feat(governance): 新增 result capture write gate review

apps/api/src/services/ai_agent_result_capture_write_gate_review.py

@@ -0,0 +1,381 @@
+"""
+AI Agent result capture write gate review snapshot.
+
+Loads the latest committed P2-121 result capture write gate review package.
+This module validates committed evidence only; it never writes result captures,
+writes learning records, updates PlayBook trust, writes reviewer / Gateway
+queues, sends Telegram messages, reads canonical runtime targets, reads secrets,
+or performs 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_result_capture_write_gate_review_*.json"
+_SCHEMA_VERSION = "ai_agent_result_capture_write_gate_review_v1"
+_RUNTIME_AUTHORITY = "result_capture_write_gate_review_only_no_live_write"
+_TARGET_WRITE_GATE = "result_capture_write_gate_review"
+
+
+def load_latest_ai_agent_result_capture_write_gate_review(
+    evaluations_dir: Path | None = None,
+) -> dict[str, Any]:
+    """Load the newest committed result capture write gate review package."""
+    directory = evaluations_dir or _DEFAULT_EVALUATIONS_DIR
+    candidates = sorted(directory.glob(_SNAPSHOT_PATTERN))
+    if not candidates:
+        raise FileNotFoundError(f"no AI Agent result capture write gate 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")
+
+    label = str(latest)
+    _require_schema(payload, label)
+    _require_prior(payload, label)
+    _require_truth(payload, label)
+    _require_write_gate_reviews(payload, label)
+    _require_approval_gates(payload, label)
+    _require_verifier_plan(payload, label)
+    _require_blocked_writes(payload, label)
+    _require_actions(payload, label)
+    _require_display_redaction(payload, label)
+    _require_no_forbidden_display_terms(payload, label)
+    _require_rollup_consistency(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 = {
+        "current_priority": "P2",
+        "current_task_id": "P2-121",
+        "next_task_id": "P2-122",
+        "read_only_mode": True,
+        "runtime_authority": _RUNTIME_AUTHORITY,
+        "overall_completion_percent": 100,
+    }
+    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_prior(payload: dict[str, Any], label: str) -> None:
+    prior = payload.get("prior_owner_approved_promotion_dry_run") or {}
+    expected = {
+        "schema_version": "ai_agent_owner_approved_result_capture_promotion_dry_run_v1",
+        "promotion_dry_run_template_count": 5,
+        "owner_acceptance_fixture_count": 5,
+        "dry_run_verifier_check_count": 5,
+        "blocked_runtime_promotion_count": 5,
+        "operator_action_count": 5,
+        "owner_approval_received_count": 0,
+        "capture_promotion_approved_count": 0,
+        "dry_run_preview_generated_count": 0,
+        "result_capture_write_count": 0,
+        "learning_write_count": 0,
+        "playbook_trust_write_count": 0,
+        "reviewer_queue_write_count": 0,
+        "gateway_queue_write_count": 0,
+        "telegram_send_count": 0,
+        "bot_api_call_count": 0,
+        "report_receipt_write_count": 0,
+    }
+    mismatches = _mismatches(prior, expected)
+    if mismatches:
+        raise ValueError(f"{label}: prior_owner_approved_promotion_dry_run mismatch: {mismatches}")
+    if not prior.get("readiness_note"):
+        raise ValueError(f"{label}: prior_owner_approved_promotion_dry_run.readiness_note is required")
+
+
+def _require_truth(payload: dict[str, Any], label: str) -> None:
+    truth = payload.get("write_gate_truth") or {}
+    required_true = {
+        "p2_120_dry_run_loaded",
+        "result_capture_writer_gate_review_ready",
+        "learning_writer_gate_review_ready",
+        "playbook_trust_writer_gate_review_ready",
+        "reviewer_queue_gate_review_ready",
+        "gateway_queue_gate_review_ready",
+        "dual_approval_required",
+        "dry_run_hash_required",
+        "post_write_verifier_required",
+        "rollback_required",
+        "redaction_required",
+    }
+    missing = sorted(field for field in required_true if truth.get(field) is not True)
+    if missing:
+        raise ValueError(f"{label}: write-gate ready flags must remain true: {missing}")
+    required_false = {
+        "result_capture_write_approved",
+        "learning_write_approved",
+        "playbook_trust_write_approved",
+        "reviewer_queue_write_approved",
+        "gateway_queue_write_approved",
+        "canonical_runtime_target_read_enabled",
+        "live_query_enabled",
+        "reviewer_queue_write_enabled",
+        "gateway_queue_write_enabled",
+        "telegram_send_enabled",
+        "bot_api_call_enabled",
+        "report_receipt_write_enabled",
+        "result_capture_write_enabled",
+        "learning_write_enabled",
+        "playbook_trust_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}: live read/send/write flags must remain false: {unsafe}")
+    zero_counts = {
+        "owner_approval_received_count",
+        "dual_approval_received_count",
+        "dry_run_hash_verified_count",
+        "post_write_verifier_pass_count",
+        "rollback_plan_verified_count",
+        "canonical_runtime_target_read_count",
+        "live_query_count",
+        "reviewer_queue_write_count",
+        "gateway_queue_write_count",
+        "telegram_send_count",
+        "bot_api_call_count",
+        "report_receipt_write_count",
+        "result_capture_write_count",
+        "learning_write_count",
+        "playbook_trust_write_count",
+        "production_write_count",
+        "secret_read_count",
+        "destructive_operation_count",
+    }
+    non_zero = sorted(field for field in zero_counts if truth.get(field) != 0)
+    if non_zero:
+        raise ValueError(f"{label}: write-gate live counters must remain zero: {non_zero}")
+    if not truth.get("truth_note"):
+        raise ValueError(f"{label}: write_gate_truth.truth_note is required")
+
+
+def _require_write_gate_reviews(payload: dict[str, Any], label: str) -> None:
+    reviews = payload.get("write_gate_reviews") or []
+    required = {
+        "gate_result_capture_writer",
+        "gate_learning_writer",
+        "gate_playbook_trust_writer",
+        "gate_reviewer_queue_writer",
+        "gate_gateway_queue_writer",
+    }
+    review_ids = {review.get("gate_id") for review in reviews}
+    if review_ids != required:
+        raise ValueError(f"{label}: write gate reviews must match {sorted(required)}")
+    for review in reviews:
+        gate_id = review.get("gate_id")
+        if review.get("target_write_gate") != _TARGET_WRITE_GATE:
+            raise ValueError(f"{label}: review {gate_id} must target {_TARGET_WRITE_GATE}")
+        if review.get("runtime_write_enabled") is not False:
+            raise ValueError(f"{label}: review {gate_id} must not enable runtime write")
+        for field in {"dual_approval_required", "dry_run_hash_required", "post_write_verifier_required"}:
+            if review.get(field) is not True:
+                raise ValueError(f"{label}: review {gate_id} must require {field}")
+        if review.get("status") not in {"ready_for_owner_review", "approval_required", "blocked_by_policy"}:
+            raise ValueError(f"{label}: review {gate_id} status is invalid")
+        if not review.get("required_before_write") or not review.get("blocked_write_action"):
+            raise ValueError(f"{label}: review {gate_id} must include required_before_write and blocked_write_action")
+        if not _is_redacted_sha256(review.get("evidence_hash")):
+            raise ValueError(f"{label}: review {gate_id} must expose redacted evidence_hash")
+
+
+def _require_approval_gates(payload: dict[str, Any], label: str) -> None:
+    gates = payload.get("approval_gates") or []
+    required = {
+        "gate_owner_dual_approval",
+        "gate_dry_run_hash",
+        "gate_post_write_verifier",
+        "gate_rollback_reverify",
+        "gate_redaction_policy",
+    }
+    gate_ids = {gate.get("gate_id") for gate in gates}
+    if gate_ids != required:
+        raise ValueError(f"{label}: approval gates must match {sorted(required)}")
+    for gate in gates:
+        gate_id = gate.get("gate_id")
