feat(governance): 新增 runtime readback 批准包

apps/api/src/services/ai_agent_runtime_readback_approval_package.py

@@ -0,0 +1,390 @@
+"""
+AI Agent runtime readback approval package snapshot.
+
+Loads the latest committed P2-109 runtime readback approval package. This
+module validates repo-committed evidence only; it never reads canonical runtime
+targets, writes result capture rows, writes scores, writes learning state,
+updates PlayBook trust, writes reviewer/Gateway queues, sends Telegram
+messages, calls Bot API, writes rollback work items, writes production targets,
+reads secrets, or runs destructive operations.
+"""
+
+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_runtime_readback_approval_package_*.json"
+_SCHEMA_VERSION = "ai_agent_runtime_readback_approval_package_v1"
+_RUNTIME_AUTHORITY = "runtime_readback_approval_package_only_no_live_write"
+
+
+def load_latest_ai_agent_runtime_readback_approval_package(
+    evaluations_dir: Path | None = None,
+) -> dict[str, Any]:
+    """Load the newest committed runtime readback approval package."""
+    directory = evaluations_dir or _DEFAULT_EVALUATIONS_DIR
+    candidates = sorted(directory.glob(_SNAPSHOT_PATTERN))
+    if not candidates:
+        raise FileNotFoundError(f"no AI Agent runtime readback approval package 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, str(latest))
+    _require_prior_readback(payload, str(latest))
+    _require_approval_truth(payload, str(latest))
+    _require_approval_packets(payload, str(latest))
+    _require_canonical_readback_plans(payload, str(latest))
+    _require_rollback_drill_lanes(payload, str(latest))
+    _require_telegram_failure_receipt_gates(payload, str(latest))
+    _require_operator_actions(payload, str(latest))
+    _require_display_redaction(payload, str(latest))
+    _require_no_forbidden_display_terms(payload, str(latest))
+    _require_rollup_consistency(payload, str(latest))
+    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 {}
+    if status.get("read_only_mode") is not True:
+        raise ValueError(f"{label}: program_status.read_only_mode must be true")
+    if status.get("runtime_authority") != _RUNTIME_AUTHORITY:
+        raise ValueError(f"{label}: runtime_authority must remain {_RUNTIME_AUTHORITY}")
+    if status.get("current_task_id") != "P2-109":
+        raise ValueError(f"{label}: current_task_id must be P2-109")
+    if status.get("next_task_id") != "P2-110":
+        raise ValueError(f"{label}: next_task_id must be P2-110")
+
+
+def _require_prior_readback(payload: dict[str, Any], label: str) -> None:
+    prior = payload.get("prior_readback_readiness") or {}
+    if prior.get("source_schema_version") != "ai_agent_owner_approved_result_capture_readback_v1":
+        raise ValueError(f"{label}: prior_readback_readiness must chain from P2-107")
+    required_counts = {
+        "readback_digest_count": 5,
+        "promotion_review_count": 5,
+        "failure_lane_count": 4,
+        "reviewer_queue_preview_count": 4,
+        "operator_action_count": 5,
+        "approval_required_digest_count": 2,
+        "blocked_total_count": 7,
+        "approved_without_execution_meta_24h": 63,
+        "execution_failed_with_matched_24h": 1,
+        "owner_approval_received_count": 0,
+        "readback_digest_generated_count": 0,
+        "promotion_approved_count": 0,
+        "reviewer_queue_write_count": 0,
+        "result_capture_write_count": 0,
+        "score_write_count": 0,
+        "learning_write_count": 0,
+        "playbook_trust_write_count": 0,
+        "gateway_queue_write_count": 0,
+        "telegram_send_count": 0,
+        "production_write_count": 0,
+    }
+    mismatches = {
+        key: {"expected": expected, "actual": prior.get(key)}
+        for key, expected in required_counts.items()
+        if prior.get(key) != expected
+    }
+    if mismatches:
+        raise ValueError(f"{label}: P2-107 prior readback counts mismatch: {mismatches}")
+
+
+def _require_approval_truth(payload: dict[str, Any], label: str) -> None:
+    truth = payload.get("approval_package_truth") or {}
+    required_true = {
+        "p2_107_readback_loaded",
+        "approval_package_ready",
+        "canonical_readback_plan_ready",
+        "rollback_drill_ready",
+        "telegram_failure_receipt_gate_ready",
+        "owner_review_required_before_runtime",
+    }
+    missing = sorted(field for field in required_true if truth.get(field) is not True)
+    if missing:
+        raise ValueError(f"{label}: approval readiness flags must remain true: {missing}")
+
+    required_false = {
+        "canonical_runtime_readback_enabled",
