feat(governance): 新增服務健康失敗限定通知合約

apps/api/src/api/v1/agents.py

@@ -74,6 +74,9 @@ from src.services.ai_provider_route_matrix import (
 from src.services.service_health_gap_matrix import (
     load_latest_service_health_gap_matrix,
 )
+from src.services.service_health_failure_notification_policy import (
+    load_latest_service_health_failure_notification_policy,
+)
 from src.services.package_supply_chain_inventory import (
     load_latest_package_supply_chain_inventory,
 )
@@ -709,6 +712,34 @@ async def get_backup_notification_policy() -> dict[str, Any]:
         ) from exc
 
 
+@router.get(
+    "/service-health-failure-notification-policy",
+    response_model=dict[str, Any],
+    summary="取得服務健康失敗限定通知合約",
+    description=(
+        "讀取最新已提交的 service health failure-only Telegram / AwoooP 通知合約；"
+        "此端點只回傳成功降噪、action-required 與 failure escalation 規則，"
+        "不送通知、不做 live probe、不重啟服務、不改 endpoint、不觸發 workflow / runtime execution、"
+        "不讀取 secret payload、不回傳工作視窗對話內容或 prompt。"
+    ),
+)
+async def get_service_health_failure_notification_policy() -> dict[str, Any]:
+    """Return the latest read-only service health failure-only notification policy."""
+    try:
+        return await asyncio.to_thread(load_latest_service_health_failure_notification_policy)
+    except FileNotFoundError as exc:
+        raise HTTPException(
+            status_code=status.HTTP_404_NOT_FOUND,
+            detail=str(exc),
+        ) from exc
+    except (json.JSONDecodeError, ValueError) as exc:
+        logger.error("service_health_failure_notification_policy_invalid", error=str(exc))
+        raise HTTPException(
+            status_code=status.HTTP_500_INTERNAL_SERVER_ERROR,
+            detail="服務健康失敗限定通知合約快照無效",
+        ) from exc
+
+
 @router.get(
     "/backup-restore-drill-approval-package-template",
     response_model=dict[str, Any],

