feat(agents): 新增高風險 owner review queue

apps/api/src/services/ai_agent_high_risk_owner_review_queue.py

@@ -0,0 +1,448 @@
+"""
+P2-409 AI Agent high-risk owner review queue snapshot.
+
+Loads the latest committed high-risk owner review queue. This module only
+validates read-only approval packets, rejection guards, and reviewer checklists.
+It does not run workers, send Telegram, write Gateway queues, read secrets, call
+paid APIs, mutate hosts, run kubectl, or write production state.
+"""
+
+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 = "ai_agent_high_risk_owner_review_queue_*.json"
+_SCHEMA_VERSION = "ai_agent_high_risk_owner_review_queue_v1"
+_RUNTIME_AUTHORITY = "high_risk_owner_review_queue_no_live_execution_committed_snapshot"
+_EXPECTED_CURRENT_TASK = "P2-409"
+_EXPECTED_NEXT_TASK = "P2-410"
+_EXPECTED_CANONICAL_ROOM = "AwoooI SRE 戰情室"
+_EXPECTED_CANONICAL_ROOM_ENV = "SRE_GROUP_CHAT_ID"
+_EXPECTED_SOURCE_SCHEMAS = {
+    "ai_agent_low_medium_risk_whitelist_v1",
+    "ai_agent_receipt_readback_owner_review_v1",
+    "ai_agent_report_source_health_v1",
+    "awoooi_work_items_report_source_gap_owner_review_v1",
+    "telegram_notification_egress_inventory_v1",
+    "telegram_notification_egress_owner_request_draft_v1",
+}
+_TRUE_TRUTH_FLAGS = {
+    "p2_408_redirects_loaded",
+    "p2_406b_receipt_owner_review_loaded",
+    "p2_110d_report_source_gap_loaded",
+    "p2_110e_work_items_owner_review_loaded",
+    "telegram_egress_inventory_loaded",
+    "telegram_owner_request_draft_loaded",
+    "all_high_risk_actions_paused",
+    "approval_packets_ready",
+    "rejection_guards_ready",
+    "reviewer_checklists_ready",
+    "high_risk_owner_review_required",
+}
+_FALSE_TRUTH_FLAGS = {
+    "auto_worker_enabled",
+    "live_execution_enabled",
+    "gateway_queue_write_enabled",
+    "telegram_send_enabled",
+    "bot_api_call_enabled",
+    "receipt_production_write_enabled",
+    "production_write_enabled",
+    "secret_read_enabled",
+    "paid_api_call_enabled",
+    "host_write_enabled",
+    "kubectl_action_enabled",
+    "destructive_operation_enabled",
+    "openclaw_replacement_allowed",
+}
+_ZERO_TRUTH_COUNTS = {
+    "auto_worker_run_count_24h",
+    "live_execution_count_24h",
+    "gateway_queue_write_count_24h",
+    "telegram_send_count_24h",
+    "bot_api_call_count_24h",
+    "receipt_production_write_count_24h",
+    "production_write_count_24h",
+    "secret_read_count_24h",
+    "paid_api_call_count_24h",
+    "host_write_count_24h",
+    "kubectl_action_count_24h",
+    "destructive_operation_count_24h",
+    "owner_response_received_count_24h",
+    "owner_response_accepted_count_24h",
+    "redacted_payload_ingested_count_24h",
+}
+_TRUE_BOUNDARY_FLAGS = {
+    "read_only_owner_review_queue_allowed",
+    "approval_packet_preview_allowed",
+    "rejection_guard_preview_allowed",
+    "reviewer_checklist_allowed",
+}
+_FALSE_BOUNDARY_FLAGS = {
+    "auto_worker_enabled",
+    "live_execution_enabled",
+    "gateway_queue_write_enabled",
+    "telegram_send_enabled",
+    "bot_api_call_enabled",
+    "receipt_production_write_enabled",
+    "production_write_enabled",
+    "secret_read_enabled",
+    "paid_api_call_enabled",
+    "host_write_enabled",
+    "kubectl_action_enabled",
+    "destructive_operation_enabled",
+    "openclaw_replacement_allowed",
+}
+_ZERO_ROLLUP_FIELDS = {
+    "owner_response_received_count",
+    "owner_response_accepted_count",
+    "owner_response_rejected_count",
+    "redacted_payload_ingested_count",
+    "auto_worker_run_count",
+    "live_execution_count",
+    "gateway_queue_write_count",
+    "telegram_send_count",
+    "bot_api_call_count",
+    "receipt_production_write_count",
+    "production_write_count",
+    "secret_read_count",
+    "paid_api_call_count",
+    "host_write_count",
+    "kubectl_action_count",
+    "destructive_operation_count",
