feat(governance): 新增 runtime readback implementation review

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

@@ -106,6 +106,9 @@ from src.services.ai_agent_redis_dry_run_gate import (
 from src.services.ai_agent_runtime_readback_approval_package import (
     load_latest_ai_agent_runtime_readback_approval_package,
 )
+from src.services.ai_agent_runtime_readback_implementation_review import (
+    load_latest_ai_agent_runtime_readback_implementation_review,
+)
 from src.services.ai_agent_report_automation_review import (
     load_latest_ai_agent_report_automation_review,
 )
@@ -1300,7 +1303,7 @@ async def get_agent_owner_approved_result_capture_readback() -> dict[str, Any]:
     response_model=dict[str, Any],
     summary="取得 AI Agent runtime readback 批准包",
     description=(
-        "讀取最新已提交的 P2-108 runtime readback approval package；"
+        "讀取最新已提交的 P2-109 runtime readback approval package；"
         "此端點只回傳 approval packet、canonical readback plan、rollback drill、"
         "Telegram 失敗收據 gate 與 operator action，"
         "不讀 canonical runtime target、不寫 score、不寫 result capture、不寫 learning、不更新 PlayBook trust、"
@@ -1326,6 +1329,36 @@ async def get_agent_runtime_readback_approval_package() -> dict[str, Any]:
         ) from exc
 
 
+@router.get(
+    "/agent-runtime-readback-implementation-review",
+    response_model=dict[str, Any],
+    summary="取得 AI Agent runtime readback 實作審查",
+    description=(
+        "讀取最新已提交的 P2-110 runtime readback implementation review；"
+        "此端點只回傳 implementation review card、no-write verifier、阻塞原因與 operator action，"
+        "不讀 canonical runtime target、不做 live query、不寫 score、不寫 result capture、不寫 learning、"
+        "不更新 PlayBook trust、不寫 reviewer queue、不寫 Gateway queue、不送 Telegram、不呼叫 Bot API、"
+        "不寫 rollback work item、不寫 production target、不讀 secret。"
+    ),
+)
+async def get_agent_runtime_readback_implementation_review() -> dict[str, Any]:
+    """Return the latest read-only runtime readback implementation review."""
+    try:
+        payload = await asyncio.to_thread(load_latest_ai_agent_runtime_readback_implementation_review)
+        return redact_public_lan_topology(payload)
+    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("ai_agent_runtime_readback_implementation_review_invalid", error=str(exc))
+        raise HTTPException(
+            status_code=status.HTTP_500_INTERNAL_SERVER_ERROR,
+            detail="AI Agent runtime readback implementation review 無效",
+        ) from exc
+
+
 @router.get(
     "/agent-owner-approved-fixture-dry-run",
     response_model=dict[str, Any],

