-02.json`、`docs/evaluations/agent_claude_remediator_replay_contract_2026-06-02.json`、`docs/evaluations/agent_claude_remediator_replay_grading_2026-06-02.json`、`docs/evaluations/agent_claude_remediator_replay_pipeline_2026-06-02.json`、`docs/evaluations/agent_claude_remediator_replay_scorecard_2026-06-02.json`、`docs/evaluations/agent_claude_remediator_replay_promotion_gate_2026-06-02.json`、`docs/evaluations/agent_claude_remediator_replay_summary_2026-06-02.json`。 裁決:Claude Agent SDK Remediator 適合作為 DevOps/code remediation specialist 候選,但本輪 deterministic adapter 未勝過 OpenClaw,不得進 shadow/canary,也不得取代 OpenClaw。若要正式挑戰,需先批准 Claude SDK/API 使用方式、成本上限、資料邊界、secret isolation、trace retention,然後用同一套 replay gate 重跑。 ## 問題陳述 如何讓兩個 AI 在 Telegram 中協作,而不會: - 訊息混亂(誰說了什麼?) - 責任不清(誰做的決策?) - 無限迴圈(互相觸發) - 增加過多延遲 ## 決策 ### 採用「仲裁-執行分工」架構 ``` OpenClaw = 仲裁者 (Arbitrator) - 決定「為什麼」和「風險等級」 Nemotron = 執行者 (Executor) - 決定「怎麼做」和「具體指令」 ``` ### 職責分離 | 角色 | OpenClaw | Nemotron | |------|----------|----------| | **任務** | Root Cause Analysis | Tool Calling | | **輸出** | 風險等級 + 責任團隊 + 原因推理 | kubectl 指令 + 參數驗證 | | **模型** | Ollama/Gemini (RCA 任務) | Nemotron-mini (Tool 任務) | | **信心度** | 0-100% (AI 分析品質) | 驗證狀態 (✅/❌) | | **備援** | Expert System 規則 | Gemini Tool Calling | ### 流程設計 ``` 1. Incident 產生 ↓ 2. OpenClaw.generate_incident_proposal() → 輸出: risk_level, reasoning, primary_responsibility ↓ 3. 判斷是否需要 Nemotron ├─ LOW 風險 → 跳過 Nemotron └─ MEDIUM/HIGH/CRITICAL → 呼叫 Nemotron ↓ 4. NvidiaProvider.tool_call() → 輸出: tool_name, arguments, validation_status ↓ 5. 組合結果 → 推送 Telegram 卡片 ↓ 6. 用戶簽核 → 執行 ``` ### 觸發條件 | 風險等級 | OpenClaw | Nemotron | 原因 | |----------|----------|----------|------| | LOW | ✅ | ❌ | 低風險操作不需要 Tool 驗證 | | MEDIUM | ✅ | ✅ | 需要 Tool 驗證操作可行性 | | HIGH | ✅ | ✅ | 高風險必須雙重驗證 | | CRITICAL | ✅ | ✅ + HITL | 危險操作必須人工介入 | ## 實作規格 ### 1. 擴展 TelegramMessage ```python @dataclass class TelegramMessage: # 現有欄位... # 新增 Nemotron 結果欄位 nemotron_enabled: bool = False nemotron_tools: list[dict] | None = None # Tool Calling 結果 nemotron_validation: str = "" # "✅ 驗證通過" / "❌ 驗證失敗" nemotron_latency_ms: float = 0.0 ``` ### 2. 擴展 generate_incident_proposal ```python async def generate_incident_proposal_with_tools( self, incident_id: str, severity: str, signals: list[dict], affected_services: list[str], ) -> tuple[dict | None, str, bool]: """ Phase 22: OpenClaw + Nemotron 協作 Returns: (proposal_dict, provider, success) proposal_dict 新增: - nemotron_tools: Tool Calling 結果 - nemotron_validation: 驗證狀態 """ # Step 1: OpenClaw 仲裁 proposal, provider, success = await self.generate_incident_proposal( incident_id, severity, signals, affected_services ) if not success: return proposal, provider, success # Step 2: 判斷是否需要 Nemotron risk_level = proposal.get("risk_level", "low").lower() if