[V10.278] 補市場情報 AI 摘要 preflight 與競價摘要快取 | market_intel, competitor_intel

TODO_NEXT_STEPS.txt

@@ -4,6 +4,7 @@
 ================================================================================
 
 【已完成】
+   - V10.278 補 PChome 競價摘要 30 分鐘共享快取與 feeder/backfill 主動清除，並新增市場情報 `candidate_queue_review_ai_summary_preflight` 預覽 gate；API 只檢查未來摘要輸入與 Ollama-first/Gemini-backup-only policy，不呼叫 LLM、不派 Telegram、不寫 DB、不掛 scheduler。
    - V10.276 修正 ElephantAlpha 價格類 Hermes prefetch timeout：`price_drop` / `market_opportunity` trigger 直接把 SQL 命中的 MOMO / PChome 價差實證轉成 HITL action lines，完整 Hermes LLM prefetch 預設關閉；無 DB 實證仍只記 suppressed telemetry / cooldown，不寫 `human_review`、不發空 Telegram。
    - V10.266 強化核心 MOMO/PChome 比價鏈路：新增 `marketplace_product_matcher.py` 身份比對、只讓 `identity_v2` 且分數 ≥ 0.76 的高信心配對進 Dashboard/AI/Excel/Daily/Growth/PPT，並建立 `competitor_intel_repository.py` 統一圖表與簡報資料出口；同品牌但不同型號、不同組數、套組/單品或多品項不一致會進待審，不進正式比價。
    - V10.267 專業化 ElephantAlpha `resource_optimization` 告警：不再讓 LLM 生成「48 小時預期效益 / 已執行」敘事，改由程式量測 action queue、P1/P2、pending_review、逾時項目與 CPU load；單純 backlog 不發 Telegram，只有可行動資源壓力才寫 `ai_insights(resource_pressure)` 並發送量測型告警。

routes/market_intel_review_post_routes.py

@@ -11,6 +11,9 @@ from routes.market_intel_review_routes import (
     market_intel_review_bp,
 )
 from services.market_intel import MarketIntelService
+from services.market_intel.candidate_queue_review_ai_summary_preflight import (
+    build_candidate_queue_review_ai_summary_preflight,
+)
 from services.market_intel.candidate_queue_review_archive_summary import (
     build_candidate_queue_review_archive_summary,
 )
@@ -106,6 +109,37 @@ def _build_review_completion_archive_stack(
     return transaction, receipt, closeout, inventory, archive
 
 
+def _build_review_archive_summary_stack(
+    *,
+    service,
+    sample_result,
+    payload_error,
+    operator_evidence,
+    writer_output,
+    smoke_result,
+    limit,
+    execute_requested,
+):
+    transaction, receipt, closeout, inventory, archive = (
+        _build_review_completion_archive_stack(
+            service=service,
+            sample_result=sample_result,
+            payload_error=payload_error,
+            operator_evidence=operator_evidence,
+            writer_output=writer_output,
+            smoke_result=smoke_result,
+            limit=limit,
+            execute_requested=execute_requested,
+        )
+    )
+    archive_summary = build_candidate_queue_review_archive_summary(
+        review_completion_archive=archive,
+        operator_evidence=operator_evidence,
+        execute_requested=execute_requested,
+    )
+    return transaction, receipt, closeout, inventory, archive, archive_summary
+
+
 @market_intel_review_bp.route(
     "/api/market_intel/manual_sample_review/"
     "candidate_queue_review_decision_post_closeout_inventory",
@@ -145,8 +179,8 @@ def market_intel_manual_sample_candidate_queue_review_archive_summary():
     sample_result, operator_evidence, writer_output, smoke_result, payload_error, limit = (
         _extract_run_payload()
     )
-    transaction, receipt, closeout, inventory, archive = (
-        _build_review_completion_archive_stack(
+    transaction, receipt, closeout, inventory, archive, archive_summary = (
+        _build_review_archive_summary_stack(
             service=service,
             sample_result=sample_result,
             payload_error=payload_error,
@@ -157,8 +191,37 @@ def market_intel_manual_sample_candidate_queue_review_archive_summary():
             execute_requested=execute_requested,
         )
     )
-    data = build_candidate_queue_review_archive_summary(
-        review_completion_archive=archive,
+    data = archive_summary
+    data["phase"] = service.phase
+    return jsonify(data), 400 if payload_error else 200
+
+
+@market_intel_review_bp.route(
+    "/api/market_intel/manual_sample_review/"
+    "candidate_queue_review_ai_summary_preflight",
+    methods=["POST"],
+)
+@login_required
+def market_intel_manual_sample_candidate_queue_review_ai_summary_preflight():
