Files
awoooi/apps/api/src/services/ai_provider_route_matrix.py
Your Name 45556f8fd1
All checks were successful
CD Pipeline / tests (push) Successful in 1m29s
Code Review / ai-code-review (push) Successful in 15s
CD Pipeline / build-and-deploy (push) Successful in 3m51s
CD Pipeline / post-deploy-checks (push) Successful in 1m30s
feat(governance): 新增 AI Provider 路由矩陣
2026-06-05 13:28:38 +08:00

255 lines
10 KiB
Python

"""
AI provider route matrix snapshot.
Loads the latest committed, read-only AI Router / Ollama / OpenClaw /
Nemotron / Gemini provider route matrix. This module never switches providers,
calls paid APIs, probes external providers, reads secrets, changes runtime
routes, enters shadow/canary, triggers workflows, or executes runtime actions.
"""
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_provider_route_matrix_*.json"
_SCHEMA_VERSION = "ai_provider_route_matrix_v1"
def load_latest_ai_provider_route_matrix(
evaluations_dir: Path | None = None,
) -> dict[str, Any]:
"""Load the newest committed AI provider route matrix snapshot."""
directory = evaluations_dir or _DEFAULT_EVALUATIONS_DIR
candidates = sorted(directory.glob(_SNAPSHOT_PATTERN))
if not candidates:
raise FileNotFoundError(f"no AI provider route matrix 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, _SCHEMA_VERSION, str(latest))
_require_read_only_boundaries(payload, str(latest))
_require_operation_boundaries(payload, str(latest))
_require_rollup_consistency(payload, str(latest))
_require_route_evidence(payload, str(latest))
_require_candidate_gates(payload, str(latest))
_require_operator_denials(payload, str(latest))
_require_no_plaintext_secret_payload_keys(payload, str(latest))
return payload
def _require_schema(payload: dict[str, Any], expected: str, label: str) -> None:
actual = payload.get("schema_version")
if actual != expected:
raise ValueError(f"{label}: expected schema_version={expected}, got {actual!r}")
def _require_read_only_boundaries(payload: dict[str, Any], label: str) -> None:
program_status = payload.get("program_status") or {}
if program_status.get("read_only_mode") is not True:
raise ValueError(f"{label}: program_status.read_only_mode must be true")
approval_boundaries = payload.get("approval_boundaries") or {}
allowed = sorted(key for key, value in approval_boundaries.items() if value is not False)
if allowed:
raise ValueError(f"{label}: approval boundaries must remain false: {allowed}")
def _require_operation_boundaries(payload: dict[str, Any], label: str) -> None:
boundaries = payload.get("operation_boundaries") or {}
if boundaries.get("read_only_api_allowed") is not True:
raise ValueError(f"{label}: read_only_api_allowed must be true")
blocked_flags = {
"provider_switch_allowed",
"production_routing_change_allowed",
"use_ai_router_toggle_allowed",
"fallback_order_change_allowed",
"ollama_endpoint_change_allowed",
"paid_api_call_allowed",
"paid_api_frequency_increase_allowed",
"external_provider_probe_allowed",
"live_benchmark_allowed",
"shadow_or_canary_allowed",
"openclaw_replacement_allowed",
"nemotron_shadow_allowed",
"gemini_direct_call_allowed",
"secret_read_allowed",
"secret_plaintext_allowed",
"notification_send_allowed",
"workflow_trigger_allowed",
"deploy_trigger_allowed",
"reload_trigger_allowed",
"runtime_execution_allowed",
}
allowed = sorted(flag for flag in blocked_flags if boundaries.get(flag) is not False)
if allowed:
raise ValueError(f"{label}: operation boundaries must remain false: {allowed}")
def _require_rollup_consistency(payload: dict[str, Any], label: str) -> None:
routes = payload.get("provider_routes") or []
gates = payload.get("candidate_gates") or []
gaps = payload.get("source_gaps") or []
rollups = payload.get("rollups") or {}
if rollups.get("total_routes") != len(routes):
raise ValueError(f"{label}: rollups.total_routes must match provider_routes")
if rollups.get("by_kind") != _count_by(routes, "kind"):
raise ValueError(f"{label}: rollups.by_kind must match provider_routes")
if rollups.get("by_status") != _count_by(routes, "status"):
raise ValueError(f"{label}: rollups.by_status must match provider_routes")
if rollups.get("by_route_gate") != _count_by(routes, "route_gate"):
raise ValueError(f"{label}: rollups.by_route_gate must match provider_routes")
requiring_action = sorted(
route.get("route_id")
for route in routes
if route.get("status") == "action_required"
)
if sorted(rollups.get("route_ids_requiring_action") or []) != requiring_action:
raise ValueError(f"{label}: rollups.route_ids_requiring_action must match routes")