+}
+_FORBIDDEN_PUBLIC_TERMS = {
+    "批准！繼續",
+    "In app browser",
+    "My request for Codex",
+    "chain_of_thought",
+    "chain-of-thought",
+    "private reasoning text",
+    "authorization_header",
+    "authorization header value",
+    "telegram token value",
+    "raw prompt",
+    "raw_payload",
+}
+
+
+def load_latest_ai_agent_high_risk_owner_review_queue(
+    evaluations_dir: Path | None = None,
+) -> dict[str, Any]:
+    """Load the newest committed P2-409 high-risk owner review queue snapshot."""
+    directory = evaluations_dir or _DEFAULT_EVALUATIONS_DIR
+    candidates = sorted(directory.glob(_SNAPSHOT_PATTERN))
+    if not candidates:
+        raise FileNotFoundError(f"no AI Agent high-risk owner review queue 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")
+
+    label = str(latest)
+    _require_schema(payload, label)
+    _require_sources(payload, label)
+    _require_queue_truth(payload, label)
+    _require_queue_items(payload, label)
+    _require_approval_packets(payload, label)
+    _require_rejection_guards(payload, label)
+    _require_reviewer_checklists(payload, label)
+    _require_routing_policy(payload, label)
+    _require_boundaries(payload, label)
+    _require_rollups(payload, label)
+    _require_no_forbidden_public_terms(payload, label)
+    return payload
+
+
+def _require_schema(payload: dict[str, Any], label: str) -> None:
+    if payload.get("schema_version") != _SCHEMA_VERSION:
+        raise ValueError(f"{label}: expected schema_version={_SCHEMA_VERSION}")
+    status = payload.get("program_status") or {}
+    expected = {
+        "overall_completion_percent": 100,
+        "current_priority": "P0",
+        "current_task_id": _EXPECTED_CURRENT_TASK,
+        "next_task_id": _EXPECTED_NEXT_TASK,
+        "read_only_mode": True,
+        "runtime_authority": _RUNTIME_AUTHORITY,
+    }
+    mismatches = _mismatches(status, expected)
+    if mismatches:
+        raise ValueError(f"{label}: program_status mismatch: {mismatches}")
+    if not status.get("status_note"):
+        raise ValueError(f"{label}: program_status.status_note is required")
+
+
+def _require_sources(payload: dict[str, Any], label: str) -> None:
+    if not payload.get("source_refs"):
+        raise ValueError(f"{label}: source_refs must not be empty")
+    sources = payload.get("source_readbacks") or []
+    schemas = {item.get("source_schema_version") for item in sources}
+    missing = sorted(_EXPECTED_SOURCE_SCHEMAS - schemas)
+    if missing:
+        raise ValueError(f"{label}: missing source schemas: {missing}")
+    for item in sources:
+        readback_id = item.get("readback_id") or "<missing>"
+        for field in ("source_ref", "endpoint", "owner_agent", "status", "key_readback", "next_action"):
+            if not item.get(field):
+                raise ValueError(f"{label}: source readback {readback_id} missing {field}")
+
+
+def _require_queue_truth(payload: dict[str, Any], label: str) -> None:
+    truth = payload.get("queue_truth") or {}
+    missing_true = sorted(flag for flag in _TRUE_TRUTH_FLAGS if truth.get(flag) is not True)
+    if missing_true:
+        raise ValueError(f"{label}: queue truth flags must remain true: {missing_true}")
+    unsafe_false = sorted(flag for flag in _FALSE_TRUTH_FLAGS if truth.get(flag) is not False)
+    if unsafe_false:
+        raise ValueError(f"{label}: queue truth flags must remain false: {unsafe_false}")
+    non_zero = sorted(field for field in _ZERO_TRUTH_COUNTS if truth.get(field) != 0)
+    if non_zero:
+        raise ValueError(f"{label}: queue live counts must remain zero: {non_zero}")
+    if not truth.get("truth_note"):
+        raise ValueError(f"{label}: queue_truth.truth_note is required")
+
+
+def _require_queue_items(payload: dict[str, Any], label: str) -> None:
+    items = payload.get("owner_review_queue_items") or []