+    service = MarketIntelService()
+    execute_requested = request.args.get("execute", "false").lower() == "true"
+    sample_result, operator_evidence, writer_output, smoke_result, payload_error, limit = (
+        _extract_run_payload()
+    )
+    transaction, receipt, closeout, inventory, archive, archive_summary = (
+        _build_review_archive_summary_stack(
+            service=service,
+            sample_result=sample_result,
+            payload_error=payload_error,
+            operator_evidence=operator_evidence,
+            writer_output=writer_output,
+            smoke_result=smoke_result,
+            limit=limit,
+            execute_requested=execute_requested,
+        )
+    )
+    data = build_candidate_queue_review_ai_summary_preflight(
+        archive_summary=archive_summary,
         operator_evidence=operator_evidence,
         execute_requested=execute_requested,
     )

services/market_intel/candidate_queue_review_ai_summary_preflight.py

@@ -0,0 +1,365 @@
+"""候選審核 queue AI summary preflight 預覽。
+
+本模組只檢查 archive summary input 是否足夠進入未來的 Ollama-first 摘要流程；
+不呼叫 LLM、不派送 Telegram、不讀 approval token、不執行 CLI、不寫檔、
+不寫 DB、不更新 review_state、不 commit、不掛 scheduler。
+"""
+
+
+FORBIDDEN_TOKEN_KEYWORDS = (
+    "approval_token",
+    "approval-token",
+    "market_intel_queue_write_approval",
+)
+SAFE_TOKEN_METADATA_KEYS = {
+    "approval_token_present",
+    "approval_token_valid",
+    "approval_token_secret_configured",
+}
+SAFE_APPROVAL_ENV_VAR = "MARKET_INTEL_QUEUE_WRITE_APPROVAL"
+TARGET_TABLE = "market_alert_review_queue"
+OLLAMA_CASCADE = (
+    {"key": "gcp_a", "label": "GCP-A", "host": "34.143.170.20:11434"},
+    {"key": "gcp_b", "label": "GCP-B", "host": "34.21.145.224:11434"},
+    {"key": "lan_111", "label": "111", "host": "192.168.0.111:11434"},
+)
+
+
+def _as_dict(value):
+    return value if isinstance(value, dict) else {}
+
+
+def _as_list(value):
+    if value is None:
+        return []
+    if isinstance(value, (list, tuple, set)):
+        return list(value)
+    return [value]
+
+
+def _contains_forbidden_token_key(value):
+    if isinstance(value, dict):
+        for key, nested in value.items():
+            normalized_key = str(key).lower()
+            if normalized_key in SAFE_TOKEN_METADATA_KEYS and isinstance(nested, bool):
+                continue
+            if normalized_key == "approval_env_var" and nested == SAFE_APPROVAL_ENV_VAR:
+                continue
+            if any(token_key in normalized_key for token_key in FORBIDDEN_TOKEN_KEYWORDS):
+                return True
+            if _contains_forbidden_token_key(nested):
+                return True
+    elif isinstance(value, list):
+        return any(_contains_forbidden_token_key(item) for item in value)
+    return False
+
+
+def _safe_text(value, limit=300):
+    if value is None:
+        return None
+    text = str(value).strip()
+    return text[:limit] if text else None
+
+
+def _operator_summary(operator_evidence):
+    operator_evidence = _as_dict(operator_evidence)
+    return {
+        "provided_keys": sorted(operator_evidence.keys()),
+        "operator_confirmed_ai_summary_preflight": bool(
+            operator_evidence.get("operator_confirmed_ai_summary_preflight")
+            or operator_evidence.get("operator_confirmed_ai_summary_review")
+        ),
+        "operator_confirmed_ollama_first": bool(
+            operator_evidence.get("operator_confirmed_ollama_first")
+            or operator_evidence.get("operator_confirmed_ollama_cascade")
+        ),
+        "operator_confirmed_gemini_backup_only": bool(
+            operator_evidence.get("operator_confirmed_gemini_backup_only")
+            or operator_evidence.get("operator_confirmed_gemini_fallback_only")
+        ),
+        "operator_confirmed_no_llm_call": bool(
+            operator_evidence.get("operator_confirmed_no_llm_call")
+        ),
+        "operator_confirmed_no_telegram_dispatch": bool(
+            operator_evidence.get("operator_confirmed_no_telegram_dispatch")
+        ),