+        "runtime_result_capture_write_enabled",
+        "runtime_score_write_enabled",
+        "runtime_learning_write_enabled",
+        "playbook_trust_write_enabled",
+        "reviewer_queue_write_enabled",
+        "gateway_queue_write_enabled",
+        "telegram_failure_receipt_send_enabled",
+        "bot_api_call_enabled",
+        "rollback_work_item_write_enabled",
+        "production_write_enabled",
+        "secret_read_enabled",
+        "destructive_operation_enabled",
+    }
+    unsafe = sorted(field for field in required_false if truth.get(field) is not False)
+    if unsafe:
+        raise ValueError(f"{label}: runtime read/write/send flags must remain false: {unsafe}")
+
+    zero_counts = {
+        "owner_approval_received_count",
+        "runtime_readback_execution_count",
+        "result_capture_write_count_24h",
+        "score_write_count_24h",
+        "learning_write_count_24h",
+        "playbook_trust_write_count_24h",
+        "reviewer_queue_write_count_24h",
+        "gateway_queue_write_count_24h",
+        "telegram_failure_receipt_send_count_24h",
+        "bot_api_call_count_24h",
+        "rollback_work_item_write_count_24h",
+        "production_write_count_24h",
+        "secret_read_count_24h",
+        "destructive_operation_count_24h",
+    }
+    non_zero = sorted(field for field in zero_counts if truth.get(field) != 0)
+    if non_zero:
+        raise ValueError(f"{label}: approval package live counters must remain zero: {non_zero}")
+
+
+def _require_approval_packets(payload: dict[str, Any], label: str) -> None:
+    packets = payload.get("approval_packet_templates") or []
+    packet_ids = {packet.get("packet_id") for packet in packets}
+    required = {
+        "approved_result_capture_runtime_readback",
+        "execution_failed_negative_learning_review",
+        "pending_human_gate_expiry_review",
+        "manual_or_noop_result_acceptance",
+        "post_write_verifier_receipt_package",
+    }
+    if packet_ids != required:
+        raise ValueError(f"{label}: approval packet templates must match {sorted(required)}")
+
+    valid_statuses = {"ready_for_owner_review", "approval_required", "blocked_by_policy"}
+    valid_tiers = {"medium", "high", "critical"}
+    for packet in packets:
+        packet_id = packet.get("packet_id")
+        if packet.get("status") not in valid_statuses:
+            raise ValueError(f"{label}: packet {packet_id} status is invalid")
+        if packet.get("risk_tier") not in valid_tiers:
+            raise ValueError(f"{label}: packet {packet_id} risk_tier is invalid")
+        if packet.get("approval_required") is not True:
+            raise ValueError(f"{label}: packet {packet_id} approval_required must remain true")
+        if packet.get("runtime_write_allowed") is not False:
+            raise ValueError(f"{label}: packet {packet_id} runtime_write_allowed must remain false")
+        if not packet.get("required_approval_fields") or not packet.get("blocked_runtime_actions"):
+            raise ValueError(f"{label}: packet {packet_id} must list approval fields and blocked actions")
+        if not _is_redacted_sha256(packet.get("evidence_hash")):
+            raise ValueError(f"{label}: packet {packet_id} must expose evidence_hash")
+
+
+def _require_canonical_readback_plans(payload: dict[str, Any], label: str) -> None:
+    plans = payload.get("canonical_readback_plans") or []
+    plan_ids = {plan.get("plan_id") for plan in plans}
+    required = {
+        "result_capture_row_readback",
+        "score_fixture_parity_readback",
+        "playbook_trust_noop_readback",
+        "verifier_receipt_readback",
+    }
+    if plan_ids != required:
+        raise ValueError(f"{label}: canonical readback plans must match {sorted(required)}")
+    for plan in plans:
+        plan_id = plan.get("plan_id")
+        if plan.get("status") not in {"planned_no_runtime_read", "blocked_by_policy"}:
+            raise ValueError(f"{label}: canonical readback plan {plan_id} status is invalid")
+        if plan.get("live_query_enabled") is not False:
+            raise ValueError(f"{label}: canonical readback plan {plan_id} live_query_enabled must remain false")
+        if plan.get("runtime_write_allowed") is not False:
+            raise ValueError(f"{label}: canonical readback plan {plan_id} runtime_write_allowed must remain false")
+        if not plan.get("required_fields") or not plan.get("redaction_checks"):
+            raise ValueError(f"{label}: canonical readback plan {plan_id} must list required fields and redaction checks")