+        if gate.get("status") not in {"ready", "approval_required", "blocked_by_policy"}:
+            raise ValueError(f"{label}: approval gate {gate_id} status is invalid")
+        if gate.get("runtime_write_allowed") is not False:
+            raise ValueError(f"{label}: approval gate {gate_id} must not allow runtime write")
+        if not gate.get("required_evidence") or not gate.get("blocked_runtime_action"):
+            raise ValueError(f"{label}: approval gate {gate_id} must include required evidence and blocked action")
+
+
+def _require_verifier_plan(payload: dict[str, Any], label: str) -> None:
+    checks = payload.get("post_write_verifier_plan") or []
+    required = {
+        "verifier_result_capture_receipt",
+        "verifier_learning_write_boundary",
+        "verifier_playbook_trust_delta",
+        "verifier_gateway_queue_noop",
+        "verifier_rollback_reverify",
+    }
+    verifier_ids = {check.get("verifier_id") for check in checks}
+    if verifier_ids != required:
+        raise ValueError(f"{label}: post-write verifier plan must match {sorted(required)}")
+    for check in checks:
+        verifier_id = check.get("verifier_id")
+        if check.get("live_execution_enabled") is not False:
+            raise ValueError(f"{label}: verifier {verifier_id} must not enable live execution")
+        if check.get("status") not in {"ready", "approval_required", "blocked_by_policy"}:
+            raise ValueError(f"{label}: verifier {verifier_id} status is invalid")
+        if not check.get("verifies") or not check.get("failure_if_missing"):
+            raise ValueError(f"{label}: verifier {verifier_id} must include verifies and failure_if_missing")
+        if not _is_redacted_sha256(check.get("evidence_hash")):
+            raise ValueError(f"{label}: verifier {verifier_id} must expose redacted evidence_hash")
+
+
+def _require_blocked_writes(payload: dict[str, Any], label: str) -> None:
+    blockers = payload.get("blocked_live_writes") or []
+    required = {
+        "blocked_result_capture_write",
+        "blocked_learning_write",
+        "blocked_playbook_trust_write",
+        "blocked_reviewer_queue_write",
+        "blocked_gateway_queue_write",
+        "blocked_telegram_send",
+    }
+    blocker_ids = {blocker.get("blocker_id") for blocker in blockers}
+    if blocker_ids != required:
+        raise ValueError(f"{label}: blocked live writes must match {sorted(required)}")
+    for blocker in blockers:
+        blocker_id = blocker.get("blocker_id")
+        if blocker.get("status") not in {"approval_required", "blocked_by_policy"}:
+            raise ValueError(f"{label}: blocker {blocker_id} status is invalid")
+        if blocker.get("severity") not in {"high", "critical"}:
+            raise ValueError(f"{label}: blocker {blocker_id} severity is invalid")
+        if not blocker.get("blocked_action") or not blocker.get("blocked_until"):
+            raise ValueError(f"{label}: blocker {blocker_id} must include blocked_action and blocked_until")
+        if not _is_redacted_sha256(blocker.get("evidence_hash")):
+            raise ValueError(f"{label}: blocker {blocker_id} must expose redacted evidence_hash")
+
+
+def _require_actions(payload: dict[str, Any], label: str) -> None:
+    actions = payload.get("operator_actions") or []
+    required = {
+        "review_write_gate_packet",
+        "collect_dual_approval",
+        "verify_dry_run_hashes",
+        "confirm_post_write_verifier_plan",
+        "promote_to_p2_122",
+    }
+    action_ids = {action.get("action_id") for action in actions}
+    if action_ids != required:
+        raise ValueError(f"{label}: operator actions must match {sorted(required)}")
+    for action in actions:
+        action_id = action.get("action_id")
+        if action.get("runtime_write_allowed") is not False:
+            raise ValueError(f"{label}: action {action_id} must not allow runtime write")
+        if not action.get("operator_instruction"):
+            raise ValueError(f"{label}: action {action_id} operator_instruction is required")
+
+
