tegwick
26934e25b9
Assistant: codex Assistant-Model: gpt-5.6-sol Assistant-Session: 01a028de-e2c8-7732-8521-46a7fc5db82f
68 lines
1.8 KiB
Python
68 lines
1.8 KiB
Python
"""Bounded, non-secret projections for durable audit surfaces."""
|
|
|
|
from __future__ import annotations
|
|
|
|
from typing import Any
|
|
|
|
_SENSITIVE_KEYS = frozenset(
|
|
{
|
|
"api_key",
|
|
"api_token",
|
|
"access_token",
|
|
"refresh_token",
|
|
"authorization",
|
|
"client_secret",
|
|
"cookie",
|
|
"credential",
|
|
"credentials",
|
|
"messages",
|
|
"model_messages",
|
|
"password",
|
|
"private_key",
|
|
"provider_payload",
|
|
"provider_response",
|
|
"raw_output",
|
|
"raw_output_preview",
|
|
"raw_prompt",
|
|
"rendered_prompt",
|
|
"secret",
|
|
"token",
|
|
"tool_error",
|
|
"tool_output",
|
|
}
|
|
)
|
|
|
|
|
|
def bounded_audit_projection(
|
|
value: Any,
|
|
*,
|
|
max_depth: int = 8,
|
|
max_items: int = 100,
|
|
max_string: int = 4000,
|
|
) -> Any:
|
|
"""Return a JSON-like audit projection without raw or credential fields.
|
|
|
|
The projection is intentionally lossy. Workflow evaluation uses the full
|
|
in-memory context; only the durable audit copy is bounded here.
|
|
"""
|
|
|
|
def project(item: Any, depth: int) -> Any:
|
|
if depth > max_depth:
|
|
return "<depth-limit>"
|
|
if item is None or isinstance(item, (bool, int, float)):
|
|
return item
|
|
if isinstance(item, str):
|
|
return item[:max_string]
|
|
if isinstance(item, dict):
|
|
result: dict[str, Any] = {}
|
|
for raw_key, child in list(item.items())[:max_items]:
|
|
key = str(raw_key)
|
|
if key.strip().lower() in _SENSITIVE_KEYS:
|
|
continue
|
|
result[key] = project(child, depth + 1)
|
|
return result
|
|
if isinstance(item, (list, tuple)):
|
|
return [project(child, depth + 1) for child in item[:max_items]]
|
|
return str(item)[:max_string]
|
|
|
|
return project(value, 0)
|