Skip to content

[health]

Provider health checks, rate-limit cooldowns and funding probes.

Key Type Default Description
health.check_interval_seconds integer 60 check_interval_seconds is how often Prism checks each provider’s health (default 60).
health.rate_limit_burst_threshold integer 3 rate_limit_burst_threshold is how many consecutive 429 answers inside rate_limit_burst_window_seconds put a provider’s model into cooldown (default 3).
health.rate_limit_burst_window_seconds integer 60 rate_limit_burst_window_seconds is the window for counting consecutive 429 answers (default 60).
health.rate_limit_initial_cooldown_seconds integer 30 rate_limit_initial_cooldown_seconds is the first cooldown of a rate-limited model (default 30); repeated cooldowns grow up to rate_limit_max_cooldown_seconds.
health.rate_limit_max_cooldown_seconds integer 600 rate_limit_max_cooldown_seconds is the longest cooldown of a rate-limited model (default 600).
health.provider_funding_probe_cooldown_seconds integer 900 One provider-account funding probe per window; successful probes clear state. The window doubles after each probe refused for funds, up to 8x (15 min -> 2 h at the default), and starts over after recovery. Cached 402s (in-window, no probe) say so in the body; POST /prism/funding/clear or clear_provider_funding with provider / all=true recovers without waiting.
health.recovery_dwell_seconds integer 120 recovery_dwell_seconds is the minimum span of the consecutive successful probes that return an unavailable provider to service (hysteresis against flapping). Defaults to 120.
health.role_check.enabled bool false enabled turns the check on (default false). It only sends a request to a role target that no live request has served in the last six hours, typically a backup target, so a backup that ran out of credit or stopped answering shows up in the role’s state before it is needed.