Enforcement

The enforcement layer runs before each API call inside the SDK wrapper. If a rule fires, the call is blocked and an error is returned to your code — the provider is never reached and you are never billed.

Configure rules at token.audit.id/app/enforcement.

Available rules

Token ceiling

Blocks calls where the input prompt exceeds a token limit. Prevents accidental context inflation from bloated prompts.

client = JournaledAnthropic(
    api_key="sk-ant-…",
    session_id="my-session",
    enforcement_config={"token_ceiling": 8000}
)

Session budget

Caps total spend per session ID. When the session crosses the threshold, further calls are blocked until a new session is started.

enforcement_config={"max_session_usd": 2.00}

Dedup guard

Hashes the prompt and blocks exact duplicate calls within a configurable TTL window. Catches retry storms and agent loops that re-submit identical requests.

enforcement_config={"dedup_ttl_seconds": 60}

Loop breaker

Counts tool calls within a session. Blocks the call if the number of tool calls exceeds the configured ceiling. Stops runaway agent loops.

enforcement_config={"max_tool_calls": 50}

Model router

Downgrades expensive model calls to a cheaper alternative when the configured cost threshold would be exceeded.

enforcement_config={"model_downgrade": {"threshold_usd": 0.01, "fallback": "claude-haiku-4-5-20251001"}}

Remote config

Rules can also be set and updated from the dashboard without redeploying. The wrapper fetches remote config on startup and caches it for the session.