feat(ea-hitl): ADR-021 EA 升級審核 pre-fetch + 競價告警金額影響量化

根治 2026-05-02 統帥反映的三層 EA escalation 訊息空泛問題:

1. _escalate_to_human 對 price_drop_alert / market_opportunity /
   threat_escalation 三類觸發,送 Telegram 前先 await Hermes 取具體
   SKU 清單覆蓋 plan 元流程文字(5s 短超時,失敗 fallback 原 plan)
2. NemoTron 競價告警新增 _compute_business_impact helper:
   過去 7 日營收流失(gap_pct>0 才算)+ 跟進競品建議價,
   dispatch 主路徑 / 防線二 / Hermes rule fallback 三條全部 Python
   獨裁注入,告警含「📉 NT$ X」「🎯 NT$ Y」具體金額
3. 補實 telegram_bot_service.handle_callback 的 momo:eig: prefix
   handler,HITL「🛑 忽略此事件」按鈕首次有對應 audit 寫入

Critic 審查通過(5 項必修全綠):
- Critical-1: user_label HTML escape 防 Telegram username XSS
- High-1: pre-fetch 改 asyncio.wait_for(5s) 防阻塞 escalation
- High-2: 全部行缺金額時 return None 觸發 plan fallback
- Medium-2: 空 event_id callback 拒絕避免 audit 污染
- Medium-3: gap_pct≤0 時 revenue_loss_7d 強制歸 0 不誤導降價

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
OoO
2026-05-03 00:03:38 +08:00
parent 650ef4c5db
commit 00591c5489
6 changed files with 462 additions and 17 deletions