+    if len(items) < 1:
+        raise ValueError(f"{label}: owner_review_queue_items must not be empty")
+    risk_tiers = {item.get("risk_tier") for item in items}
+    if not {"high", "critical"}.issubset(risk_tiers):
+        raise ValueError(f"{label}: owner_review_queue_items must include high and critical items")
+    for item in items:
+        item_id = item.get("queue_item_id") or "<missing>"
+        if item.get("risk_tier") not in {"high", "critical"}:
+            raise ValueError(f"{label}: queue item {item_id} must be high or critical")
+        if item.get("queue_status") not in {
+            "paused_owner_review_required",
+            "blocked_missing_owner_response",
+            "approval_packet_preview_ready",
+        }:
+            raise ValueError(f"{label}: queue item {item_id} status is invalid")
+        for field in (
+            "source_readback_ids",
+            "approval_packet_id",
+            "rejection_guard_ids",
+            "reviewer_checklist_ids",
+            "required_owner_fields",
+            "blocked_runtime_actions",
+            "next_gate",
+        ):
+            if not item.get(field):
+                raise ValueError(f"{label}: queue item {item_id} missing {field}")
+        for flag in ("owner_response_required", "rollback_owner_required", "postcheck_required"):
+            if item.get(flag) is not True:
+                raise ValueError(f"{label}: queue item {item_id}.{flag} must remain true")
+        for flag in (
+            "live_execution_allowed",
+            "gateway_queue_write_allowed",
+            "telegram_send_allowed",
+            "production_write_allowed",
+        ):
+            if item.get(flag) is not False:
+                raise ValueError(f"{label}: queue item {item_id}.{flag} must remain false")
+        if item.get("side_effect_count") != 0:
+            raise ValueError(f"{label}: queue item {item_id} side_effect_count must remain zero")
+
+
+def _require_approval_packets(payload: dict[str, Any], label: str) -> None:
+    packets = payload.get("approval_packets") or []
+    packet_ids = {item.get("approval_packet_id") for item in packets}
+    queue_packet_ids = {item.get("approval_packet_id") for item in payload.get("owner_review_queue_items") or []}
+    missing = sorted(queue_packet_ids - packet_ids)
+    if missing:
+        raise ValueError(f"{label}: missing approval packets referenced by queue items: {missing}")
+    queue_ids = {item.get("queue_item_id") for item in payload.get("owner_review_queue_items") or []}
+    for packet in packets:
+        packet_id = packet.get("approval_packet_id") or "<missing>"
+        if packet.get("queue_item_id") not in queue_ids:
+            raise ValueError(f"{label}: approval packet {packet_id} references unknown queue item")
+        if packet.get("packet_status") not in {"draft_ready_owner_response_required", "blocked_missing_owner_response"}:
+            raise ValueError(f"{label}: approval packet {packet_id} status is invalid")
+        for field in ("required_owner_fields", "required_evidence_refs", "reviewer_checklist_id", "rejection_guard_ids"):
+            if not packet.get(field):
+                raise ValueError(f"{label}: approval packet {packet_id} missing {field}")
+        for flag in ("rollback_owner_required", "postcheck_required"):
+            if packet.get(flag) is not True:
+                raise ValueError(f"{label}: approval packet {packet_id}.{flag} must remain true")
+        for flag in (
+            "sensitive_payload_allowed",
+            "live_execution_allowed",
+            "gateway_queue_write_allowed",
+            "telegram_send_allowed",
+            "production_write_allowed",
+        ):
+            if packet.get(flag) is not False:
+                raise ValueError(f"{label}: approval packet {packet_id}.{flag} must remain false")
+
+
