feat(governance): 新增 runtime verifier evidence review

apps/api/src/services/ai_agent_runtime_verifier_evidence_review.py

@@ -0,0 +1,152 @@
+"""
+AI Agent runtime verifier evidence review snapshot.
+
+Loads the latest committed P2-403I runtime verifier evidence implementation
+review. This module never implements or executes the verifier, reads canonical
+targets, writes rollback work items, sends Telegram receipts, or starts a
+runtime worker.
+"""
+
+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_verifier_evidence_review_*.json"
+_SCHEMA_VERSION = "ai_agent_runtime_verifier_evidence_review_v1"
+
+
+def load_latest_ai_agent_runtime_verifier_evidence_review(
+    evaluations_dir: Path | None = None,
+) -> dict[str, Any]:
+    """Load the newest committed AI Agent runtime verifier evidence review."""
+    directory = evaluations_dir or _DEFAULT_EVALUATIONS_DIR
+    candidates = sorted(directory.glob(_SNAPSHOT_PATTERN))
+    if not candidates:
+        raise FileNotFoundError(
+            f"no AI Agent runtime verifier evidence 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_runtime_boundaries(payload, str(latest))
+    _require_review_contract(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_verifier_evidence_review_only_no_live_execution":
+        raise ValueError(
+            f"{label}: runtime_authority must remain runtime_verifier_evidence_review_only_no_live_execution"
+        )
+
+
+def _require_runtime_boundaries(payload: dict[str, Any], label: str) -> None:
+    boundaries = payload.get("approval_boundaries") or {}
+    enabled = sorted(key for key, value in boundaries.items() if value is not False)
+    if enabled:
+        raise ValueError(f"{label}: approval boundaries must remain false: {enabled}")
+
+    truth = payload.get("review_truth") or {}
+    false_flags = {
+        "runtime_verifier_implementation_allowed",
+        "post_write_verifier_execution_allowed",
+        "canonical_readback_allowed",
+    }
+    unsafe = sorted(flag for flag in false_flags if truth.get(flag) is not False)
+    if unsafe:
+        raise ValueError(f"{label}: verifier review runtime flags must remain false: {unsafe}")
+
+    zero_counts = {
+        "runtime_verifier_executed_count",
+        "canonical_readback_executed_count",
+        "rollback_work_item_created_count",
+        "telegram_failure_receipt_sent_count",
+        "learning_writeback_after_verifier_count",
+    }
+    non_zero = sorted(key for key in zero_counts if truth.get(key) != 0)
+    if non_zero:
+        raise ValueError(f"{label}: verifier review counts must remain zero: {non_zero}")
+
+
+def _require_review_contract(payload: dict[str, Any], label: str) -> None:
+    package = payload.get("review_package") or {}
+    required_evidence = set(package.get("required_evidence") or [])
+    required_minimum = {
+        "approved_write_event_id",
+        "dry_run_preview_hash",
+        "target_write_surface",
+        "canonical_readback_query_plan",
+        "rollback_work_item_template",
+        "failure_receipt_template",
+        "redacted_evidence_refs",
+    }
+    missing = sorted(required_minimum - required_evidence)
+    if missing:
+        raise ValueError(f"{label}: runtime verifier review missing required evidence: {missing}")
+
+    if not payload.get("evidence_checks"):
+        raise ValueError(f"{label}: evidence checks must not be empty")
+    if not payload.get("implementation_review_lanes"):
+        raise ValueError(f"{label}: implementation review lanes must not be empty")
+    if not payload.get("operator_actions"):
+        raise ValueError(f"{label}: operator actions must not be empty")
+
+    redaction = payload.get("display_redaction_contract") or {}
+    if redaction.get("redaction_required") is not True:
+        raise ValueError(f"{label}: frontend redaction must be required")
+    for flag in ("raw_payload_display_allowed", "private_reasoning_display_allowed", "secret_value_display_allowed"):
+        if redaction.get(flag) is not False:
+            raise ValueError(f"{label}: {flag} must remain false")
+
+
+def _require_rollup_consistency(payload: dict[str, Any], label: str) -> None:
+    rollups = payload.get("rollups") or {}
+    checks = payload.get("evidence_checks") or []
+    lanes = payload.get("implementation_review_lanes") or []
+    actions = payload.get("operator_actions") or []
+    package = payload.get("review_package") or {}
+    expected_counts = {
+        "evidence_check_count": len(checks),
+        "implementation_review_lane_count": len(lanes),
+        "operator_action_count": len(actions),
+        "blocked_runtime_action_count": len({
+            *(check.get("blocked_runtime_action") for check in checks),
+            *(lane.get("blocked_runtime_action") for lane in lanes),
+            *(action.get("blocked_runtime_action") for action in actions),
+        }),
+        "required_evidence_count": len(package.get("required_evidence") or []),
+        "forbidden_evidence_count": len(package.get("forbidden_evidence") or []),
+    }
+    mismatched = {
+        key: {"expected": expected, "actual": rollups.get(key)}
+        for key, expected in expected_counts.items()
+        if rollups.get(key) != expected
+    }
+    if mismatched:
+        raise ValueError(f"{label}: rollup counts must match payload sections: {mismatched}")
+
+    approval_required = sorted(
+        action.get("action_id")
+        for action in actions
+        if action.get("status") == "approval_required"
+    )
+    if sorted(rollups.get("approval_required_action_ids") or []) != approval_required:
+        raise ValueError(f"{label}: rollups.approval_required_action_ids mismatch")
+    if rollups.get("live_verifier_execution_count") != 0:
+        raise ValueError(f"{label}: live verifier execution count must remain zero")