View File

@@ -0,0 +1,124 @@
# ADR-021: EA HITL Pre-fetch + 競價告警必填金額影響量化
- **Status**: Accepted
- **Date**: 2026-05-03
- **Deciders**: 統帥
- **Related**: 補強 ADR-012Agent Action Ladder L3 HITL對 escalation 訊息內容的要求;不取代任何既有 ADR
- **Affects**: `services/elephant_alpha_autonomous_engine.py``services/nemoton_dispatcher_service.py``services/hermes_analyst_service.py``services/telegram_bot_service.py`
## Context
2026-05-02 統帥收到一條 EA 升級審核 Telegram 告警commit `b5a2b094` 時段),內容為:
> 自主決策信心度 0.83 低於門檻,需人工批准
> AI 摘要:競爭對手價格溢價 >5% 且有庫存充足…
> 步驟 1[OpenClaw] 基於市區拉貨區市場特性,生成包含動態定價模型與競爭對手分析的策略建議
> 步驟 2[Hermes] 識別具體競爭商品並量化價格差異(目標確認至少 20 個高價溢價 SKU
> 步驟 3[NemoTron] 將生成的定價策略與競爭分析結果同步至人工審核管道
統帥指出此告警**毫無決策價值**,原因:
1. 沒有具體 SKU、沒有金額影響、沒有可批准/駁回的具體動作
2. 三步驟「建議行動」實為「申請開始分析」的元流程描述OpenClaw 生成策略 → Hermes 識別 SKU → NemoTron 同步),人類審核者沒有判斷依據
3. HITL 攔截點放錯位置:應在「具體動作出爐後、執行前」攔,而非「分析計畫產出但尚未跑」
### 程式碼層證據
**根因 A — `_escalate_to_human` 的 ai_actions 來源是 plan 階段**[`services/elephant_alpha_autonomous_engine.py:528-557`](services/elephant_alpha_autonomous_engine.py#L528-L557) 的 `_execute_autonomous_decision``elephant_orchestrator.analyze_and_coordinate` 產出 `StrategicDecision` 後,**信心度不足即 escalate**——此時 `decision.execution_plan` 還只是 Gemini 寫的 plan 文字Hermes/NemoTron 從未實際跑過取得具體 SKU。`_escalate_to_human` 把 plan 的前 3 個 step description 直接灌進 `triaged_alert.ai_actions` 是空泛元流程的根源。
**根因 B — Hermes/NemoTron 已有具體告警,但缺金額影響量化**[`services/nemoton_dispatcher_service.py:683-700`](services/nemoton_dispatcher_service.py#L683-L700) 的 `_exec_trigger_price_alert` 已經能輸出「[SKU] 商品MOMO $X / PChome $Y價差 ±%|銷量 ±%」,但**缺絕對金額影響**——人類看到「價差 22.4%、銷量 -35%」仍需自己換算「我這週實際少賺多少」才能決策。
**根因 C — `momo:eig:` callback 按鈕從未實作**[`services/telegram_templates.py:466-467`](services/telegram_templates.py#L466-L467) 的 `triaged_alert` 鍵盤產出「🛑 忽略此事件」按鈕callback_data = `momo:eig:{event_id}`),但 [`services/telegram_bot_service.py:512`](services/telegram_bot_service.py#L512) 的 `handle_callback` 只 dispatch `menu:/cmd:/await:` 三個 prefix**`momo:` prefix 完全沒處理**——統帥點擊忽略按鈕後永遠沒反應HITL 流程閉環缺一環。
## Decision
對 EA L3 HITL 升級審核訊息與 NemoTron 競價告警採取三項根治措施:
### 規則 1 — EA 升級審核 pre-fetch Hermes 具體威脅清單
`_PRICE_RELATED_TRIGGERS = {price_drop_alert, market_opportunity, threat_escalation}` 三類觸發,`_escalate_to_human` **送 Telegram 前先呼叫 Hermes 取得具體 SKU 清單**,將前 5 筆格式化為:
```
[SKU] 商品名稱MOMO $X vs PChome $Y (±%)|近 7 日流失 NT$ Z建議跟進 NT$ W
```
蓋掉原本的 plan 元流程文字。**強制配套限制**
- `asyncio.wait_for(timeout=5)` 短超時Hermes 熱駐留 < 10s但冷啟動會拖到 30s+HITL 訊息延遲不可大於 10s
- Pre-fetch 失敗timeout / 0 threats / 全部缺金額)→ fallback 回原 plan 文字,**不中斷 escalation 主流程**
- 「全部行皆缺金額」也視同無料 fallback避免「乾巴巴兩行 MOMO/PChome 比價」比 plan 文字更空泛
### 規則 2 — NemoTron 告警必填金額影響量化
新增模組級 helper `_compute_business_impact(threat) -> {revenue_loss_7d, recommended_price}`
- **`revenue_loss_7d`** = `max(0, sales_7d_prev_amount - sales_7d_curr_amount)` 且**僅在 `gap_pct > 0` 時計算**
- 語意:「我方比競品貴,且過去 7 日銷量金額下滑 → 推估價格因素導致的流失」
- `gap_pct ≤ 0`(我方便宜或持平)即使銷量下滑亦歸 0避免把季節性/商品壽命終結等非價格因素誤標為「流失」誘導降價
- **`recommended_price`** = `round(pchome_price)``gap_pct > 0`;否則 `None`
- 語意:「跟進競品的最低調價金額」;統帥可基於此再依毛利策略加溢價
`_fmt_price_alert` / `_fmt_human_review` 加入「📉 過去 7 日營收流失NT$ X」「🎯 跟進競品建議價NT$ Y毛利策略可再加溢價」區塊。dispatch() 主路徑、防線二強制覆核、`_hermes_rule_fallback` 三條路徑**全部走 Python 獨裁注入**(同 Bug-1 防線二原則:客觀數字不過 LLM
為支援此計算,`PriceThreat` dataclass 新增 `sales_7d_curr_amount` / `sales_7d_prev_amount` 兩個欄位,由 `_batch_analyze` 從 SQL 結果直接帶出(**不**餵給 LLM 推理,避免 token 爆炸與 LLM 嘗試引用幻覺)。
### 規則 3 — 補實 `momo:eig:` callback handler
`telegram_bot_service.py.handle_callback` 在既有 `menu:/cmd:/await:` dispatch 之前加 `momo:eig:` 處理,呼叫 `_handle_event_ignore_callback`
1. 解析 `event_id`**空 id 立即拒絕**(防 audit 污染)
2. 寫入 `ai_insights``status='ignored', insight_type='human_review', metadata_json``event_id / decided_by / decided_at`ADR-012 §③ audit trail
3. 編輯原訊息加「🛑 已忽略 by `<user>` @ `<ts>`」尾註
4. **`user_label` / `ts_label` 寫入 HTML 前必須 `html.escape()`**user-controlled Telegram username 透過 `<a href>/<pre>` 注入超連結與破版的 XSS 防線)
5. 任何例外都 best-effort不阻斷 UI
### Critic 審查(必修)
| 編號 | 嚴重度 | 內容 | 修復 |
|------|--------|------|------|
| Critical-1 | CRITICAL | `user_label` 直接 HTML 拼接 → username 注入 `<a>/<pre>` 破版 | `html.escape()` 雙重 escape |
| High-1 | HIGH | Pre-fetch Hermes 同步阻塞 escalation cooldown 視窗30-60s | `asyncio.wait_for(timeout=5)` |
| High-2 | HIGH | Hermes 有 threats 但全部缺金額時 → 兩行乾巴巴比價反而更空泛 | `any_concrete` 判斷,全缺則 `return None` 觸發 plan fallback |
| Medium-2 | MEDIUM | 空 `event_id` callback 寫入 `'unknown'` 污染 audit | prefix 解析後即拒絕 |
| Medium-3 | MEDIUM | `gap_pct ≤ 0``prev > curr` 仍顯示「流失」誤導降價 | `revenue_loss_7d` 條件改為 `if gap_pct > 0` |
## Alternatives Considered
| 方案 | 為何不選 |
|------|---------|
| A. 廢掉 `market_opportunity` 觸發類型,全交 NemoTron pipeline | EA 同時負責 `price_drop_alert / threat_escalation / resource_optimization / code_exception` 多類觸發,不可單刪一類;且 EA orchestrator 是 SOT廢觸發類型需大改 |
| B. 把 EA 信心度門檻降到 0.5讓「market_opportunity」全部自主執行 | 違反 ADR-012 L3 HITL 對「實際修改價格」的安全網;統帥未授權自動調價 |
| C. Pre-fetch 改為背景 taskescalation 訊息分兩次發 | 兩次告警分裂語境,使用者體驗差;且 Hermes 結果晚到時 escalation 已 cooldown |
| D. 不 pre-fetch要求人類自己用 `/menu` 查 SKU | 違反「告警必須 actionable」精神且增加人類認知負荷 |
## Consequences
### 正面
- EA 升級審核 Telegram 內容從元流程描述變為「具體 SKU + 價格 + 金額流失 + 建議調價」HITL 真正可決策
- NemoTron 既有告警再升級,每筆都帶可批准/駁回的金額判斷依據
- `momo:eig:` 按鈕首次有對應 handlerHITL 流程閉環完整
- pre-fetch 改用 5s 短超時 + fallback最壞情況退回原 plan 文字,不破壞既有行為
### 負面 / 風險
- 每次價格類 escalation 多花 ≤ 5sHermes 熱駐留實測 < 10s 但有 timeout整體告警延遲略增
- Hermes 在 5s 內若沒回應,告警內容降級回 plan 文字(仍維持原行為,無新增風險)
- `gap_pct ≤ 0` 案例的銷量下滑(非價格因素)將完全不顯示流失金額——若統帥需追蹤「非價格流失」需另開告警類型(待後續 ADR
### 監控指標
- Telegram 「EA 升級審核」訊息含「📉 過去 7 日營收流失」比例(部署後應 → 每筆價格類觸發都有,除非 5s timeout
- `_handle_event_ignore_callback` audit 寫入頻率(觀察人類實際使用 HITL 比例)
- Hermes pre-fetch timeout rate>10% 代表 Ollama 冷啟動嚴重,需檢查 keep_alive
## 實作 checklist
- [x] `services/hermes_analyst_service.py` PriceThreat 新增絕對金額欄位
- [x] `services/nemoton_dispatcher_service.py` `_compute_business_impact` helper + 三條 dispatch 路徑注入
- [x] `services/elephant_alpha_autonomous_engine.py` `_fetch_hermes_threats_summary` + 5s timeout + fallback
- [x] `services/telegram_bot_service.py` `_handle_event_ignore_callback` + HTML escape + 空 id 拒絕
- [x] Critic 審查通過Critical-1 / High-1 / High-2 / Medium-2 / Medium-3 全修)
- [x] Smoke test`_compute_business_impact` 對 gap≤0 / gap=0 / 銷量回升 / bogus type 四案例驗證
- [x] `docs/adr/README.md` 索引加 ADR-021
- [ ] 部署到 188 + 觀察首日 EA escalation 內容(人工驗證)