+        "operator_confirmed_no_scheduler_attach": bool(
+            operator_evidence.get("operator_confirmed_no_scheduler_attach")
+        ),
+        "operator_confirmed_no_api_db_write": bool(
+            operator_evidence.get("operator_confirmed_no_api_db_write")
+        ),
+        "ai_summary_preflight_artifact_path": _safe_text(
+            operator_evidence.get("ai_summary_preflight_artifact_path")
+            or operator_evidence.get("summary_preflight_artifact_path")
+        ),
+        "approval_token_submitted_to_api": _contains_forbidden_token_key(
+            operator_evidence
+        ),
+    }
+
+
+def _summary_payload(archive_summary):
+    summary = _as_dict(archive_summary)
+    sections = _as_list(summary.get("summary_sections"))
+    artifact_paths = _as_dict(summary.get("artifact_paths"))
+    expected_keys = sorted(str(key) for key in _as_list(summary.get("expected_dedupe_keys")))
+    found_keys = sorted(str(key) for key in _as_list(summary.get("found_dedupe_keys")))
+    missing_keys = sorted(str(key) for key in _as_list(summary.get("missing_dedupe_keys")))
+    return {
+        "provided": bool(summary),
+        "mode": summary.get("mode"),
+        "archive_summary_ready": bool(summary.get("archive_summary_ready")),
+        "summary_input_ready": bool(summary.get("summary_input_ready")),
+        "ready_for_ai_summary_review": bool(summary.get("ready_for_ai_summary_review")),
+        "expected_dedupe_keys": expected_keys,
+        "found_dedupe_keys": found_keys,
+        "missing_dedupe_keys": missing_keys,
+        "state_mismatches": _as_list(summary.get("state_mismatches")),
+        "row_summary_count": int(summary.get("row_summary_count") or 0),
+        "summary_sections": sections,
+        "summary_section_keys": [
+            _as_dict(section).get("key")
+            for section in sections
+            if _as_dict(section).get("key")
+        ],
+        "artifact_paths": artifact_paths,
+        "read_only_query_executed": bool(summary.get("read_only_query_executed")),
+        "database_connection_opened": bool(summary.get("database_connection_opened")),
+        "database_write_executed": bool(summary.get("database_write_executed")),
+        "database_commit_executed": bool(summary.get("database_commit_executed")),
+        "review_state_update_executed": bool(summary.get("review_state_update_executed")),
+        "scheduler_attached": bool(summary.get("scheduler_attached")),
+        "llm_call_executed": bool(summary.get("llm_call_executed")),
+        "telegram_dispatched": bool(summary.get("telegram_dispatched")),
+        "ai_summary_generated": bool(summary.get("ai_summary_generated")),
+        "blocked_reasons": _as_list(summary.get("blocked_reasons")),
+    }
+
+
+def _side_effects_clear(summary):
+    blocked_keys = (
+        "api_executes_cli",
+        "api_writes_file",
+        "api_writes_database",
+        "api_updates_review_state",
+        "summary_file_written",
+        "summary_record_written",
+        "summary_manifest_written",
+        "database_write_executed",
+        "database_commit_executed",
+        "review_state_update_executed",
+        "llm_call_executed",
+        "ollama_call_executed",
+        "gemini_call_executed",
+        "telegram_dispatched",
+        "ai_summary_generated",
+        "external_network_executed",
+        "scheduler_attached",
+        "writes_executed",
+        "would_write_database",
+    )
+    return all(not _as_dict(summary).get(key) for key in blocked_keys)
+
+
+def _model_route_policy():
+    return {
+        "primary_policy": "ollama_first",
+        "primary_cascade": list(OLLAMA_CASCADE),
+        "fallback_policy": "gemini_backup_only_after_ollama_cascade_failure",
+        "app_host_forbidden_as_ollama_node": "192.168.0.188",
+        "api_executes_llm": False,
+        "api_dispatches_telegram": False,
+        "expected_future_caller": "market_intel_review_ai_summary",
+    }
+
+
+def _preflight_gates(*, summary, operator, route_policy):