+        if not _is_redacted_sha256(plan.get("evidence_hash")):
+            raise ValueError(f"{label}: canonical readback plan {plan_id} must expose evidence_hash")
+
+
+def _require_rollback_drill_lanes(payload: dict[str, Any], label: str) -> None:
+    lanes = payload.get("rollback_drill_lanes") or []
+    lane_ids = {lane.get("lane_id") for lane in lanes}
+    required = {
+        "result_capture_insert_rollback_drill",
+        "score_write_rollback_drill",
+        "learning_write_rollback_drill",
+        "telegram_receipt_rollback_drill",
+    }
+    if lane_ids != required:
+        raise ValueError(f"{label}: rollback drill lanes must match {sorted(required)}")
+    for lane in lanes:
+        lane_id = lane.get("lane_id")
+        if lane.get("status") not in {"planned_no_write", "blocked_by_policy"}:
+            raise ValueError(f"{label}: rollback drill lane {lane_id} status is invalid")
+        if lane.get("rollback_write_enabled") is not False:
+            raise ValueError(f"{label}: rollback drill lane {lane_id} rollback_write_enabled must remain false")
+        if lane.get("creates_runtime_write") is not False:
+            raise ValueError(f"{label}: rollback drill lane {lane_id} creates_runtime_write must remain false")
+        if not lane.get("rollback_owner") or not lane.get("required_steps"):
+            raise ValueError(f"{label}: rollback drill lane {lane_id} must list owner and required steps")
+
+
+def _require_telegram_failure_receipt_gates(payload: dict[str, Any], label: str) -> None:
+    gates = payload.get("telegram_failure_receipt_gates") or []
+    gate_ids = {gate.get("gate_id") for gate in gates}
+    required = {
+        "receipt_payload_redaction",
+        "sre_war_room_route_lock",
+        "delivery_correlation",
+        "duplicate_suppression",
+    }
+    if gate_ids != required:
+        raise ValueError(f"{label}: Telegram failure receipt gates must match {sorted(required)}")
+    for gate in gates:
+        gate_id = gate.get("gate_id")
+        if gate.get("status") not in {"ready_for_owner_review", "blocked_by_policy"}:
+            raise ValueError(f"{label}: Telegram failure gate {gate_id} status is invalid")
+        if gate.get("queue_write_enabled") is not False:
+            raise ValueError(f"{label}: Telegram failure gate {gate_id} queue_write_enabled must remain false")
+        if gate.get("telegram_send_enabled") is not False:
+            raise ValueError(f"{label}: Telegram failure gate {gate_id} telegram_send_enabled must remain false")
+        if gate.get("bot_api_call_enabled") is not False:
+            raise ValueError(f"{label}: Telegram failure gate {gate_id} bot_api_call_enabled must remain false")
+        if not gate.get("required_evidence"):
+            raise ValueError(f"{label}: Telegram failure gate {gate_id} must list required evidence")
+        if not _is_redacted_sha256(gate.get("evidence_hash")):
+            raise ValueError(f"{label}: Telegram failure gate {gate_id} must expose evidence_hash")
+
+
+def _require_operator_actions(payload: dict[str, Any], label: str) -> None:
+    actions = payload.get("operator_actions") or []
+    action_types = {action.get("action_type") for action in actions}
+    required = {
+        "review_approval_package",
+        "validate_canonical_plan",
+        "review_rollback_drill",
+        "review_failure_receipt_gate",
+        "reject_or_promote",
+    }
+    if action_types != required:
+        raise ValueError(f"{label}: operator actions must match {sorted(required)}")
+    for action in actions:
+        if action.get("runtime_write_allowed") is not False:
+            raise ValueError(f"{label}: operator action {action.get('action_id')} must not allow runtime write")
+        if not action.get("operator_instruction"):
+            raise ValueError(f"{label}: operator action {action.get('action_id')} must include instruction")
+
+
+def _require_display_redaction(payload: dict[str, Any], label: str) -> None:
+    contract = payload.get("display_redaction_contract") or {}
+    if contract.get("redaction_required") is not True:
+        raise ValueError(f"{label}: display redaction must remain required")
+    required_false = {
+        "raw_prompt_display_allowed",
+        "private_reasoning_display_allowed",
+        "secret_value_display_allowed",
+        "raw_telegram_payload_display_allowed",
+        "work_window_transcript_display_allowed",
+    }
+    unsafe = sorted(field for field in required_false if contract.get(field) is not False)
+    if unsafe:
+        raise ValueError(f"{label}: display redaction fields must remain false: {unsafe}")