View File

@@ -42,6 +42,7 @@
| [018](ADR-018-four-agent-ai-automation-control-plane.md) | 四 AI Agent 自動化控制面Hermes/NemoTron/OpenClaw/ElephantAlpha | Accepted | 2026-04-29 | | [018](ADR-018-four-agent-ai-automation-control-plane.md) | 四 AI Agent 自動化控制面Hermes/NemoTron/OpenClaw/ElephantAlpha | Accepted | 2026-04-29 |
| [019](ADR-019-telegram-bot-agentic-conversation-layer.md) | Telegram Bot Agentic Conversation Layer菜單→Agent 決策統一入口) | Accepted | 2026-05-02 | | [019](ADR-019-telegram-bot-agentic-conversation-layer.md) | Telegram Bot Agentic Conversation Layer菜單→Agent 決策統一入口) | Accepted | 2026-05-02 |
| [020](ADR-020-code-review-full-autoheal.md) | Code Review 全自動修復政策(局部覆寫 ADR-012 HITL | Accepted | 2026-05-02 | | [020](ADR-020-code-review-full-autoheal.md) | Code Review 全自動修復政策(局部覆寫 ADR-012 HITL | Accepted | 2026-05-02 |
| [021](ADR-021-ea-hitl-prefetch-and-alert-impact.md) | EA HITL Pre-fetch + 競價告警必填金額影響量化 | Accepted | 2026-05-03 |
## 規範 ## 規範

View File

@@ -115,6 +115,14 @@ _PRICE_ADJUSTMENT_REVIEW_ACTIONS = frozenset({
"dispatch_price_updates", "dispatch_price_updates",
}) })
# A' 軌價格相關觸發類型HITL 前需 pre-fetch Hermes 具體威脅清單
# 取代 Gemini plan 階段的元流程文字(「步驟 1:[OpenClaw] 生成策略」這類)
_PRICE_RELATED_TRIGGERS = frozenset({
"price_drop_alert",
"market_opportunity",
"threat_escalation",
})
def _zh_trigger(trigger_type: str) -> str: def _zh_trigger(trigger_type: str) -> str:
return _TRIGGER_ZH.get(trigger_type, trigger_type) return _TRIGGER_ZH.get(trigger_type, trigger_type)
@@ -623,6 +631,78 @@ class ElephantAlphaAutonomousEngine:
from services.hermes_analyst_service import HermesAnalystService from services.hermes_analyst_service import HermesAnalystService
return await self._run_with_timeout(HermesAnalystService().run, timeout=SSH_COMMAND_TIMEOUT) return await self._run_with_timeout(HermesAnalystService().run, timeout=SSH_COMMAND_TIMEOUT)
async def _fetch_hermes_threats_summary(self, top_n: int = 5) -> Optional[List[str]]:
"""A'HITL escalation 前 pre-fetch Hermes 具體威脅清單,
將「步驟 1: [OpenClaw] 生成策略」這類元流程文字換成
「[SKU] 商品MOMO $X / PChome $Y流失 NT$ Z建議 NT$ W」具體可決策行動。
失敗回 None由呼叫端 fallback 至既有 execution_plan 文字。
本方法為 best-effort任何例外都不阻斷 escalation 主流程。
Critic High-1 fix: 加 5 秒短超時防止阻塞 escalation cooldown 視窗
Hermes 完整 run 可能 30-60sHITL 訊息應快速送出)
Critic High-2 fix: 若每筆都缺 loss/rec_price視同無料、return None 觸發 fallback
"""
# 使用 5s 短超時Hermes 熱駐留時實測 < 10s但若需冷啟動會拖到 30s+
# HITL 訊息延遲不可大於 10s影響統帥決策時效性寧可 fallback 到原 plan 文字
try:
result = await asyncio.wait_for(self._hermes_analyze(), timeout=5)
except asyncio.TimeoutError:
self._log.warning("Pre-fetch Hermes 5s timeout; falling back to plan text")
return None
except Exception as e:
self._log.warning("Pre-fetch Hermes threats failed (non-blocking): %s", e)
return None
threats = getattr(result, "threats", None) or []
if not threats:
self._log.info("Pre-fetch Hermes returned 0 threats; falling back to plan text")
return None
# 模組頂部 import 較乾淨,但這裡保留 lazy import 避免兩服務循環依賴
# nemoton 也需 hermes_analyst 的 PriceThreat dataclass
try:
from services.nemoton_dispatcher_service import _compute_business_impact
except Exception as imp_err:
self._log.error("import _compute_business_impact failed: %s", imp_err)
_compute_business_impact = None
lines: List[str] = []
any_concrete = False # Critic High-2: 至少一筆有金額才算具體
for t in threats[:top_n]:
sku = getattr(t, "sku", "?")
name = getattr(t, "name", "")[:24]
momo = float(getattr(t, "momo_price", 0) or 0)
pchome = float(getattr(t, "pchome_price", 0) or 0)
gap_pct = float(getattr(t, "gap_pct", 0) or 0)
impact = _compute_business_impact(t) if _compute_business_impact else {
"revenue_loss_7d": 0.0, "recommended_price": None,
}
loss = impact.get("revenue_loss_7d", 0.0) or 0.0
rec_price = impact.get("recommended_price")
parts = [
f"[{sku}] {name}",
f"MOMO ${momo:,.0f} vs PChome ${pchome:,.0f} ({gap_pct:+.1f}%)",
]
if loss > 0:
parts.append(f"近 7 日流失 NT$ {loss:,.0f}")
any_concrete = True
if rec_price is not None and rec_price > 0:
parts.append(f"建議跟進 NT$ {rec_price:,.0f}")
any_concrete = True
lines.append("".join(parts))
if not any_concrete:
# Critic High-2: 全部都只有「MOMO $X vs PChome $Y」乾巴巴兩行
# 比原本「步驟 1:OpenClaw 生成策略」更空泛。返回 None 觸發 plan fallback
self._log.info("Pre-fetch threats lacked impact figures on all rows; falling back")
return None
self._log.info("Pre-fetch Hermes threats produced %d concrete actions", len(lines))
return lines
async def _dispatch_alerts(self, threats: List[Any]) -> Any: async def _dispatch_alerts(self, threats: List[Any]) -> Any:
from services.nemoton_dispatcher_service import NemotronDispatcher from services.nemoton_dispatcher_service import NemotronDispatcher
return await self._run_with_timeout( return await self._run_with_timeout(
@@ -789,6 +869,26 @@ class ElephantAlphaAutonomousEngine:
cooldown_min = self._get_cooldown_min(trigger.trigger_type) cooldown_min = self._get_cooldown_min(trigger.trigger_type)
if not dedup_ts or (datetime.now().timestamp() - dedup_ts) / 60 >= cooldown_min: if not dedup_ts or (datetime.now().timestamp() - dedup_ts) / 60 >= cooldown_min:
self._store_escalation(trigger.trigger_type) self._store_escalation(trigger.trigger_type)
# A' 軌:價格類觸發前 pre-fetch Hermes 具體威脅清單,
# 取代「步驟 1:[OpenClaw] 生成策略」這類元流程文字。
# — Claude Opus 4.7 (2026-05-02)
concrete_actions: Optional[List[str]] = None
if trigger.trigger_type in _PRICE_RELATED_TRIGGERS:
try:
concrete_actions = await self._fetch_hermes_threats_summary(top_n=5)
except Exception as e:
self._log.warning("Pre-fetch threats raised (non-blocking): %s", e)
concrete_actions = None
if concrete_actions:
ai_actions_payload = concrete_actions
else:
ai_actions_payload = [
f"步驟 {s.get('step', i+1)}{_zh_step(s)}"
for i, s in enumerate(decision.execution_plan[:3])
] or ["無具體執行計畫"]
try: try:
from services.telegram_templates import triaged_alert, _send_telegram_raw from services.telegram_templates import triaged_alert, _send_telegram_raw
msg, keyboard = triaged_alert( msg, keyboard = triaged_alert(
@@ -807,13 +907,13 @@ class ElephantAlphaAutonomousEngine:
f"信心度:{decision.confidence:.2f} | " f"信心度:{decision.confidence:.2f} | "
f"參與模組:{', '.join(_AGENT_LABEL.get(a.lower(), a) for a in decision.agents_required)}" f"參與模組:{', '.join(_AGENT_LABEL.get(a.lower(), a) for a in decision.agents_required)}"
), ),
ai_actions=[ ai_actions=ai_actions_payload,
f"步驟 {s.get('step', i+1)}{_zh_step(s)}"
for i, s in enumerate(decision.execution_plan[:3])
] or ["無具體執行計畫"],
) )
await self._run_with_timeout(_send_telegram_raw, msg, timeout=10, reply_markup=keyboard) await self._run_with_timeout(_send_telegram_raw, msg, timeout=10, reply_markup=keyboard)
self._log.info("Human escalation Telegram sent: %s", trigger.trigger_type) self._log.info(
"Human escalation Telegram sent: %s (concrete=%s)",
trigger.trigger_type, bool(concrete_actions),
)
except Exception as e: except Exception as e:
self._log.error("Telegram escalation failed (non-blocking): %s", e) self._log.error("Telegram escalation failed (non-blocking): %s", e)