risk_level == "low": proposal["nemotron_enabled"] = False return proposal, provider, True # Step 3: Nemotron Tool Calling from src.services.nvidia_provider import get_nvidia_provider nvidia = get_nvidia_provider() tool_result = await nvidia.tool_call( messages=[{ "role": "user", "content": f""" 根據以下分析,生成對應的 kubectl 操作: - Incident: {incident_id} - 原因: {proposal.get('reasoning', '')} - 目標資源: {proposal.get('target_resource', '')} - 建議操作: {proposal.get('action', '')} """ }], tools=K8S_OPERATION_TOOLS, ) # Step 4: 驗證 Tool Calling 結果 validation = await self._validate_tool_calls(tool_result.tool_calls) proposal["nemotron_enabled"] = True proposal["nemotron_tools"] = [ {"tool": tc.tool_name, "args": tc.arguments, "valid": tc.valid} for tc in tool_result.tool_calls ] proposal["nemotron_validation"] = validation proposal["nemotron_latency_ms"] = tool_result.latency_ms return proposal, provider, True ``` ### 3. Telegram 卡片格式 ```python def format_with_nemotron(self) -> str: """格式化含 Nemotron 結果的訊息""" # OpenClaw 區塊 openclaw_block = f""" 🤖 OpenClaw 仲裁 ├ 📊 信心: {self.confidence_emoji} {self.confidence_pct}% ├ 👥 責任: {self.primary_responsibility} └ 💡 原因: {self.root_cause[:50]} """ # Nemotron 區塊 (如果啟用) nemotron_block = "" if self.nemotron_enabled and self.nemotron_tools: tools_str = "\n".join([ f" {'✅' if t['valid'] else '❌'} {t['tool']}: {t['args'][:30]}" for t in self.nemotron_tools[:3] # 最多顯示 3 個 ]) nemotron_block = f""" ━━━━━━━━━━━━━━━━━━━ 🔧 Nemotron 執行方案 {tools_str} └ 驗證: {self.nemotron_validation} """ return f"{openclaw_block}{nemotron_block}" ``` ### 4. 異步執行 (非阻塞) ```python async def _push_decision_to_telegram_async( incident: Incident, proposal_data: dict, ) -> None: """ 異步推送,不阻塞主流程 Phase 22: 如果 Nemotron 延遲過長 (>10s),先推送 OpenClaw 結果, Nemotron 結果後續用 edit_message 更新 """ # 先推送 OpenClaw 結果 message_id = await gateway.send_approval_card( # ... OpenClaw 結果 ) # 如果需要 Nemotron,異步執行並更新 if proposal_data.get("risk_level") in ["medium", "high", "critical"]: asyncio.create_task( _update_with_nemotron_result(message_id, incident, proposal_data) ) ``` ## 後果 ### 正面 - **清晰分工**: OpenClaw 和 Nemotron 職責明確 - **可追蹤**: 每個 AI 的貢獻獨立顯示 - **容錯性**: 備援鏈清晰 (Nemotron → Gemini → Expert) - **效能**: 低風險操作不觸發 Nemotron,節省延遲 ### 負面 - **延遲增加**: 高風險操作需要兩輪 LLM - **複雜度**: 訊息格式需要擴展 ### 風險緩解 | 風險 | 緩解 | |------|------| | Nemotron 延遲 11-45s | 異步執行,先推送 OpenClaw 結果 | | Tool Calling 失敗 | Fallback 到 Gemini,再失敗則只顯示 OpenClaw | | 訊息超長 | 縮寫 Tool 參數,完整內容放 SignOz Link | ## 併發控制 (與 ADR-038 整合) > **首席架構師 P1 必修項** (2026-03-31) ### 雙 Semaphore 策略 ```python # apps/api/src/core/circuit_breaker.py 擴展 class OpenClawGuard: def __init__(self): self.openclaw_semaphore = asyncio.Semaphore(3) # 原有 self.nemotron_semaphore = asyncio.Semaphore(2) # 新增 (NVIDIA API 較慢) ``` **設計原因**: - Nemotron 併發限制為 2 (低於 OpenClaw 的 3) - NVIDIA NIM 免費 tier 有 RPM 限制 - Nemotron 延遲較高 (11-45s),過多並發無益 ### 並行執行優化 ```python # Step 3 優化: OpenClaw + Nemotron 並行而非串行 import asyncio async def generate_incident_proposal_with_tools(...): # 並行啟動 OpenClaw 和 Nemotron (減少延遲) openclaw_task = asyncio.create_task( self.generate_incident_proposal(incident_id, severity, signals, affected_services) ) # 先等待 OpenClaw 完成,判斷是否需要 Nemotron proposal, provider, success = await openclaw_task if not success or proposal.get("risk_level", "low").lower() == "low": return proposal, provider, success # 需要 Nemotron - 此時 OpenClaw 已完成,立即啟動 Nemotron nemotron_result = await self._call_nemotron_tools(proposal) # 組合結果 return self._combine_results(proposal, nemotron_result), provider, True ``` **延遲對比**: | 場景 | 串行 | 並行 | 改善 | |------|------|------|------| | MEDIUM 風險 | 3s + 15s = 18s | max(3s, 15s) = 15s | -3s | | HIGH 風險 | 5s + 30s = 35s | max(5s, 30s) = 30s | -5s | --- ## Circuit Breaker 整合 ### 雙層 Circuit Breaker 協調 ``` ┌─────────────────────────────────────────┐ │ OpenClawGuard (ADR-038) │ │ - 管理請求佇列 │ │ - 長期熔斷 (5 分鐘) │ └─────────────────────────────────────────┘ │ ▼ ┌─────────────────────────────────────────┐ │ NvidiaProvider.CircuitBreaker │ │ - NVIDIA API 短期熔斷 (60s) │ │ - 失敗 3 次後 OPEN │ └─────────────────────────────────────────┘ ``` ### 熔斷策略 | 層級 | 觸發條件 | 恢復時間 | 影響 | |------|---------|---------|------| | OpenClawGuard | 佇列滿 (>10) | 5 分鐘 | 停止新請求 | | NvidiaProvider | 連續 3 失敗 | 60 秒 | Fallback 到 Gemini | --- ## Feature Flag 支援 > **首席架構師 P1 必修項** ### 環境變數 ```bash # 啟用/停用 Nemotron 協作 (預設 true) ENABLE_NEMOTRON_COLLABORATION=true # Nemotron 呼叫超時 (預設 45s) NEMOTRON_TIMEOUT_SECONDS=45 # 強制使用異步更新 (先推 OpenClaw,後更新 Nemotron) NEMOTRON_ASYNC_UPDATE=true ``` ### 回滾計畫 ```python async def generate_incident_proposal_with_tools(...): # Feature Flag 檢查 if not settings.ENABLE_NEMOTRON_COLLABORATION: return await self.generate_incident_proposal(...) # 原流程 # ... 協作邏輯 ``` **回滾步驟**: 1. 設置 `ENABLE_NEMOTRON_COLLABORATION=false` 2. Rollout restart awoooi-api 3. 無需代碼回滾 --- ## DI 模式重構 > **首席架構師 P1 必修項** - 避免函數內 import ### 修改前 (❌ 違反 DI) ```python # Step 3: Nemotron Tool Calling from src.services.nvidia_provider import get_nvidia_provider # ❌ 函數內 import nvidia = get_nvidia_provider() ``` ### 修改後 (✅ DI 模式) ```python # apps/api/src/services/openclaw.py from src.services.nvidia_provider import INvidiaProvider class OpenClawService: def __init__( self, nvidia_provider: INvidiaProvider | None = None, # DI 注入 ): self._nvidia = nvidia_provider or get_nvidia_provider() async def generate_incident_proposal_with_tools( self, incident_id: str, severity: str, signals: list[dict], affected_services: list[str], ) -> tuple[dict | None, str, bool]: # ... 