+def _require_display_redaction(payload: dict[str, Any], label: str) -> None:
+    contract = payload.get("display_redaction_contract") or {}
+    expected = {
+        "redaction_required": True,
+        "raw_prompt_display_allowed": False,
+        "private_reasoning_display_allowed": False,
+        "secret_value_display_allowed": False,
+        "raw_runtime_payload_display_allowed": False,
+        "internal_collaboration_content_display_allowed": False,
+    }
+    mismatches = _mismatches(contract, expected)
+    if mismatches:
+        raise ValueError(f"{label}: display_redaction_contract mismatch: {mismatches}")
+    if not contract.get("frontend_display_policy"):
+        raise ValueError(f"{label}: display_redaction_contract.frontend_display_policy is required")
+
+
+def _require_rollup_consistency(payload: dict[str, Any], label: str) -> None:
+    rollups = payload.get("rollups") or {}
+    reviews = payload.get("write_gate_reviews") or []
+    gates = payload.get("approval_gates") or []
+    verifiers = payload.get("post_write_verifier_plan") or []
+    blockers = payload.get("blocked_live_writes") or []
+    actions = payload.get("operator_actions") or []
+    expected = {
+        "write_gate_review_count": len(reviews),
+        "approval_gate_count": len(gates),
+        "post_write_verifier_plan_count": len(verifiers),
+        "blocked_live_write_count": len(blockers),
+        "operator_action_count": len(actions),
+        "approval_required_review_count": sum(1 for item in reviews if item.get("status") == "approval_required"),
+        "blocked_review_count": sum(1 for item in reviews if item.get("status") == "blocked_by_policy"),
+        "approval_required_gate_count": sum(1 for item in gates if item.get("status") == "approval_required"),
+        "blocked_gate_count": sum(1 for item in gates if item.get("status") == "blocked_by_policy"),
+        "approval_required_verifier_count": sum(1 for item in verifiers if item.get("status") == "approval_required"),
+        "critical_blocker_count": sum(1 for item in blockers if item.get("severity") == "critical"),
+        "owner_approval_received_count": 0,
+        "dual_approval_received_count": 0,
+        "dry_run_hash_verified_count": 0,
+        "post_write_verifier_pass_count": 0,
+        "rollback_plan_verified_count": 0,
+        "canonical_runtime_target_read_count": 0,
+        "live_query_count": 0,
+        "reviewer_queue_write_count": 0,
+        "gateway_queue_write_count": 0,
+        "telegram_send_count": 0,
+        "bot_api_call_count": 0,
+        "report_receipt_write_count": 0,
+        "result_capture_write_count": 0,
+        "learning_write_count": 0,
+        "playbook_trust_write_count": 0,
+        "production_write_count": 0,
+        "secret_read_count": 0,
+        "destructive_operation_count": 0,
+    }
+    mismatches = _mismatches(rollups, expected)
+    if mismatches:
+        raise ValueError(f"{label}: rollup counts mismatch: {mismatches}")
+
+
+def _require_no_forbidden_display_terms(payload: dict[str, Any], label: str) -> None:
+    serialized = json.dumps(payload, ensure_ascii=False)
+    forbidden = {
+        "work_window_transcript",
+        "session_id",
+        "browser_context",
+        "authorization_header",
+        "raw Telegram payload",
+        "private reasoning",
+        "raw prompt",
+        "chain-of-thought",
+    }
+    hits = sorted(term for term in forbidden if term in serialized)
+    if hits:
+        raise ValueError(f"{label}: forbidden display terms present: {hits}")
+
+
+def _is_redacted_sha256(value: Any) -> bool:
+    if not isinstance(value, str) or not value.startswith("sha256:"):
+        return False
+    digest = value.removeprefix("sha256:")
+    return len(digest) == 64 and all(char in "0123456789abcdef" for char in digest)
+
+
+def _mismatches(payload: dict[str, Any], expected: dict[str, Any]) -> dict[str, dict[str, Any]]:
+    return {
+        key: {"expected": value, "actual": payload.get(key)}
+        for key, value in expected.items()
+        if payload.get(key) != value
+    }