View File

@@ -45,6 +45,8 @@ class PriceThreat:
risk: str # HIGH / MED / LOW risk: str # HIGH / MED / LOW
recommended_action: str recommended_action: str
confidence: float confidence: float
sales_7d_curr_amount: float = 0.0 # 過去 7 日營收金額NT$),供下游金額影響量化
sales_7d_prev_amount: float = 0.0 # 前 7 日營收金額NT$),供「可挽回營收」估算
@dataclass @dataclass
@@ -375,6 +377,9 @@ class HermesAnalystService:
"pchome": pchome_price, "pchome": pchome_price,
"gap_pct": gap_pct, # Python 預算好Hermes 只做分類 "gap_pct": gap_pct, # Python 預算好Hermes 只做分類
"sales_delta": delta_pct, "sales_delta": delta_pct,
# 絕對營收金額(不傳給 Hermes 推理,只在 Python 端保留供下游金額影響量化)
"_sales_curr": sales_curr,
"_sales_prev": sales_prev,
} }
if raw_tags: if raw_tags:
item["competitor_tags"] = raw_tags # 語意情境給 Hermes 加分 item["competitor_tags"] = raw_tags # 語意情境給 Hermes 加分
@@ -386,10 +391,15 @@ class HermesAnalystService:
mcp_ctx = build_mcp_context(topics=["market_trends", "holiday_calendar", "seasonal_insights"]) mcp_ctx = build_mcp_context(topics=["market_trends", "holiday_calendar", "seasonal_insights"])
# 餵給 LLM 的版本:剝除底線開頭的 Python-only 欄位(避免 token 爆炸 + 防 LLM 嘗試引用)
items_for_llm = [
{k: v for k, v in item.items() if not k.startswith("_")}
for item in items
]
prompt = ( prompt = (
f"【市場外部情報 (MCP)】\n{mcp_ctx}\n\n" f"【市場外部情報 (MCP)】\n{mcp_ctx}\n\n"
f"分析以下 {len(items)} 支商品的競價威脅,回傳前 {TOP_N} 個最高風險商品。\n\n" f"分析以下 {len(items_for_llm)} 支商品的競價威脅,回傳前 {TOP_N} 個最高風險商品。\n\n"
f"資料:{json.dumps(items, ensure_ascii=False)}\n\n" f"資料:{json.dumps(items_for_llm, ensure_ascii=False)}\n\n"
f"輸出格式JSON 陣列,每筆含):\n" f"輸出格式JSON 陣列,每筆含):\n"
f'[{{"sku": string, "name": string, "category": string, ' f'[{{"sku": string, "name": string, "category": string, '
f'"momo_price": number, "pchome_price": number, ' f'"momo_price": number, "pchome_price": number, '
@@ -488,6 +498,9 @@ class HermesAnalystService:
risk=t.get("risk", "LOW"), # LLM 分類 risk=t.get("risk", "LOW"), # LLM 分類
recommended_action=t.get("recommended_action", ""), # LLM 洞察 recommended_action=t.get("recommended_action", ""), # LLM 洞察
confidence=float(t.get("confidence", 0.5)), # LLM 信心度 confidence=float(t.get("confidence", 0.5)), # LLM 信心度
# 絕對營收金額:純 Python truth供下游金額影響量化B' 軌)
sales_7d_curr_amount=float(ground.get("_sales_curr", 0) or 0),
sales_7d_prev_amount=float(ground.get("_sales_prev", 0) or 0),
)) ))
hermes_stats = getattr(self, "_last_stats", {}) hermes_stats = getattr(self, "_last_stats", {})

