feat(governance): 新增 result capture promotion gate
This commit is contained in:
@@ -118,6 +118,9 @@ from src.services.ai_agent_reviewer_queue_no_write_readback import (
|
||||
from src.services.ai_agent_result_capture_no_write_readback import (
|
||||
load_latest_ai_agent_result_capture_no_write_readback,
|
||||
)
|
||||
from src.services.ai_agent_result_capture_promotion_approval_gate import (
|
||||
load_latest_ai_agent_result_capture_promotion_approval_gate,
|
||||
)
|
||||
from src.services.ai_agent_runtime_readback_approval_package import (
|
||||
load_latest_ai_agent_runtime_readback_approval_package,
|
||||
)
|
||||
@@ -1623,6 +1626,36 @@ async def get_agent_result_capture_no_write_readback() -> dict[str, Any]:
|
||||
) from exc
|
||||
|
||||
|
||||
@router.get(
|
||||
"/agent-result-capture-promotion-approval-gate",
|
||||
response_model=dict[str, Any],
|
||||
summary="取得 AI Agent result capture promotion approval gate",
|
||||
description=(
|
||||
"讀取最新已提交的 P2-119 result capture promotion approval gate;"
|
||||
"此端點只回傳 promotion approval packet、acceptance template、verifier、"
|
||||
"blocked promotion write 與 operator handoff,不寫 result capture、learning、"
|
||||
"PlayBook trust、Gateway queue,不送 Telegram、不呼叫 Bot API、不讀 canonical runtime target、"
|
||||
"不讀 secret。"
|
||||
),
|
||||
)
|
||||
async def get_agent_result_capture_promotion_approval_gate() -> dict[str, Any]:
|
||||
"""Return the latest read-only result capture promotion approval gate package."""
|
||||
try:
|
||||
payload = await asyncio.to_thread(load_latest_ai_agent_result_capture_promotion_approval_gate)
|
||||
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_result_capture_promotion_approval_gate_invalid", error=str(exc))
|
||||
raise HTTPException(
|
||||
status_code=status.HTTP_500_INTERNAL_SERVER_ERROR,
|
||||
detail="AI Agent result capture promotion approval gate 無效",
|
||||
) from exc
|
||||
|
||||
|
||||
@router.get(
|
||||
"/agent-owner-approved-fixture-dry-run",
|
||||
response_model=dict[str, Any],
|
||||
|
||||
@@ -0,0 +1,365 @@
|
||||
"""
|
||||
AI Agent result capture promotion approval gate snapshot.
|
||||
|
||||
Loads the latest committed P2-119 result capture promotion approval package. This
|
||||
module validates committed evidence only; it never writes result captures,
|
||||
writes learning records, updates PlayBook trust, writes 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_promotion_approval_gate_*.json"
|
||||
_SCHEMA_VERSION = "ai_agent_result_capture_promotion_approval_gate_v1"
|
||||
_RUNTIME_AUTHORITY = "result_capture_promotion_approval_gate_only_no_capture_or_learning_write"
|
||||
_TARGET_PROMOTION = "result_capture_promotion_preview"
|
||||
|
||||
|
||||
def load_latest_ai_agent_result_capture_promotion_approval_gate(
|
||||
evaluations_dir: Path | None = None,
|
||||
) -> dict[str, Any]:
|
||||
"""Load the newest committed result capture promotion approval gate 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 promotion approval gate 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_promotion_packets(payload, label)
|
||||
_require_acceptance_templates(payload, label)
|
||||
_require_verifier_checks(payload, label)
|
||||
_require_blockers(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": "P2-119",
|
||||
"next_task_id": "P2-120",
|
||||
"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_result_capture_no_write_readback") or {}
|
||||
expected = {
|
||||
"schema_version": "ai_agent_result_capture_no_write_readback_v1",
|
||||
"result_capture_readback_fixture_count": 5,
|
||||
"capture_field_mapping_count": 5,
|
||||
"readback_verifier_check_count": 5,
|
||||
"blocked_result_capture_write_count": 5,
|
||||
"operator_action_count": 5,
|
||||
"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_result_capture_no_write_readback mismatch: {mismatches}")
|
||||
if not prior.get("readiness_note"):
|
||||
raise ValueError(f"{label}: prior_result_capture_no_write_readback.readiness_note is required")
|
||||
|
||||
|
||||
def _require_truth(payload: dict[str, Any], label: str) -> None:
|
||||
truth = payload.get("promotion_truth") or {}
|
||||
required_true = {
|
||||
"p2_118_result_capture_readback_loaded",
|
||||
"promotion_approval_package_ready",
|
||||
"acceptance_gate_template_ready",
|
||||
"promotion_verifier_ready",
|
||||
"operator_handoff_ready",
|
||||
"rollback_and_reverify_required",
|
||||
}
|
||||
missing = sorted(field for field in required_true if truth.get(field) is not True)
|
||||
if missing:
|
||||
raise ValueError(f"{label}: promotion ready flags must remain true: {missing}")
|
||||
for field in {"owner_approval_received", "capture_promotion_approved"}:
|
||||
if truth.get(field) is not False:
|
||||
raise ValueError(f"{label}: {field} must remain false before promotion write")
|
||||
|
||||
required_false = {
|
||||
"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",
|
||||
"capture_promotion_approved_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",
|
||||
}
|
||||
non_zero = sorted(field for field in zero_counts if truth.get(field) != 0)
|
||||
if non_zero:
|
||||
raise ValueError(f"{label}: promotion live counters must remain zero: {non_zero}")
|
||||
if not truth.get("truth_note"):
|
||||
raise ValueError(f"{label}: promotion_truth.truth_note is required")
|
||||
|
||||
|
||||
def _require_promotion_packets(payload: dict[str, Any], label: str) -> None:
|
||||
packets = payload.get("promotion_approval_packets") or []
|
||||
required = {
|