+
+
+def _require_no_forbidden_display_terms(payload: dict[str, Any], label: str) -> None:
+    forbidden_terms = {
+        "工作視窗",
+        "對話內容",
+        "批准！繼續",
+        "In app browser",
+        "My request for Codex",
+        "browser_context",
+        "codex_user_message",
+        "prompt_text",
+        "raw prompt",
+        "private reasoning",
+        "chain of thought",
+        "private_reasoning",
+        "chain_of_thought",
+        "authorization_header",
+        "work window transcript",
+        "internal collaboration transcript",
+    }
+    hits: list[str] = []
+
+    def walk(value: Any, path: str) -> None:
+        if isinstance(value, dict):
+            for key, nested in value.items():
+                walk(nested, f"{path}.{key}" if path else str(key))
+            return
+        if isinstance(value, list):
+            for index, nested in enumerate(value):
+                walk(nested, f"{path}[{index}]")
+            return
+        if isinstance(value, str):
+            matched = sorted(term for term in forbidden_terms if term in value)
+            if matched:
+                hits.append(f"{path}: {', '.join(matched)}")
+
+    walk(payload, "")
+    if hits:
+        raise ValueError(f"{label}: forbidden display terms found: {hits}")
+
+
+def _require_rollup_consistency(payload: dict[str, Any], label: str) -> None:
+    rollups = payload.get("rollups") or {}
+    truth = payload.get("approval_package_truth") or {}
+    prior = payload.get("prior_readback_readiness") or {}
+    packets = payload.get("approval_packet_templates") or []
+    plans = payload.get("canonical_readback_plans") or []
+    rollback_lanes = payload.get("rollback_drill_lanes") or []
+    telegram_gates = payload.get("telegram_failure_receipt_gates") or []
+    actions = payload.get("operator_actions") or []
+    expected = {
+        "approval_packet_count": len(packets),
+        "canonical_readback_plan_count": len(plans),
+        "rollback_drill_lane_count": len(rollback_lanes),
+        "telegram_failure_receipt_gate_count": len(telegram_gates),
+        "operator_action_count": len(actions),
+        "approval_required_packet_count": sum(1 for item in packets if item.get("status") == "approval_required"),
+        "blocked_packet_count": sum(1 for item in packets if item.get("status") == "blocked_by_policy"),
+        "blocked_readback_plan_count": sum(1 for item in plans if item.get("status") == "blocked_by_policy"),
+        "blocked_rollback_lane_count": sum(1 for item in rollback_lanes if item.get("status") == "blocked_by_policy"),
+        "blocked_telegram_gate_count": sum(1 for item in telegram_gates if item.get("status") == "blocked_by_policy"),
+        "approved_without_execution_meta_24h": prior.get("approved_without_execution_meta_24h"),
+        "execution_failed_with_matched_24h": prior.get("execution_failed_with_matched_24h"),
+        "owner_approval_received_count": truth.get("owner_approval_received_count"),
+        "runtime_readback_execution_count": truth.get("runtime_readback_execution_count"),
+        "result_capture_write_count": truth.get("result_capture_write_count_24h"),
+        "score_write_count": truth.get("score_write_count_24h"),
+        "learning_write_count": truth.get("learning_write_count_24h"),
+        "playbook_trust_write_count": truth.get("playbook_trust_write_count_24h"),
+        "reviewer_queue_write_count": truth.get("reviewer_queue_write_count_24h"),
+        "gateway_queue_write_count": truth.get("gateway_queue_write_count_24h"),
+        "telegram_failure_receipt_send_count": truth.get("telegram_failure_receipt_send_count_24h"),
+        "bot_api_call_count": truth.get("bot_api_call_count_24h"),
+        "rollback_work_item_write_count": truth.get("rollback_work_item_write_count_24h"),
+        "production_write_count": truth.get("production_write_count_24h"),
+        "secret_read_count": truth.get("secret_read_count_24h"),
+        "destructive_operation_count": truth.get("destructive_operation_count_24h"),
+    }
+    mismatches = {
+        key: {"expected": expected_value, "actual": rollups.get(key)}
+        for key, expected_value in expected.items()
+        if rollups.get(key) != expected_value
+    }
+    if mismatches:
+        raise ValueError(f"{label}: rollup counts mismatch: {mismatches}")
+
+
+def _is_redacted_sha256(value: Any) -> bool:
+    if not isinstance(value, str):
+        return False
+    if not value.startswith("sha256:") or len(value) != 71:
+        return False
+    return all(char in "0123456789abcdef" for char in value.removeprefix("sha256:"))