View File

@@ -259,6 +259,51 @@ TOOLS = [
] ]
# ── 金額影響量化B' 軌:告警必須攜帶可決策的金額數字) ──
def _compute_business_impact(threat) -> dict:
"""從 PriceThreat 計算「過去 7 日營收流失」與「建議調價金額」。
回傳純 Python 客觀計算結果,由 dispatcher 強制注入告警 — LLM 不得碰觸這些數字。
revenue_loss_7d
= max(0, sales_7d_prev_amount - sales_7d_curr_amount) **僅在 gap_pct > 0 時**
語意:「我方比競品貴,且過去 7 日銷量金額下滑 → 推估價格因素導致的流失」
若 gap_pct ≤ 0我方已便宜或持平即使銷量下滑亦歸 0避免把
季節性/商品壽命終結等非價格因素誤標為「流失」誘導降價Critic Medium-3 fix
recommended_price
= round(pchome_price)
語意:「跟進競品的最低調價金額」;統帥可基於此再依毛利策略加溢價
gap_pct ≤ 0我方已便宜或持平→ recommended_price=None不需調價
"""
try:
gap_pct = float(getattr(threat, "gap_pct", 0) or 0)
except (TypeError, ValueError):
gap_pct = 0.0
revenue_loss_7d = 0.0
if gap_pct > 0:
try:
prev = float(getattr(threat, "sales_7d_prev_amount", 0) or 0)
curr = float(getattr(threat, "sales_7d_curr_amount", 0) or 0)
revenue_loss_7d = max(0.0, prev - curr)
except (TypeError, ValueError):
revenue_loss_7d = 0.0
recommended_price = None
try:
pchome = float(getattr(threat, "pchome_price", 0) or 0)
if pchome > 0 and gap_pct > 0:
recommended_price = round(pchome)
except (TypeError, ValueError):
pass
return {
"revenue_loss_7d": revenue_loss_7d,
"recommended_price": recommended_price,
}
# ── 語意化 Emoji 字典 ────────────────────────────────── # ── 語意化 Emoji 字典 ──────────────────────────────────
# 身份識別 # 身份識別
HEADER_DISPATCHER = "⚡ NemoTron 派發器" HEADER_DISPATCHER = "⚡ NemoTron 派發器"
@@ -515,6 +560,10 @@ class NemotronDispatcher:
for t in threats: for t in threats:
try: try:
# B' 軌:每個 threat 預先算金額影響,所有路徑統一注入
impact = _compute_business_impact(t)
rl, rp = impact["revenue_loss_7d"], impact["recommended_price"]
if t.gap_pct < 5 and t.sales_7d_delta_pct < -30: if t.gap_pct < 5 and t.sales_7d_delta_pct < -30:
# Rule 1價差微小但銷量大跌 → 非定價問題,人工確認 # Rule 1價差微小但銷量大跌 → 非定價問題,人工確認
self._exec_flag_for_human_review( self._exec_flag_for_human_review(
@@ -528,6 +577,7 @@ class NemotronDispatcher:
footprint=footprint, footprint=footprint,
momo_price=t.momo_price, comp_price=t.pchome_price, momo_price=t.momo_price, comp_price=t.pchome_price,
gap_pct=t.gap_pct, sales_delta=t.sales_7d_delta_pct, gap_pct=t.gap_pct, sales_delta=t.sales_7d_delta_pct,
revenue_loss_7d=rl, recommended_price=rp,
) )
elif t.gap_pct >= 5 and t.risk == "HIGH": elif t.gap_pct >= 5 and t.risk == "HIGH":
# Rule 2高價差 HIGH 風險 → 競價告警 # Rule 2高價差 HIGH 風險 → 競價告警
@@ -538,6 +588,7 @@ class NemotronDispatcher:
t.confidence, t.confidence,
momo_price=t.momo_price, comp_price=t.pchome_price, momo_price=t.momo_price, comp_price=t.pchome_price,
footprint=footprint, footprint=footprint,
revenue_loss_7d=rl, recommended_price=rp,
) )
elif t.gap_pct < 0 and t.sales_7d_delta_pct > 0: elif t.gap_pct < 0 and t.sales_7d_delta_pct > 0:
# Rule 3我方具競爭力 + 銷量正成長 → 推薦 # Rule 3我方具競爭力 + 銷量正成長 → 推薦
@@ -563,6 +614,7 @@ class NemotronDispatcher:
footprint=footprint, footprint=footprint,
momo_price=t.momo_price, comp_price=t.pchome_price, momo_price=t.momo_price, comp_price=t.pchome_price,
gap_pct=t.gap_pct, sales_delta=t.sales_7d_delta_pct, gap_pct=t.gap_pct, sales_delta=t.sales_7d_delta_pct,
revenue_loss_7d=rl, recommended_price=rp,
) )
dispatched += 1 dispatched += 1
except Exception as e: except Exception as e:
@@ -585,16 +637,22 @@ class NemotronDispatcher:
gap_pct: float, sales_delta: float, gap_pct: float, sales_delta: float,
action: str, confidence: float, action: str, confidence: float,
footprint: str, footprint: str,
revenue_loss_7d: float = 0.0,
recommended_price: Optional[float] = None,
) -> str: ) -> str:
""" """
類別一:緊急告警 類別一:緊急告警
倒金字塔:結論先行 → 核心數據 → AI 洞察 → 建議行動 → 運算足跡 倒金字塔:結論先行 → 核心數據 → 金額影響 → AI 洞察 → 運算足跡
[2026-04-18 台北] Bug-3 防線三 UI 物理隔離: [2026-04-18 台北] Bug-3 防線三 UI 物理隔離:
- 核心問題 = Python 客觀組字(價差 X% / 銷量 Y%),不碰 AI 文字 - 核心問題 = Python 客觀組字(價差 X% / 銷量 Y%),不碰 AI 文字
- 關鍵數據 = Python 獨裁注入None/0 降級為 N/A 而非 $0 - 關鍵數據 = Python 獨裁注入None/0 降級為 N/A 而非 $0
- AI 洞察 = action 唯一使用位置移除假冒「Hermes 分析師研判」標籤 - AI 洞察 = action 唯一使用位置移除假冒「Hermes 分析師研判」標籤
action 實為 NemoTron 輸出,非 Hermes — Claude Opus 4.7 action 實為 NemoTron 輸出,非 Hermes — Claude Opus 4.7
[2026-05-02 台北] B' 軌:金額影響量化 — Claude Opus 4.7
- revenue_loss_7d / recommended_price 純 Python 計算_compute_business_impact
- 解決「告警內容空泛、人類無可批准的具體動作」根因
""" """
conf_pct = int(confidence * 100) conf_pct = int(confidence * 100)
@@ -605,6 +663,17 @@ class NemotronDispatcher:
# 核心問題Python 客觀組字,不碰 AI 文字 # 核心問題Python 客觀組字,不碰 AI 文字
core_issue = f"價差 {gap_pct:+.1f}% / 近七天銷量 {sales_delta:+.1f}%" core_issue = f"價差 {gap_pct:+.1f}% / 近七天銷量 {sales_delta:+.1f}%"
# 金額影響區塊B' 軌新增)
impact_lines = []
if revenue_loss_7d and revenue_loss_7d > 0:
impact_lines.append(f"📉 過去 7 日營收流失NT$ {revenue_loss_7d:,.0f}")
if recommended_price is not None and recommended_price > 0:
impact_lines.append(
f"🎯 跟進競品建議價NT$ {recommended_price:,.0f}"
f"(毛利策略可再加溢價)"
)
impact_block = ("\n".join(impact_lines) + "\n\n") if impact_lines else ""
# AI 洞察:唯一允許 LLM 文字進入的欄位 # AI 洞察:唯一允許 LLM 文字進入的欄位
ai_insight = _sanitize_text(action, fallback="請人工評估議價空間") ai_insight = _sanitize_text(action, fallback="請人工評估議價空間")
@@ -616,6 +685,7 @@ class NemotronDispatcher:
f"• 我方價格:{mp_str}\n" f"• 我方價格:{mp_str}\n"
f"• 競品價格:{cp_str}\n" f"• 競品價格:{cp_str}\n"