+    required_sections = {
+        "review_scope",
+        "review_state_result",
+        "evidence_artifacts",
+        "execution_safety",
+    }
+    section_keys = set(summary["summary_section_keys"])
+    artifact_paths = summary["artifact_paths"]
+    return [
+        {
+            "key": "archive_summary_preview_provided",
+            "label": "必須提供 archive summary preview",
+            "passed": bool(
+                summary["provided"]
+                and summary["mode"] == "candidate_queue_review_archive_summary_preview"
+            ),
+        },
+        {
+            "key": "archive_summary_ready",
+            "label": "archive summary input 必須已通過",
+            "passed": bool(
+                summary["archive_summary_ready"]
+                and summary["summary_input_ready"]
+                and summary["ready_for_ai_summary_review"]
+            ),
+        },
+        {
+            "key": "summary_sections_complete",
+            "label": "摘要輸入必須包含必要 section",
+            "passed": required_sections.issubset(section_keys),
+        },
+        {
+            "key": "summary_dedupe_scope_complete",
+            "label": "summary dedupe key 必須完整且無缺漏",
+            "passed": bool(
+                summary["expected_dedupe_keys"]
+                and summary["expected_dedupe_keys"] == summary["found_dedupe_keys"]
+                and not summary["missing_dedupe_keys"]
+            ),
+        },
+        {
+            "key": "summary_state_clean",
+            "label": "AI 摘要前不可存在 review_state mismatch",
+            "passed": bool(summary["row_summary_count"] > 0 and not summary["state_mismatches"]),
+        },
+        {
+            "key": "summary_artifact_paths_present",
+            "label": "摘要前必須保留上游 artifact path",
+            "passed": all(artifact_paths.values()),
+        },
+        {
+            "key": "ollama_first_policy_locked",
+            "label": "未來摘要路由必須 Ollama-first 且 Gemini 只能備援",
+            "passed": bool(
+                route_policy["primary_policy"] == "ollama_first"
+                and len(route_policy["primary_cascade"]) == 3
+                and route_policy["fallback_policy"]
+                == "gemini_backup_only_after_ollama_cascade_failure"
+                and route_policy["app_host_forbidden_as_ollama_node"]
+                == "192.168.0.188"
+            ),
+        },
+        {
+            "key": "operator_confirmed_ai_summary_preflight",
+            "label": "操作員確認本步不呼叫模型且後續須 Ollama-first",
+            "passed": bool(
+                operator["operator_confirmed_ai_summary_preflight"]
+                and operator["operator_confirmed_ollama_first"]
+                and operator["operator_confirmed_gemini_backup_only"]
+                and operator["operator_confirmed_no_llm_call"]
+                and operator["operator_confirmed_no_telegram_dispatch"]
+                and operator["operator_confirmed_no_api_db_write"]
+                and operator["operator_confirmed_no_scheduler_attach"]
+            ),
+        },
+        {
+            "key": "ai_summary_preflight_artifact_path_recorded",
+            "label": "preflight artifact path 必須由操作員提供",
+            "passed": bool(operator["ai_summary_preflight_artifact_path"]),
+        },
+        {
+            "key": "ai_summary_preflight_no_approval_token_submitted_to_api",
+            "label": "payload 不得包含一次性 approval token key",
+            "passed": not operator["approval_token_submitted_to_api"],
+        },
+        {
+            "key": "ai_summary_preflight_has_no_side_effects",
+            "label": "preflight 不得呼叫 LLM、派送 Telegram、寫檔、寫 DB 或掛 scheduler",
+            "passed": _side_effects_clear(summary),
+        },
+    ]
+
+
+def build_candidate_queue_review_ai_summary_preflight(
+    *,
+    archive_summary,
+    operator_evidence=None,
+    execute_requested=False,
+):
+    """建立 AI summary preflight 預覽；不執行 LLM、檔案、網路或 DB 副作用。"""
+    summary = _summary_payload(archive_summary)
+    operator = _operator_summary(operator_evidence)