使用 self._nvidia 而非 import ``` --- ## 測試策略 ### E2E 測試案例 ```python # tests/test_openclaw_nemotron_collaboration.py @pytest.mark.asyncio async def test_low_risk_skips_nemotron(): """LOW 風險不觸發 Nemotron""" result = await openclaw.generate_incident_proposal_with_tools(...) assert result[0]["nemotron_enabled"] is False @pytest.mark.asyncio async def test_medium_risk_enables_nemotron(): """MEDIUM 風險啟用 Nemotron""" result = await openclaw.generate_incident_proposal_with_tools(...) assert result[0]["nemotron_enabled"] is True assert result[0]["nemotron_tools"] is not None @pytest.mark.asyncio async def test_nemotron_failure_fallback(): """Nemotron 失敗時 fallback 到 Gemini""" # Mock NVIDIA 失敗 with patch("nvidia_provider.tool_call", side_effect=Exception): result = await openclaw.generate_incident_proposal_with_tools(...) # 應該有結果 (來自 Gemini fallback) assert result[2] is True @pytest.mark.asyncio async def test_feature_flag_disabled(): """Feature Flag 停用時走原流程""" with patch.dict(os.environ, {"ENABLE_NEMOTRON_COLLABORATION": "false"}): result = await openclaw.generate_incident_proposal_with_tools(...) assert "nemotron_enabled" not in result[0] ``` ### 整合測試 ```python @pytest.mark.integration async def test_telegram_message_with_nemotron(): """Telegram 訊息包含 Nemotron 區塊""" msg = TelegramMessage( nemotron_enabled=True, nemotron_tools=[{"tool": "restart_deployment", "args": {...}, "valid": True}], ) formatted = msg.format_with_nemotron() assert "Nemotron 執行方案" in formatted assert "✅ restart_deployment" in formatted ``` --- ## 實作排程 (詳細) | 階段 | 內容 | 時間 | 檔案 | 依賴 | |------|------|------|------|------| | **22.1** | TelegramMessage 擴展 | 2h | `telegram_gateway.py` | 無 | | **22.2a** | OpenClawGuard 雙 Semaphore | 1h | `circuit_breaker.py` | 無 | | **22.2b** | DI 模式重構 | 1h | `openclaw.py` | 22.2a | | **22.2c** | `generate_incident_proposal_with_tools` | 2h | `openclaw.py` | 22.2a, 22.2b | | **22.3a** | Feature Flag 支援 | 1h | `config.py` | 無 | | **22.3b** | 異步推送邏輯 | 2h | `decision_manager.py` | 22.1, 22.2c | | **22.4a** | 單元測試 | 2h | `test_openclaw_nemotron*.py` | 22.2c | | **22.4b** | E2E 測試 | 2h | `test_e2e_collaboration.py` | 22.3b | | **總計** | | **13h (~1.5 天)** | | | --- ## 首席架構師審查結論 > **審查日期**: 2026-03-31 (台北時區) > **分數**: 83/100 → **條件通過** ### P1 必修項 (已補充) | 編號 | 項目 | 狀態 | |------|------|------| | P1-1 | 併發控制整合 | ✅ 已補充 | | P1-2 | DI 模式 | ✅ 已補充 | | P1-3 | Feature Flag | ✅ 已補充 | ### P2 建議項 (後續迭代) | 編號 | 項目 | 說明 | |------|------|------| | P2-1 | 並行優化 | 已納入設計 | | P2-2 | Pydantic Model | Phase 22.5 | | P2-3 | NemotronBlock | Phase 22.5 | --- ## 相關文件 - ADR-036: Nemotron Tool Calling 整合 - ADR-038: OpenClaw 併發治理 - Phase 18: 失敗自動修復閉環 - `feedback_ai_rate_limiter.md`: AI 用量控制 --- **Co-Authored-By: Claude Opus 4.5 **