apps/api/src/services/ai_agent_runtime_readback_implementation_review.py

@@ -0,0 +1,364 @@
+"""
+AI Agent runtime readback implementation review snapshot.
+
+Loads the latest committed P2-110 runtime readback implementation review.
+This module validates 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_implementation_review_*.json"
+_SCHEMA_VERSION = "ai_agent_runtime_readback_implementation_review_v1"
+_RUNTIME_AUTHORITY = "runtime_readback_implementation_review_only_no_live_read_or_write"
+
+
+def load_latest_ai_agent_runtime_readback_implementation_review(
+    evaluations_dir: Path | None = None,
+) -> dict[str, Any]:
+    """Load the newest committed runtime readback implementation review."""
+    directory = evaluations_dir or _DEFAULT_EVALUATIONS_DIR
+    candidates = sorted(directory.glob(_SNAPSHOT_PATTERN))
+    if not candidates:
+        raise FileNotFoundError(f"no AI Agent runtime readback implementation review 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_package(payload, str(latest))
+    _require_review_truth(payload, str(latest))
+    _require_review_cards(payload, str(latest))
+    _require_verifier_checks(payload, str(latest))
+    _require_blockers(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") != "P2-110":
+        raise ValueError(f"{label}: current_task_id must be P2-110")
+    if status.get("next_task_id") != "P2-111":
+        raise ValueError(f"{label}: next_task_id must be P2-111")
+
+
+def _require_prior_package(payload: dict[str, Any], label: str) -> None:
+    prior = payload.get("prior_approval_package") or {}
+    if prior.get("source_schema_version") != "ai_agent_runtime_readback_approval_package_v1":
+        raise ValueError(f"{label}: prior_approval_package must chain from P2-109")
+    required_counts = {
+        "approval_packet_count": 5,
+        "canonical_readback_plan_count": 4,
+        "rollback_drill_lane_count": 4,
+        "telegram_failure_receipt_gate_count": 4,
+        "operator_action_count": 5,
+        "approval_required_packet_count": 3,
+        "blocked_total_count": 5,
+        "owner_approval_received_count": 0,
+        "runtime_readback_execution_count": 0,
+        "telegram_failure_receipt_send_count": 0,
+        "bot_api_call_count": 0,
+        "rollback_work_item_write_count": 0,
+        "production_write_count": 0,
+        "secret_read_count": 0,
+        "destructive_operation_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}: P2-109 prior approval package counts mismatch: {mismatches}")
+
+
+def _require_review_truth(payload: dict[str, Any], label: str) -> None:
+    truth = payload.get("implementation_review_truth") or {}
+    required_true = {
+        "p2_109_approval_package_loaded",
+        "implementation_review_ready",
+        "adapter_contract_ready",
+        "verifier_contract_ready",
+        "redaction_contract_ready",
+        "telemetry_receipt_contract_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}: implementation review flags must remain true: {missing}")
+
+    required_false = {
+        "canonical_runtime_readback_enabled",
+        "live_query_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",
+        "live_query_count_24h",
+        "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}: implementation review live counters must remain zero: {non_zero}")
+
+
+def _require_review_cards(payload: dict[str, Any], label: str) -> None:
+    cards = payload.get("implementation_review_cards") or []
+    card_ids = {card.get("card_id") for card in cards}
+    required = {
+        "runtime_readback_adapter_contract",
+        "result_capture_writer_noop_contract",
+        "critic_reviewer_score_noop_contract",
+        "telegram_failure_receipt_adapter_contract",
+        "rollback_work_item_noop_contract",
+    }
+    if card_ids != required:
+        raise ValueError(f"{label}: implementation review cards must match {sorted(required)}")
+
+    valid_statuses = {"ready_for_owner_review", "approval_required", "blocked_by_policy"}
+    valid_tiers = {"high", "critical"}
+    for card in cards:
+        card_id = card.get("card_id")
+        if card.get("status") not in valid_statuses:
+            raise ValueError(f"{label}: card {card_id} status is invalid")
+        if card.get("risk_tier") not in valid_tiers:
+            raise ValueError(f"{label}: card {card_id} risk_tier is invalid")
+        if card.get("owner_review_required") is not True:
+            raise ValueError(f"{label}: card {card_id} owner_review_required must remain true")
+        if card.get("no_write_mode") is not True:
+            raise ValueError(f"{label}: card {card_id} no_write_mode must remain true")
+        if not card.get("required_interfaces") or not card.get("blocked_runtime_actions"):
+            raise ValueError(f"{label}: card {card_id} must list interfaces and blocked runtime actions")
+        if not _is_redacted_sha256(card.get("evidence_hash")):
+            raise ValueError(f"{label}: card {card_id} must expose evidence_hash")
+
+
+def _require_verifier_checks(payload: dict[str, Any], label: str) -> None:
+    checks = payload.get("no_write_verifier_checks") or []
+    check_ids = {check.get("check_id") for check in checks}
+    required = {
+        "schema_field_parity_check",
+        "redaction_payload_check",
+        "zero_write_counter_check",
+        "failure_receipt_route_check",
+        "rollback_reference_check",
+    }
+    if check_ids != required:
+        raise ValueError(f"{label}: no-write verifier checks must match {sorted(required)}")
+    for check in checks:
+        check_id = check.get("check_id")
+        if check.get("status") not in {"ready", "blocked_by_policy"}:
+            raise ValueError(f"{label}: verifier check {check_id} status is invalid")
+        if check.get("live_query_enabled") is not False:
+            raise ValueError(f"{label}: verifier check {check_id} live_query_enabled must remain false")
+        if check.get("runtime_write_allowed") is not False:
+            raise ValueError(f"{label}: verifier check {check_id} runtime_write_allowed must remain false")
+        if not check.get("required_evidence"):
+            raise ValueError(f"{label}: verifier check {check_id} must list required evidence")
+        if not _is_redacted_sha256(check.get("evidence_hash")):
+            raise ValueError(f"{label}: verifier check {check_id} must expose evidence_hash")
+
+
+def _require_blockers(payload: dict[str, Any], label: str) -> None:
+    blockers = payload.get("implementation_blockers") or []
+    blocker_ids = {blocker.get("blocker_id") for blocker in blockers}
+    required = {
+        "owner_acceptance_missing",
+        "canonical_runtime_readback_disabled",
+        "reviewer_queue_write_disabled",
+        "telegram_receipt_send_disabled",
+        "rollback_work_item_write_disabled",
+    }
+    if blocker_ids != required:
+        raise ValueError(f"{label}: implementation blockers must match {sorted(required)}")
+    valid_statuses = {
+        "blocking_runtime",
+        "blocking_live_read",
+        "blocking_queue_write",
+        "blocking_notification",
+        "blocking_production_write",
+    }
+    for blocker in blockers:
+        blocker_id = blocker.get("blocker_id")
+        if blocker.get("severity") not in {"high", "critical"}:
+            raise ValueError(f"{label}: blocker {blocker_id} severity is invalid")
+        if blocker.get("status") not in valid_statuses:
+            raise ValueError(f"{label}: blocker {blocker_id} status is invalid")
+        if not blocker.get("blocked_action") or not blocker.get("required_resolution"):
+            raise ValueError(f"{label}: blocker {blocker_id} must explain blocked action and resolution")
+        if not _is_redacted_sha256(blocker.get("evidence_hash")):
+            raise ValueError(f"{label}: blocker {blocker_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_implementation_contract",
+        "validate_adapter_mapping",
+        "validate_zero_write_counters",
+        "review_failure_receipt_mapping",
+        "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("implementation_review_truth") or {}
+    cards = payload.get("implementation_review_cards") or []
+    checks = payload.get("no_write_verifier_checks") or []
+    blockers = payload.get("implementation_blockers") or []
+    actions = payload.get("operator_actions") or []
+    expected = {
+        "implementation_review_card_count": len(cards),
+        "no_write_verifier_check_count": len(checks),
+        "implementation_blocker_count": len(blockers),
+        "operator_action_count": len(actions),
+        "approval_required_card_count": sum(1 for item in cards if item.get("status") == "approval_required"),
+        "blocked_card_count": sum(1 for item in cards if item.get("status") == "blocked_by_policy"),
+        "blocked_verifier_check_count": sum(1 for item in checks if item.get("status") == "blocked_by_policy"),
+        "critical_blocker_count": sum(1 for item in blockers if item.get("severity") == "critical"),
+        "owner_approval_received_count": truth.get("owner_approval_received_count"),
+        "runtime_readback_execution_count": truth.get("runtime_readback_execution_count"),
+        "live_query_count": truth.get("live_query_count_24h"),
+        "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:"))