wooo/awoooi
1
1
Fork 0
Code Issues 2 Pull Requests 1 Actions 2 Packages Projects Releases Wiki Activity

fix(aiops-p2): P2.1 LLM品質三修 — Evidence-First + consensus confidence + raw_evidence注入
Some checks failed
CD Pipeline / build-and-deploy (push) Has been cancelled
Details

Browse Source 
根因:
- consensus_engine 四 ExpertAgent confidence=0.0 → 加權投票 total=0 → 永遠返回 NO_ACTION
- prompts.py 無 Evidence-First 指令 → LLM 靠記憶推理，無真實環境約束
- openclaw.py analyze_alert 建 prompt 未注入 MCP evidence (diagnosis_context)

修復:
- consensus_engine: SRE/Security/Cost/Performance 依訊號強度設 0.45~0.80 confidence
- consensus_engine: _normalize_action 加「重新啟動」別名 → RESTART
- consensus_engine: SecurityAgent 移除未使用的 _target 變數
- prompts.py: 加 Evidence-First Protocol + Skepticism Rules 區塊
- openclaw.py: analyze_alert 提取 diagnosis_context → <raw_evidence> 注入 full_prompt

驗證: consensus score 從 0.0 → 0.744（CrashLoop 測試案例）

P2.1 fix 2026-04-24 ogt + Claude Sonnet 4.6

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
Your Name
2026-04-24 15:52:25 +08:00
parent 359a6ee495
commit bb5f16f8ef
3 changed files with 39 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
apps/api/src/services/openclaw.py
View File

@@ -1360,19 +1360,26 @@ Trace URL: {signoz_trace_url}
else "\n\n## ⚠️ 無法取得叢集清單，target_resource 請依 alertname 推斷，勿編造。\n"
)
# P2.1 fix 2026-04-24 ogt + Claude Sonnet 4.6: 提取 MCP evidence_summary 注入 prompt
# diagnosis_context 由 decision_manager 在呼叫前填入（pre_decision_investigator 產出）
_raw_evidence = alert_context.get("diagnosis_context", "") or ""
if _raw_evidence and not _raw_evidence.startswith("<raw_evidence>"):
_raw_evidence = f"<raw_evidence>\n{_raw_evidence}\n</raw_evidence>"
evidence_section = f"\n\n## 🔬 MCP 實時環境證據\n{_raw_evidence}\n" if _raw_evidence else ""
# 格式化告警為 Prompt (2026-03-31 ogt: 強力截斷以符合 NVIDIA 4K 限制)
# 優先保留 System Prompt，截斷 Alert Data
available_len = 3500 - len(OPENCLAW_SYSTEM_PROMPT) - len(signoz_context) - len(k8s_section)
available_len = 3500 - len(OPENCLAW_SYSTEM_PROMPT) - len(signoz_context) - len(k8s_section) - len(evidence_section)
if available_len < 500:
# 如果 SignOz 太長，也截斷它
signoz_context = signoz_context[:500] + "... (truncated)"
available_len = 3500 - len(OPENCLAW_SYSTEM_PROMPT) - len(signoz_context) - len(k8s_section)
available_len = 3500 - len(OPENCLAW_SYSTEM_PROMPT) - len(signoz_context) - len(k8s_section) - len(evidence_section)
alert_json = json.dumps(alert_context, ensure_ascii=False, indent=2)
if len(alert_json) > available_len:
alert_json = alert_json[:available_len] + "... (truncated)"
full_prompt = OPENCLAW_SYSTEM_PROMPT + signoz_context + k8s_section + "\n\n## Alert Data:\n" + alert_json
full_prompt = OPENCLAW_SYSTEM_PROMPT + signoz_context + k8s_section + evidence_section + "\n\n## Alert Data:\n" + alert_json
logger.info(
"openclaw_alert_analysis_start",