||||
"promotion_packet_action_required",
|
||||
"promotion_packet_no_action_review",
|
||||
"promotion_packet_verifier_degraded",
|
||||
"promotion_packet_route_lock",
|
||||
"promotion_packet_owner_acceptance",
|
||||
}
|
||||
packet_ids = {packet.get("packet_id") for packet in packets}
|
||||
if packet_ids != required:
|
||||
raise ValueError(f"{label}: promotion approval packets must match {sorted(required)}")
|
||||
for packet in packets:
|
||||
packet_id = packet.get("packet_id")
|
||||
if packet.get("promotion_mode") != "approval_gate_only":
|
||||
raise ValueError(f"{label}: packet {packet_id} must remain approval_gate_only")
|
||||
if packet.get("result_capture_write_enabled") is not False:
|
||||
raise ValueError(f"{label}: packet {packet_id} must not enable result capture write")
|
||||
if packet.get("status") not in {"ready_for_owner_review", "approval_required", "blocked_by_policy"}:
|
||||
raise ValueError(f"{label}: packet {packet_id} status is invalid")
|
||||
if not packet.get("required_decision") or not packet.get("approval_lane"):
|
||||
raise ValueError(f"{label}: packet {packet_id} must include approval lane and required decision")
|
||||
if not _is_redacted_sha256(packet.get("evidence_hash")):
|
||||
raise ValueError(f"{label}: packet {packet_id} must expose redacted evidence_hash")
|
||||
|
||||
|
||||
def _require_acceptance_templates(payload: dict[str, Any], label: str) -> None:
|
||||
templates = payload.get("acceptance_gate_templates") or []
|
||||
required = {
|
||||
"owner_acceptance_record_template",
|
||||
"capture_write_scope_template",
|
||||
"learning_write_scope_template",
|
||||
"playbook_trust_scope_template",
|
||||
"rollback_and_verifier_template",
|
||||
}
|
||||
template_ids = {template.get("template_id") for template in templates}
|
||||
if template_ids != required:
|
||||
raise ValueError(f"{label}: acceptance gate templates must match {sorted(required)}")
|
||||
for template in templates:
|
||||
template_id = template.get("template_id")
|
||||
if template.get("target_promotion") != _TARGET_PROMOTION:
|
||||
raise ValueError(f"{label}: template {template_id} must target {_TARGET_PROMOTION}")
|
||||
if template.get("promotion_write_enabled") is not False:
|
||||
raise ValueError(f"{label}: template {template_id} must not enable promotion write")
|
||||
if template.get("status") not in {"ready", "approval_required", "blocked_by_policy"}:
|
||||
raise ValueError(f"{label}: template {template_id} status is invalid")
|
||||
if not template.get("required_owner") or not template.get("required_fields"):
|
||||
raise ValueError(f"{label}: template {template_id} required owner and fields are required")
|
||||
if not _is_redacted_sha256(template.get("evidence_hash")):
|
||||
raise ValueError(f"{label}: template {template_id} must expose redacted evidence_hash")
|
||||
|
||||
|
||||
def _require_verifier_checks(payload: dict[str, Any], label: str) -> None:
|
||||
checks = payload.get("promotion_verifier_checks") or []
|
||||
required = {
|
||||
"no_result_capture_write_before_approval",
|
||||
"no_learning_write_before_approval",
|
||||
"no_playbook_trust_before_approval",
|
||||
"no_gateway_queue_before_approval",
|
||||
"promotion_redaction_completeness",
|
||||
}
|
||||
verifier_ids = {check.get("verifier_id") for check in checks}
|
||||
if verifier_ids != required:
|
||||
raise ValueError(f"{label}: promotion verifier checks 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_blockers(payload: dict[str, Any], label: str) -> None:
|
||||
blockers = payload.get("blocked_promotion_writes") or []
|
||||
required = {
|
||||
"result_capture_promotion_not_authorized",
|
||||
"learning_promotion_not_authorized",
|
||||
"playbook_trust_promotion_not_authorized",
|
||||
"gateway_queue_promotion_not_authorized",
|
||||
"production_promotion_not_authorized",
|
||||
}
|
||||
blocker_ids = {blocker.get("blocker_id") for blocker in blockers}
|
||||
if blocker_ids != required:
|
||||
raise ValueError(f"{label}: blocked promotion 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_promotion_approval_packet",
|
||||
"verify_no_write_promotion_counts",
|
||||
"confirm_acceptance_scope",
|
||||
"check_promotion_redaction_contract",
|
||||
"promote_to_p2_120",
|
||||
}
|
||||
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_promotion_write_allowed") is not False:
|
||||
raise ValueError(f"{label}: action {action_id} must not allow runtime promotion 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 {}
|
||||
packets = payload.get("promotion_approval_packets") or []
|
||||
templates = payload.get("acceptance_gate_templates") or []
|
||||
verifiers = payload.get("promotion_verifier_checks") or []
|
||||
blockers = payload.get("blocked_promotion_writes") or []
|
||||
actions = payload.get("operator_actions") or []
|
||||
expected = {
|
||||
"promotion_approval_packet_count": len(packets),
|
||||
"acceptance_gate_template_count": len(templates),
|
||||
"promotion_verifier_check_count": len(verifiers),
|
||||
"blocked_promotion_write_count": len(blockers),
|
||||
"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"),
|
||||
"approval_required_template_count": sum(1 for item in templates if item.get("status") == "approval_required"),
|
||||
"blocked_template_count": sum(1 for item in templates 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,
|
||||
"capture_promotion_approved_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
|
||||
}
|
||||
Reference in New Issue
Block a user