f"• 銷量變化:{sales_delta:+.1f}%\n\n" f"• 銷量變化:{sales_delta:+.1f}%\n\n"
f"{impact_block}"
f"{ICON_AI} AI 洞察(信心度 {conf_pct}%\n" f"{ICON_AI} AI 洞察(信心度 {conf_pct}%\n"
f"{ai_insight}\n\n" f"{ai_insight}\n\n"
f"{footprint}" f"{footprint}"
@@ -627,10 +697,13 @@ class NemotronDispatcher:
concern: str, footprint: str, concern: str, footprint: str,
momo_price: float = None, comp_price: float = None, momo_price: float = None, comp_price: float = None,
gap_pct: float = None, sales_delta: float = None, gap_pct: float = None, sales_delta: float = None,
revenue_loss_7d: float = 0.0,
recommended_price: Optional[float] = None,
) -> str: ) -> str:
""" """
類別二:人工覆核 類別二:人工覆核
客觀數據由 Python 注入防幻覺AI 診斷隔離在獨立欄位 客觀數據由 Python 注入防幻覺AI 診斷隔離在獨立欄位
B' 軌:補金額影響欄位
""" """
# 客觀數據快照100% Python不經 LLM # 客觀數據快照100% Python不經 LLM
if momo_price is not None and comp_price is not None: if momo_price is not None and comp_price is not None:
@@ -643,10 +716,19 @@ class NemotronDispatcher:
else: else:
data_block = f"{ICON_REPORT} 客觀數據:(無競品比價數據)\n" data_block = f"{ICON_REPORT} 客觀數據:(無競品比價數據)\n"
# 金額影響B' 軌新增)
impact_lines = []
if revenue_loss_7d and revenue_loss_7d > 0:
impact_lines.append(f"📉 過去 7 日營收流失NT$ {revenue_loss_7d:,.0f}")
if recommended_price is not None and recommended_price > 0:
impact_lines.append(f"🎯 跟進競品建議價NT$ {recommended_price:,.0f}")
impact_block = ("\n".join(f"{l}" for l in impact_lines) + "\n") if impact_lines else ""
return ( return (
f"{ICON_WARNING} [{HEADER_DISPATCHER}] 異常波動需人工覆核\n\n" f"{ICON_WARNING} [{HEADER_DISPATCHER}] 異常波動需人工覆核\n\n"
f"🔍 待查商品:[{sku}] {name}\n\n" f"🔍 待查商品:[{sku}] {name}\n\n"
f"{data_block}\n" f"{data_block}"
f"{impact_block}\n"
f"🧠 AI 診斷:\n" f"🧠 AI 診斷:\n"
f"{concern}\n\n" f"{concern}\n\n"
f"👉 建議行動:請營運人員立即進行前台走查。\n\n" f"👉 建議行動:請營運人員立即進行前台走查。\n\n"
@@ -686,26 +768,38 @@ class NemotronDispatcher:
gap_pct: float, sales_delta: float, action: str, confidence: float, gap_pct: float, sales_delta: float, action: str, confidence: float,
momo_price=None, comp_price=None, momo_price=None, comp_price=None,
footprint: str = "", footprint: str = "",
revenue_loss_7d: float = 0.0,
recommended_price: Optional[float] = None,
): ):
"""發送語意化競價高危險預警 """發送語意化競價高危險預警
[2026-04-18 台北] Bug-1 防線一 保險default 改 None避免 LLM 漏吐 [2026-04-18 台北] Bug-1 防線一 保險default 改 None避免 LLM 漏吐
→ 舊版 default=0 → Telegram 顯示 $0。Layer A Hermes 已根治,這層是第二道屏障 → 舊版 default=0 → Telegram 顯示 $0。Layer A Hermes 已根治,這層是第二道屏障
— Claude Opus 4.7 — Claude Opus 4.7
[2026-05-02 台北] B'revenue_loss_7d / recommended_price 純 Python 注入
— Claude Opus 4.7
""" """
msg = self._fmt_price_alert( msg = self._fmt_price_alert(
sku, name, momo_price, comp_price, sku, name, momo_price, comp_price,
gap_pct, sales_delta, action, confidence, footprint, gap_pct, sales_delta, action, confidence, footprint,
revenue_loss_7d=revenue_loss_7d,
recommended_price=recommended_price,
) )
self._send_telegram(msg) self._send_telegram(msg)
logger.info(f"[Dispatcher] 競價告警 → {sku} gap={gap_pct:.1f}% sales={sales_delta:.1f}%") logger.info(
f"[Dispatcher] 競價告警 → {sku} gap={gap_pct:.1f}% sales={sales_delta:.1f}% "
f"loss=${revenue_loss_7d:,.0f} rec_price={recommended_price}"
)
# ADR-007 雙寫:沉澱到 ai_insights 供日後 RAG # ADR-007 雙寫:沉澱到 ai_insights 供日後 RAG
self._sink_insight_to_km( self._sink_insight_to_km(
insight_type="price_alert", insight_type="price_alert",
sku=sku, name=name, sku=sku, name=name,
content=f"[高危險告警] {name} 價差 {gap_pct:+.1f}% / 銷量 {sales_delta:+.1f}%。行動:{action}", content=f"[高危險告警] {name} 價差 {gap_pct:+.1f}% / 銷量 {sales_delta:+.1f}%。行動:{action}",
metadata={"gap_pct": gap_pct, "sales_delta": sales_delta, "confidence": confidence, metadata={"gap_pct": gap_pct, "sales_delta": sales_delta, "confidence": confidence,
"momo_price": momo_price, "comp_price": comp_price}, "momo_price": momo_price, "comp_price": comp_price,
"revenue_loss_7d": revenue_loss_7d,
"recommended_price": recommended_price},
) )
def _exec_add_to_recommendation( def _exec_add_to_recommendation(
@@ -776,6 +870,8 @@ class NemotronDispatcher:
footprint: str = "", footprint: str = "",
momo_price: float = None, comp_price: float = None, momo_price: float = None, comp_price: float = None,
gap_pct: float = None, sales_delta: float = None, gap_pct: float = None, sales_delta: float = None,
revenue_loss_7d: float = 0.0,
recommended_price: Optional[float] = None,
): ):
"""發送語意化人工覆核請求""" """發送語意化人工覆核請求"""
concern = _sanitize_text(concern, fallback=f"數據走勢違背常理,疑似缺貨或前台異常。") concern = _sanitize_text(concern, fallback=f"數據走勢違背常理,疑似缺貨或前台異常。")
@@ -783,16 +879,22 @@ class NemotronDispatcher:
sku, name, concern, footprint, sku, name, concern, footprint,
momo_price=momo_price, comp_price=comp_price, momo_price=momo_price, comp_price=comp_price,
gap_pct=gap_pct, sales_delta=sales_delta, gap_pct=gap_pct, sales_delta=sales_delta,
revenue_loss_7d=revenue_loss_7d,
recommended_price=recommended_price,
) )
self._send_telegram(msg) self._send_telegram(msg)
logger.info(f"[Dispatcher] 人工覆核請求 → {sku}") logger.info(
f"[Dispatcher] 人工覆核請求 → {sku} loss=${revenue_loss_7d:,.0f}"
)
# ADR-007 雙寫 # ADR-007 雙寫
self._sink_insight_to_km( self._sink_insight_to_km(