apps/api/src/services/service_health_failure_notification_policy.py

@@ -0,0 +1,269 @@
+"""
+Service health failure-only notification policy snapshot.
+
+Loads the latest committed, read-only service health notification policy. The
+policy defines success-noise suppression, failure/action-required escalation,
+message redaction, and frontend display limits. It never sends Telegram or
+AwoooP notifications, writes operator events, probes live systems, restarts
+services, changes endpoints, triggers workflows, reads secrets, or displays
+work-window conversation transcripts.
+"""
+
+from __future__ import annotations
+
+import json
+from pathlib import Path
+from typing import Any
+
+from src.services.snapshot_paths import default_evaluations_dir
+
+_DEFAULT_EVALUATIONS_DIR = default_evaluations_dir(Path(__file__))
+_SNAPSHOT_PATTERN = "service_health_failure_notification_policy_*.json"
+_SCHEMA_VERSION = "service_health_failure_notification_policy_v1"
+
+_CONVERSATION_TRANSCRIPT_MARKERS = {
+    "# In app browser",
+    "My request for Codex",
+    "Current URL:",
+    "AGENTS.md instructions",
+    "<environment_context>",
+    "批准！繼續",
+}
+
+
+def load_latest_service_health_failure_notification_policy(
+    evaluations_dir: Path | None = None,
+) -> dict[str, Any]:
+    """Load the newest committed service health failure notification policy."""
+    directory = evaluations_dir or _DEFAULT_EVALUATIONS_DIR
+    candidates = sorted(directory.glob(_SNAPSHOT_PATTERN))
+    if not candidates:
+        raise FileNotFoundError(
+            f"no service health failure notification policy snapshots found in {directory}"
+        )
+
+    latest = candidates[-1]
+    with latest.open(encoding="utf-8") as handle:
+        payload = json.load(handle)
+
+    if not isinstance(payload, dict):
+        raise ValueError(f"{latest}: expected JSON object")
+    _require_schema(payload, _SCHEMA_VERSION, str(latest))
+    _require_read_only_boundaries(payload, str(latest))
+    _require_operation_boundaries(payload, str(latest))
+    _require_rollup_consistency(payload, str(latest))
+    _require_success_noise_suppression(payload, str(latest))
+    _require_failure_only_escalation(payload, str(latest))
+    _require_frontend_redaction_contract(payload, str(latest))
+    _require_no_plaintext_secret_payload_keys(payload, str(latest))
+    _require_no_conversation_transcript_content(payload, str(latest))
+    return payload
+
+
+def _require_schema(payload: dict[str, Any], expected: str, label: str) -> None:
+    actual = payload.get("schema_version")
+    if actual != expected:
+        raise ValueError(f"{label}: expected schema_version={expected}, got {actual!r}")
+
+
+def _require_read_only_boundaries(payload: dict[str, Any], label: str) -> None:
+    program_status = payload.get("program_status") or {}
+    if program_status.get("read_only_mode") is not True:
+        raise ValueError(f"{label}: program_status.read_only_mode must be true")
+
+    approval_boundaries = payload.get("approval_boundaries") or {}
+    allowed = sorted(key for key, value in approval_boundaries.items() if value is not False)
+    if allowed:
+        raise ValueError(f"{label}: approval boundaries must remain false: {allowed}")
+
+
+def _require_operation_boundaries(payload: dict[str, Any], label: str) -> None:
+    boundaries = payload.get("operation_boundaries") or {}
+    if boundaries.get("read_only_policy_allowed") is not True:
+        raise ValueError(f"{label}: read_only_policy_allowed must be true")
+
+    blocked_flags = {
+        "notification_send_allowed",
+        "telegram_test_message_allowed",
+        "awooop_event_write_allowed",
+        "live_probe_allowed",
+        "external_health_probe_allowed",
+        "service_restart_allowed",
+        "endpoint_change_allowed",
+        "workflow_trigger_allowed",
+        "runtime_execution_allowed",
+        "secret_plaintext_allowed",
+    }
+    allowed = sorted(flag for flag in blocked_flags if boundaries.get(flag) is not False)
+    if allowed:
+        raise ValueError(f"{label}: operation boundaries must remain false: {allowed}")
+
+
+def _require_rollup_consistency(payload: dict[str, Any], label: str) -> None:
+    rules = payload.get("policy_rules") or []
+    rollups = payload.get("rollups") or {}
+    if rollups.get("total_rules") != len(rules):
+        raise ValueError(f"{label}: rollups.total_rules must match policy_rules")
+
+    by_decision: dict[str, int] = {}
+    for rule in rules:
+        decision = str(rule.get("decision"))
+        by_decision[decision] = by_decision.get(decision, 0) + 1
+    if rollups.get("by_decision") != by_decision:
+        raise ValueError(f"{label}: rollups.by_decision must match policy rule decisions")
+
+    immediate_ids = {
+        rule.get("rule_id")
+        for rule in rules
+        if rule.get("decision") == "escalate_immediate"
+    }
+    if set(rollups.get("immediate_escalation_rule_ids") or []) != immediate_ids:
+        raise ValueError(f"{label}: rollups.immediate_escalation_rule_ids must match immediate rules")
+
+    suppressed_success_ids = {
+        rule.get("rule_id")
+        for rule in rules
+        if rule.get("service_state") == "verified"
+        and rule.get("decision") == "suppress_immediate_success"
+    }
+    if set(rollups.get("suppressed_success_rule_ids") or []) != suppressed_success_ids:
+        raise ValueError(f"{label}: rollups.suppressed_success_rule_ids must match suppressed success rules")
+
+    action_required_ids = {
+        rule.get("rule_id")
+        for rule in rules
+        if rule.get("decision") == "create_action_required"
+    }
+    if set(rollups.get("action_required_rule_ids") or []) != action_required_ids:
+        raise ValueError(f"{label}: rollups.action_required_rule_ids must match action-required rules")
+
+    if rollups.get("notification_send_allowed_count") != 0:
+        raise ValueError(f"{label}: rollups.notification_send_allowed_count must remain 0")
+
+
+def _require_success_noise_suppression(payload: dict[str, Any], label: str) -> None:
+    channels = payload.get("notification_channels") or []
+    noisy_channels = [
+        channel.get("channel_id")
+        for channel in channels
+        if channel.get("success_immediate_allowed") is not False
+    ]
+    if noisy_channels:
+        raise ValueError(f"{label}: channels must not allow success immediate notifications: {noisy_channels}")
+
+    success_escalations = [
+        rule.get("rule_id")
+        for rule in payload.get("policy_rules") or []
+        if rule.get("service_state") == "verified"
+        and rule.get("decision") != "suppress_immediate_success"
+    ]
+    if success_escalations:
+        raise ValueError(f"{label}: verified service health rules must suppress immediate notifications")
+
+    template_contract = payload.get("message_template_contract") or {}
+    success_policy = str(template_contract.get("success_message_policy") or "")
+    if "不得即時" not in success_policy or "Telegram / AwoooP" not in success_policy:
+        raise ValueError(f"{label}: success_message_policy must suppress Telegram / AwoooP success noise")
+
+
+def _require_failure_only_escalation(payload: dict[str, Any], label: str) -> None:
+    escalation_rules = [
+        rule
+        for rule in payload.get("policy_rules") or []
+        if rule.get("decision") == "escalate_immediate"
+    ]
+    invalid_states = sorted(
+        rule.get("rule_id")
+        for rule in escalation_rules
+        if rule.get("service_state") not in {"failed", "blocked"}
+    )
+    if invalid_states:
+        raise ValueError(f"{label}: immediate escalation must be failure-only: {invalid_states}")
+
+    telegram_non_failure = sorted(
+        rule.get("rule_id")
+        for rule in payload.get("policy_rules") or []
+        if "telegram_ops" in (rule.get("channels") or [])
+        and rule.get("decision") != "escalate_immediate"
+    )
+    if telegram_non_failure:
+        raise ValueError(f"{label}: telegram_ops must only appear on immediate failure escalation rules")
+
+    template_contract = payload.get("message_template_contract") or {}
+    required_fields = set(template_contract.get("required_fields") or [])
+    required = {"stage", "next_action", "blocked_reason", "target_id", "severity", "evidence_ref"}
+    if not required.issubset(required_fields):
+        raise ValueError(f"{label}: message_template_contract.required_fields missing failure context")
+
+    forbidden_fields = set(template_contract.get("forbidden_fields") or [])
+    required_forbidden_fields = {
+        "secret_value",
+        "token",
+        "authorization_header",
+        "work_window_transcript",
+        "codex_user_message",
+        "prompt_text",
+        "chain_of_thought",
+        "session_id",
+        "browser_context",
+    }
+    if not required_forbidden_fields.issubset(forbidden_fields):
+        raise ValueError(f"{label}: message_template_contract.forbidden_fields missing redaction boundary")
+
+
+def _require_frontend_redaction_contract(payload: dict[str, Any], label: str) -> None:
+    contract = payload.get("display_redaction_contract") or {}
+    if contract.get("conversation_transcript_display_allowed") is not False:
+        raise ValueError(f"{label}: conversation transcript display must remain false")
+    if contract.get("redaction_required") is not True:
+        raise ValueError(f"{label}: frontend redaction must be required")
+
+    forbidden = set(contract.get("forbidden_frontend_content") or [])
+    required_forbidden = {
+        "工作視窗對話內容",
+        "Codex / user 訊息逐字稿",
+        "prompt / chain-of-thought",
+        "session id / browser context",
+        "secret / token / authorization header",
+    }
+    if not required_forbidden.issubset(forbidden):
+        raise ValueError(f"{label}: display_redaction_contract is missing required forbidden content")
+
+    allowed_fields = set(contract.get("allowed_frontend_fields") or [])
+    if "committed evidence ref" not in allowed_fields or "policy rule summary" not in allowed_fields:
+        raise ValueError(f"{label}: display_redaction_contract must limit frontend to committed policy evidence")
+
+
+def _require_no_plaintext_secret_payload_keys(value: Any, label: str, path: str = "$") -> None:
+    if isinstance(value, dict):
+        forbidden_key_fragments = {
+            "secret_value",
+            "token_value",
+            "authorization_header",
+            "private_key",
+            "webhook_secret",
+            "runner_token",
+            "api_key_value",
+            "password_value",
+        }
+        for key, nested in value.items():
+            lowered = str(key).lower()
+            if any(fragment in lowered for fragment in forbidden_key_fragments):
+                raise ValueError(f"{label}: forbidden secret payload key at {path}.{key}")
+            _require_no_plaintext_secret_payload_keys(nested, label, f"{path}.{key}")
+    elif isinstance(value, list):
+        for index, nested in enumerate(value):
+            _require_no_plaintext_secret_payload_keys(nested, label, f"{path}[{index}]")
+
+
+def _require_no_conversation_transcript_content(value: Any, label: str, path: str = "$") -> None:
+    if isinstance(value, dict):
+        for key, nested in value.items():
+            _require_no_conversation_transcript_content(nested, label, f"{path}.{key}")
+    elif isinstance(value, list):
+        for index, nested in enumerate(value):
+            _require_no_conversation_transcript_content(nested, label, f"{path}[{index}]")
+    elif isinstance(value, str):
+        for marker in _CONVERSATION_TRANSCRIPT_MARKERS:
+            if marker in value:
+                raise ValueError(f"{label}: forbidden work-window conversation content at {path}")