approval_gates = sorted(
gate.get("gate_id")
for gate in gates
if gate.get("approval_required") is True
)
if sorted(rollups.get("candidate_gate_ids_requiring_approval") or []) != approval_gates:
raise ValueError(
f"{label}: rollups.candidate_gate_ids_requiring_approval must match candidate_gates"
)
if sorted(rollups.get("source_gap_ids") or []) != sorted(gap.get("gap_id") for gap in gaps):
raise ValueError(f"{label}: rollups.source_gap_ids must match source_gaps")
zero_count_fields = {
"provider_switch_allowed_count",
"paid_api_call_allowed_count",
"shadow_or_canary_allowed_count",
"runtime_route_change_allowed_count",
}
non_zero = sorted(field for field in zero_count_fields if rollups.get(field) != 0)
if non_zero:
raise ValueError(f"{label}: route permission rollup counts must remain 0: {non_zero}")
def _require_route_evidence(payload: dict[str, Any], label: str) -> None:
routes = payload.get("provider_routes") or []
missing = sorted(
route.get("route_id")
for route in routes
if not route.get("current_policy")
or not route.get("provider_order")
or not route.get("fallback_policy")
or not route.get("evidence_refs")
or not route.get("next_action")
)
if missing:
raise ValueError(f"{label}: provider_routes must include policy, order, evidence, next_action: {missing}")
required_route_ids = {
"ai_router_execution_policy",
"ollama_global_endpoint_order",
"alert_ai_ollama_first_lane",
"openclaw_nemo_rca_lane",
"nemotron_tool_calling_candidate",
"gemini_final_fallback_policy",
}
present = {route.get("route_id") for route in routes}
missing_required = sorted(required_route_ids - present)
if missing_required:
raise ValueError(f"{label}: missing required provider routes: {missing_required}")
blocked_candidates = sorted(
route.get("route_id")
for route in routes
if route.get("kind") == "nemotron_candidate" and route.get("status") != "blocked"
)
if blocked_candidates:
raise ValueError(f"{label}: Nemotron candidates must remain blocked: {blocked_candidates}")
def _require_candidate_gates(payload: dict[str, Any], label: str) -> None:
gates = payload.get("candidate_gates") or []
required_ids = {
"provider_switch_gate",
"nemotron_replay_gate",
"paid_provider_call_gate",
}
present = {gate.get("gate_id") for gate in gates}
missing = sorted(required_ids - present)
if missing:
raise ValueError(f"{label}: missing required candidate gates: {missing}")
not_approval_required = sorted(
gate.get("gate_id")
for gate in gates
if gate.get("approval_required") is not True
)
if not_approval_required:
raise ValueError(f"{label}: candidate gates must require approval: {not_approval_required}")
def _require_operator_denials(payload: dict[str, Any], label: str) -> None:
contract = payload.get("operator_contract") or {}
must_not_interpret_as = set(contract.get("must_not_interpret_as") or [])
required_denials = {
"provider 切換批准",
"production routing change 批准",
"Gemini / Claude / NVIDIA 付費呼叫批准",
"OpenClaw 取代或降級批准",
"Nemotron 進 shadow / canary 批准",
"fallback order 修改批准",
"Ollama endpoint / ConfigMap 修改批准",
"Secret payload 已讀取或可輸出",
"外部 live probe 或 benchmark 批准",
"workflow / deploy / reload 觸發批准",
"runtime execution 授權",
}
if not required_denials.issubset(must_not_interpret_as):
raise ValueError(f"{label}: operator_contract.must_not_interpret_as is missing required denials")
provider_switch_policy = str(contract.get("provider_switch_policy") or "")
if "P1-004" not in provider_switch_policy or "只能盤點與顯示" not in provider_switch_policy:
raise ValueError(f"{label}: provider_switch_policy must preserve P1-004 read-only boundary")
def _require_no_plaintext_secret_payload_keys(value: Any, label: str, path: str = "$") -> None:
if isinstance(value, dict):
forbidden_key_fragments = {
"secret_value",
"token_value",
"authorization_header",
"private_key",
"webhook_secret",
"runner_token",
"gemini_api_key_value",
"nvidia_api_key_value",
"claude_api_key_value",
}
for key, nested in value.items():
lowered = str(key).lower()
if any(fragment in lowered for fragment in forbidden_key_fragments):
raise ValueError(f"{label}: forbidden secret payload key at {path}.{key}")
_require_no_plaintext_secret_payload_keys(nested, label, f"{path}.{key}")
elif isinstance(value, list):
for index, nested in enumerate(value):
_require_no_plaintext_secret_payload_keys(nested, label, f"{path}[{index}]")
def _count_by(items: list[dict[str, Any]], key: str) -> dict[str, int]:
counts: dict[str, int] = {}
for item in items:
value = item.get(key)
counts[value] = counts.get(value, 0) + 1
return counts