insight_type="human_review", insight_type="human_review",
sku=sku, name=name, sku=sku, name=name,
content=f"[人工覆核] {name}。疑慮:{concern}", content=f"[人工覆核] {name}。疑慮:{concern}",
metadata={"confidence": confidence, "gap_pct": gap_pct, "sales_delta": sales_delta, metadata={"confidence": confidence, "gap_pct": gap_pct, "sales_delta": sales_delta,
"momo_price": momo_price, "comp_price": comp_price}, "momo_price": momo_price, "comp_price": comp_price,
"revenue_loss_7d": revenue_loss_7d,
"recommended_price": recommended_price},
) )
def _exec_route_to_km( def _exec_route_to_km(
@@ -1059,6 +1161,7 @@ class NemotronDispatcher:
"疑似缺貨、下架、或前台異常,請營運人員立即走查。" "疑似缺貨、下架、或前台異常,請營運人員立即走查。"
) )
try: try:
impact = _compute_business_impact(t)
self._exec_flag_for_human_review( self._exec_flag_for_human_review(
sku=t.sku, sku=t.sku,
name=t.name, name=t.name,
@@ -1069,6 +1172,8 @@ class NemotronDispatcher:
comp_price=t.pchome_price, comp_price=t.pchome_price,
gap_pct=t.gap_pct, gap_pct=t.gap_pct,
sales_delta=t.sales_7d_delta_pct, sales_delta=t.sales_7d_delta_pct,
revenue_loss_7d=impact["revenue_loss_7d"],
recommended_price=impact["recommended_price"],
) )
dispatched += 1 dispatched += 1
except Exception as e: except Exception as e:
@@ -1174,17 +1279,24 @@ class NemotronDispatcher:
# [2026-04-18 台北] Bug-1 防線一 保險:所有客觀數字強制由 Python 從 threat_map 注入, # [2026-04-18 台北] Bug-1 防線一 保險:所有客觀數字強制由 Python 從 threat_map 注入,
# 覆蓋 LLM 可能回吐的幻覺數字(例如 $0。Layer A Hermes 根治是主防線, # 覆蓋 LLM 可能回吐的幻覺數字(例如 $0。Layer A Hermes 根治是主防線,
# 此處為二道屏障(萬一 ground_items 有漏網,或未來走 bypass — Claude Opus 4.7 # 此處為二道屏障(萬一 ground_items 有漏網,或未來走 bypass — Claude Opus 4.7
# [2026-05-02 台北] B' 軌:金額影響量化亦走 Python 獨裁注入 — Claude Opus 4.7
t = threat_map.get(args.get("sku")) t = threat_map.get(args.get("sku"))
if tool_name == "trigger_price_alert" and t: if tool_name == "trigger_price_alert" and t:
args["momo_price"] = getattr(t, "momo_price", None) args["momo_price"] = getattr(t, "momo_price", None)
args["comp_price"] = getattr(t, "pchome_price", None) args["comp_price"] = getattr(t, "pchome_price", None)
args["gap_pct"] = getattr(t, "gap_pct", None) args["gap_pct"] = getattr(t, "gap_pct", None)
args["sales_delta"] = getattr(t, "sales_7d_delta_pct", None) args["sales_delta"] = getattr(t, "sales_7d_delta_pct", None)
impact = _compute_business_impact(t)
args["revenue_loss_7d"] = impact["revenue_loss_7d"]
args["recommended_price"] = impact["recommended_price"]
elif tool_name == "flag_for_human_review" and t: elif tool_name == "flag_for_human_review" and t:
args["momo_price"] = getattr(t, "momo_price", None) args["momo_price"] = getattr(t, "momo_price", None)
args["comp_price"] = getattr(t, "pchome_price", None) args["comp_price"] = getattr(t, "pchome_price", None)
args["gap_pct"] = getattr(t, "gap_pct", None) args["gap_pct"] = getattr(t, "gap_pct", None)
args["sales_delta"] = getattr(t, "sales_7d_delta_pct", None) args["sales_delta"] = getattr(t, "sales_7d_delta_pct", None)
impact = _compute_business_impact(t)
args["revenue_loss_7d"] = impact["revenue_loss_7d"]
args["recommended_price"] = impact["recommended_price"]
elif tool_name == "add_to_recommendation": elif tool_name == "add_to_recommendation":
args["footprint_data"] = footprint_data args["footprint_data"] = footprint_data
args["threat"] = t args["threat"] = t
@@ -1235,17 +1347,22 @@ if __name__ == "__main__":
risk: str risk: str
recommended_action: str recommended_action: str
confidence: float confidence: float
sales_7d_curr_amount: float = 0.0
sales_7d_prev_amount: float = 0.0
fake_threats = [ fake_threats = [
FakeThreat("A003", "舒特膚AD乳液200ml", "美妝保養", FakeThreat("A003", "舒特膚AD乳液200ml", "美妝保養",
1200, 980, 22.4, -35.0, "HIGH", 1200, 980, 22.4, -35.0, "HIGH",
"建議立即降價至 $1,000 迎戰,或發放 $200 專屬折價券", 0.85), "建議立即降價至 $1,000 迎戰,或發放 $200 專屬折價券", 0.85,
sales_7d_curr_amount=78000, sales_7d_prev_amount=120000),
FakeThreat("A001", "玻尿酸面膜10片裝", "美妝保養", FakeThreat("A001", "玻尿酸面膜10片裝", "美妝保養",
320, 280, 14.3, -42.0, "HIGH", 320, 280, 14.3, -42.0, "HIGH",
"建議跟進降價至 $285配合限時加購活動", 0.78), "建議跟進降價至 $285配合限時加購活動", 0.78,
sales_7d_curr_amount=58000, sales_7d_prev_amount=100000),
FakeThreat("A009", "美白化妝水150ml", "美妝保養", FakeThreat("A009", "美白化妝水150ml", "美妝保養",
420, 350, 20.0, -22.0, "HIGH", 420, 350, 20.0, -22.0, "HIGH",
"價格差距過大,建議優先調降或捆包促銷", 0.45), "價格差距過大,建議優先調降或捆包促銷", 0.45,
sales_7d_curr_amount=78000, sales_7d_prev_amount=100000),
] ]
# 模擬 Hermes 運算足跡 # 模擬 Hermes 運算足跡
@@ -1259,11 +1376,13 @@ if __name__ == "__main__":
footprint = _build_footprint_block(fake_hermes_stats, fake_nim_stats) footprint = _build_footprint_block(fake_hermes_stats, fake_nim_stats)
# 測試三種訊息格式 # 測試三種訊息格式
print("=== 類別一:緊急告警 ===") print("=== 類別一:緊急告警(含 B' 金額影響) ===")
print(NemotronDispatcher._fmt_price_alert( print(NemotronDispatcher._fmt_price_alert(
"A003", "舒特膚AD乳液200ml", 1200, 980, "A003", "舒特膚AD乳液200ml", 1200, 980,
22.4, -35.0, "建議立即降價至 $1,000 迎戰,或發放 $200 專屬折價券", 22.4, -35.0, "建議立即降價至 $1,000 迎戰,或發放 $200 專屬折價券",
0.85, footprint, 0.85, footprint,
revenue_loss_7d=42000.0, # B' 軌驗證120k - 78k = 42k 流失
recommended_price=980, # B' 軌驗證:跟進競品價
)) ))
print() print()
print("=== 類別二:人工覆核 ===") print("=== 類別二:人工覆核 ===")