+def _require_rejection_guards(payload: dict[str, Any], label: str) -> None:
+    guards = payload.get("rejection_guards") or []
+    guard_ids = {item.get("guard_id") for item in guards}
+    referenced_ids = {
+        guard_id
+        for item in payload.get("owner_review_queue_items") or []
+        for guard_id in (item.get("rejection_guard_ids") or [])
+    } | {
+        guard_id
+        for item in payload.get("approval_packets") or []
+        for guard_id in (item.get("rejection_guard_ids") or [])
+    }
+    missing = sorted(referenced_ids - guard_ids)
+    if missing:
+        raise ValueError(f"{label}: missing rejection guards referenced by packets or queue items: {missing}")
+    for guard in guards:
+        guard_id = guard.get("guard_id") or "<missing>"
+        tiers = set(guard.get("applies_to_risk_tiers") or [])
+        if not tiers or not tiers.issubset({"high", "critical"}):
+            raise ValueError(f"{label}: rejection guard {guard_id} tiers are invalid")
+        for field in ("rejection_condition", "blocked_runtime_actions", "reviewer_action"):
+            if not guard.get(field):
+                raise ValueError(f"{label}: rejection guard {guard_id} missing {field}")
+
+
+def _require_reviewer_checklists(payload: dict[str, Any], label: str) -> None:
+    checklists = payload.get("reviewer_checklists") or []
+    checklist_ids = {item.get("checklist_id") for item in checklists}
+    referenced_ids = {
+        checklist_id
+        for item in payload.get("owner_review_queue_items") or []
+        for checklist_id in (item.get("reviewer_checklist_ids") or [])
+    } | {item.get("reviewer_checklist_id") for item in payload.get("approval_packets") or []}
+    missing = sorted(referenced_ids - checklist_ids)
+    if missing:
+        raise ValueError(f"{label}: missing reviewer checklists referenced by packets or queue items: {missing}")
+    for checklist in checklists:
+        checklist_id = checklist.get("checklist_id") or "<missing>"
+        if not checklist.get("required_checks") or not checklist.get("pass_condition"):
+            raise ValueError(f"{label}: reviewer checklist {checklist_id} missing checks or pass condition")
+        for flag in ("approval_decision_allowed", "checklist_write_allowed"):
+            if checklist.get(flag) is not False:
+                raise ValueError(f"{label}: reviewer checklist {checklist_id}.{flag} must remain false")
+        if checklist.get("side_effect_count") != 0:
+            raise ValueError(f"{label}: reviewer checklist {checklist_id} side_effect_count must remain zero")
+
+
+def _require_routing_policy(payload: dict[str, Any], label: str) -> None:
+    policy = payload.get("routing_policy") or {}
+    expected = {
+        "high_risk_default_route": "pause_to_owner_review_queue",
+        "critical_risk_default_route": "pause_to_owner_review_queue",
+        "low_medium_runtime_route": "pause_until_owner_approved_runtime_gate",
+        "owner_response_required": True,
+        "verbal_approval_accepted": False,
+        "redacted_payload_only": True,
+    }
+    mismatches = _mismatches(policy, expected)
+    if mismatches:
+        raise ValueError(f"{label}: routing_policy mismatch: {mismatches}")
+
+
+def _require_boundaries(payload: dict[str, Any], label: str) -> None:
+    boundaries = payload.get("activation_boundaries") or {}
+    missing_true = sorted(flag for flag in _TRUE_BOUNDARY_FLAGS if boundaries.get(flag) is not True)
+    if missing_true:
+        raise ValueError(f"{label}: activation boundaries must remain true: {missing_true}")
+    unsafe_false = sorted(flag for flag in _FALSE_BOUNDARY_FLAGS if boundaries.get(flag) is not False)
+    if unsafe_false:
+        raise ValueError(f"{label}: activation boundaries must remain false: {unsafe_false}")
+
+    telegram = payload.get("telegram_policy") or {}
+    expected_telegram = {
