Rejection Reasons

X-Fairvisor-Reason is set on reject responses.

Core reasons

Reason	Meaning
`no_bundle_loaded`	No active bundle is loaded
`kill_switch`	Request matched an active kill switch
`token_bucket_exceeded`	Token bucket denied request
`budget_exceeded`	Cost budget exhausted
`circuit_breaker_open`	Policy circuit breaker is open
`loop_detected`	Loop detector threshold exceeded

LLM-specific reasons

Reason	Meaning
`prompt_tokens_exceeded`	Prompt estimate exceeded max prompt limit
`max_tokens_per_request_exceeded`	Prompt + completion reservation exceeded per-request cap
`tpm_exceeded`	Tokens-per-minute budget exhausted
`tpd_exceeded`	Tokens-per-day budget exhausted

Streaming reason

completion_tokens_exceeded may appear in streaming SSE error payloads when mid-stream completion limit is hit. It is not emitted as X-Fairvisor-Reason header in the normal reject path.

Shadow mode

In shadow mode, would-reject reasons are kept in internal decision metadata while HTTP response is allow (200).