View File

@@ -509,6 +509,12 @@ class TrendTelegramBot:
await query.answer() await query.answer()
# ADR-012 / A' 軌HITL escalation 按鈕momo:eig:{event_id} = event_ignore
# triaged_alert 鍵盤按鈕「🛑 忽略此事件」會 callback momo:eig:<id>
if data.startswith('momo:eig:'):
await self._handle_event_ignore_callback(query, data)
return
if data.startswith(('menu:', 'cmd:', 'await:')): if data.startswith(('menu:', 'cmd:', 'await:')):
await self._handle_openclaw_callback(query, context, data) await self._handle_openclaw_callback(query, context, data)
return return
@@ -579,6 +585,88 @@ class TrendTelegramBot:
elif data.startswith("settings_"): elif data.startswith("settings_"):
await self._handle_settings_callback(query, data) await self._handle_settings_callback(query, data)
async def _handle_event_ignore_callback(self, query, data: str):
"""A'HITL 升級審核「🛑 忽略此事件」按鈕處理。
callback_data 格式momo:eig:<event_id>
動作:
1. 寫入 ai_insights 一筆 status='ignored' 紀錄(含 event_id 供 audit
2. 編輯原訊息:標記「已忽略 by <user> @ <ts>」
3. 失敗 best-effort不阻斷 user 體驗
ADR-012 §③:人工決策必有 audit trailai_insights 為 SOT。
Critic Critical-1 fix: user_label / ts_label 寫入 HTML 前須 escape
Critic Medium-2 fix: 空 event_id 直接拒絕,避免污染 audit 資料。
"""
from datetime import datetime as _dt
from html import escape as _html_escape
# Critic Medium-2: 空 event_id 即拒絕(避免攻擊者送 `momo:eig:` 污染 audit
parts_in = data.split(':', 2)
event_id = parts_in[2].strip() if len(parts_in) >= 3 else ''
if not event_id:
try:
await query.answer("event_id 缺失,忽略動作未生效", show_alert=False)
except Exception:
pass
logger.warning("[EA HITL] empty event_id callback rejected: %r", data)
return
user = getattr(query, 'from_user', None)
user_label_raw = (
getattr(user, 'username', None)
or getattr(user, 'first_name', None)
or str(getattr(user, 'id', '?'))
)
ts_label_raw = _dt.now().strftime('%Y-%m-%d %H:%M')
try:
from database.manager import get_session
from sqlalchemy import text
session = get_session()
try:
session.execute(
text("""
INSERT INTO ai_insights
(insight_type, content, confidence, created_by, status, metadata_json)
VALUES (:type, :content, :conf, :by, :status, :meta)
"""),
{
"type": "human_review",
"content": f"[EA HITL] 事件 {event_id}{user_label_raw} 忽略",
"conf": 1.0,
"by": f"telegram:{user_label_raw}",
"status": "ignored",
"meta": json.dumps({
"event_id": event_id,
"decided_by": user_label_raw,
"decided_at": ts_label_raw,
"decision": "ignored",
}, ensure_ascii=False),
},
)
session.commit()
finally:
session.close()
except Exception as audit_err:
logger.warning(f"[EA HITL] ai_insights audit 寫入失敗(不阻斷 UI: {audit_err}")
# Critic Critical-1: user_label / ts_label 須 HTML escape避免攻擊者
# 透過 Telegram username 注入 <a>/<pre>/破版標籤
user_label_safe = _html_escape(str(user_label_raw))
ts_label_safe = _html_escape(ts_label_raw)
try:
original = query.message.text_html if query.message and getattr(query.message, 'text_html', None) else (query.message.text if query.message else "")
footer = f"\n\n🛑 <b>已忽略</b> by {user_label_safe} @ {ts_label_safe}"
await query.edit_message_text(
(original or '事件已忽略') + footer,
parse_mode='HTML',
)
except Exception as ui_err:
logger.debug(f"[EA HITL] edit_message 失敗(不阻斷): {ui_err}")
logger.info(f"[EA HITL] event_ignore event_id={event_id} by={user_label_raw}")
async def _handle_openclaw_callback(self, query, context, data: str): async def _handle_openclaw_callback(self, query, context, data: str):
"""轉接 OpenClaw 完整菜單 callback避免長輪詢 Bot 吃掉 /menu。""" """轉接 OpenClaw 完整菜單 callback避免長輪詢 Bot 吃掉 /menu。"""
chat_id = query.message.chat_id chat_id = query.message.chat_id