+        "canonical_room": _EXPECTED_CANONICAL_ROOM,
+        "canonical_room_env": _EXPECTED_CANONICAL_ROOM_ENV,
+        "gateway_queue_write_allowed": False,
+        "direct_bot_api_allowed": False,
+        "telegram_send_allowed": False,
+        "receipt_write_allowed": False,
+    }
+    mismatches = _mismatches(telegram, expected_telegram)
+    if mismatches:
+        raise ValueError(f"{label}: telegram_policy mismatch: {mismatches}")
+
+    redaction = payload.get("display_redaction_contract") or {}
+    if redaction.get("redaction_required") is not True:
+        raise ValueError(f"{label}: display redaction must remain required")
+    for flag in (
+        "unsafe_payload_display_allowed",
+        "private_reasoning_display_allowed",
+        "secret_value_display_allowed",
+        "work_window_transcript_display_allowed",
+    ):
+        if redaction.get(flag) is not False:
+            raise ValueError(f"{label}: display redaction flag {flag} must remain false")
+
+
+def _require_rollups(payload: dict[str, Any], label: str) -> None:
+    rollups = payload.get("rollups") or {}
+    items = payload.get("owner_review_queue_items") or []
+    packets = payload.get("approval_packets") or []
+    guards = payload.get("rejection_guards") or []
+    checklists = payload.get("reviewer_checklists") or []
+    sources = payload.get("source_readbacks") or []
+    blocked_actions = {
+        *(
+            action
+            for item in items
+            for action in (item.get("blocked_runtime_actions") or [])
+        ),
+        *(
+            action
+            for guard in guards
+            for action in (guard.get("blocked_runtime_actions") or [])
+        ),
+    }
+    blocked_actions.discard(None)
+    expected = {
+        "source_readback_count": len(sources),
+        "queue_item_count": len(items),
+        "high_risk_queue_count": sum(1 for item in items if item.get("risk_tier") == "high"),
+        "critical_queue_count": sum(1 for item in items if item.get("risk_tier") == "critical"),
+        "approval_packet_count": len(packets),
+        "rejection_guard_count": len(guards),
+        "reviewer_checklist_count": len(checklists),
+        "approval_packet_required_count": len(items),
+        "rejection_guard_required_queue_count": sum(1 for item in items if item.get("rejection_guard_ids")),
+        "rollback_owner_required_count": sum(1 for item in items if item.get("rollback_owner_required") is True),
+        "postcheck_required_count": sum(1 for item in items if item.get("postcheck_required") is True),
+        "blocked_runtime_action_count": len(blocked_actions),
+    }
+    mismatches = {
+        key: {"expected": value, "actual": rollups.get(key)}
+        for key, value in expected.items()
+        if rollups.get(key) != value
+    }
+    if mismatches:
+        raise ValueError(f"{label}: rollup counts must match payload sections: {mismatches}")
+
+    non_zero = sorted(field for field in _ZERO_ROLLUP_FIELDS if rollups.get(field) != 0)
+    if non_zero:
+        raise ValueError(f"{label}: live rollup counts must remain zero: {non_zero}")
+
+
+def _require_no_forbidden_public_terms(payload: dict[str, Any], label: str) -> None:
+    public_text = json.dumps(payload, ensure_ascii=False)
+    lower_public_text = public_text.lower()
+    leaked_terms = sorted(
+        term
+        for term in _FORBIDDEN_PUBLIC_TERMS
+        if (term.lower() if term.isascii() else term) in lower_public_text
+    )
+    if leaked_terms:
+        raise ValueError(f"{label}: forbidden public terms present: {leaked_terms}")
+
+
+def _mismatches(actual: dict[str, Any], expected: dict[str, Any]) -> dict[str, dict[str, Any]]:
+    return {
+        key: {"expected": expected_value, "actual": actual.get(key)}
+        for key, expected_value in expected.items()
+        if actual.get(key) != expected_value
+    }