+    route_policy = _model_route_policy()
+    gates = _preflight_gates(
+        summary=summary,
+        operator=operator,
+        route_policy=route_policy,
+    )
+    blocked_reasons = [gate["key"] for gate in gates if not gate["passed"]]
+    preflight_ready = bool(not blocked_reasons)
+
+    return {
+        "mode": "candidate_queue_review_ai_summary_preflight_preview",
+        "target_table": TARGET_TABLE,
+        "target_operation": "preflight_ollama_first_ai_summary",
+        "execute_requested": bool(execute_requested),
+        "ai_summary_preflight_ready": preflight_ready,
+        "ready_for_manual_ollama_summary_run": preflight_ready,
+        "ready_for_next_manual_phase": preflight_ready,
+        "ready_for_ai_summary_generation": False,
+        "ready_for_llm_call": False,
+        "ready_for_telegram_dispatch": False,
+        "ready_for_api_review_state_update": False,
+        "ready_for_api_database_write": False,
+        "ready_for_scheduler_attach": False,
+        "api_executes_cli": False,
+        "api_reads_approval_token": False,
+        "api_writes_file": False,
+        "api_writes_database": False,
+        "api_updates_review_state": False,
+        "approval_record_written": False,
+        "decision_record_written": False,
+        "ai_summary_generated": False,
+        "llm_call_executed": False,
+        "ollama_call_executed": False,
+        "gemini_call_executed": False,
+        "telegram_dispatched": False,
+        "summary_file_written": False,
+        "summary_record_written": False,
+        "summary_manifest_written": False,
+        "review_state_update_executed": False,
+        "read_only_query_executed": summary["read_only_query_executed"],
+        "database_connection_opened": summary["database_connection_opened"],
+        "database_session_created": False,
+        "explicit_transaction_opened": False,
+        "transaction_opened": False,
+        "transaction_committed": False,
+        "database_write_executed": False,
+        "database_commit_executed": False,
+        "database_rollback_executed": False,
+        "external_network_executed": False,
+        "scheduler_attached": False,
+        "writes_executed": False,
+        "would_write_database": False,
+        "expected_dedupe_keys": summary["expected_dedupe_keys"],
+        "found_dedupe_keys": summary["found_dedupe_keys"],
+        "missing_dedupe_keys": summary["missing_dedupe_keys"],
+        "state_mismatches": summary["state_mismatches"],
+        "row_summary_count": summary["row_summary_count"],
+        "summary_section_keys": summary["summary_section_keys"],
+        "model_route_policy": route_policy,
+        "operator_ai_summary_preflight": operator,
+        "blocked_reasons": blocked_reasons,
+        "gates": gates,
+        "next_operator_steps": [
+            "確認 archive summary input 已通過且 artifact path 完整",
+            "後續若要產生 AI 摘要，只能另以 OllamaService 三主機級聯執行",
+            "Gemini 只能在 Ollama cascade 全失敗後作為備援或 ADR 鎖定場景",
+            "preflight 本身不得呼叫模型、不得派送 Telegram、不得寫 DB",
+        ],
+        "safe_boundaries": [
+            "do_not_call_llm_from_ai_summary_preflight",
+            "do_not_dispatch_telegram_from_ai_summary_preflight",
+            "do_not_write_summary_file_from_ai_summary_preflight",
+            "do_not_insert_summary_record_from_ai_summary_preflight",
+            "do_not_update_review_state_from_ai_summary_preflight",
+            "do_not_read_approval_token_from_ai_summary_preflight",
+            "do_not_execute_cli_from_ai_summary_preflight",
+            "do_not_attach_scheduler_from_ai_summary_preflight",
+            "ollama_first_required_for_future_summary",
+            "ollama_cascade_must_use_gcp_a_then_gcp_b_then_111",
+            "gemini_backup_only_after_ollama_cascade_failure",
+            "host_188_forbidden_as_ollama_node",
+            "ai_summary_preflight_preview_only",
+            "no_remove_orphans",
+            "no_momo_db_lifecycle_change",
+        ],
+    }