Struct LimitsConfig

Source

pub struct LimitsConfig {Show 17 fields
    pub max_concurrent_inferences: Option<usize>,
    pub max_concurrent_tools: usize,
    pub default_max_iterations: Option<usize>,
    pub exact_token_counting: bool,
    pub script_shell_timeout_secs: u64,
    pub stall_timeout_secs: u64,
    pub dead_cycles_before_relief: u32,
    pub finished_retention_secs: u64,
    pub mcp_idle_disconnect_secs: u64,
    pub wedge_timeout_secs: u64,
    pub provider_failures_before_open: u32,
    pub provider_circuit_cooldown_secs: u64,
    pub interaction_timeout_secs: u64,
    pub inference_retry_attempts: u32,
    pub inference_retry_base_ms: u64,
    pub max_tool_call_write_bytes: Option<u64>,
    pub max_run_write_bytes: Option<u64>,
}

Expand description

Runtime resource limits with safe defaults baked in.

Both fields default to a bounded value so a fresh install can’t accidentally run unbounded inference concurrency or an unbounded agent loop. Set a field explicitly in [limits] to raise or lower it.

Fields§

§max_concurrent_inferences: Option<usize>

Global fallback cap on concurrent inference requests for any model without its own per-model pool entry. Defaults to Some(8); omit or set a large number to effectively unbound it.

One physical bound sits behind this for script providers: each of their in-flight calls occupies a blocking-pool thread, and the daemon’s runtime provisions 2048 of those. Pools above 2048 only run that wide for HTTP providers, whose calls are fully async.

§max_concurrent_tools: usize

Size of the shared tool-execution worker pool - the number of agents whose tool batches may run concurrently across the whole daemon (the tool-lane counterpart of max_concurrent_inferences). Defaults to 8. Clamped to at least 1.

§default_max_iterations: Option<usize>

Fallback max_iterations applied to a stage that does not set its own, so an agent can’t loop forever with no completion signal. Defaults to Some(50). A stage’s explicit max_iterations always wins.

§exact_token_counting: bool

Opt-in exact pre-inference token budgeting. When true, each agent inference is preceded by an exact token count of the assembled request (via the provider’s count_tokens, which uses a remote endpoint for Anthropic/Gemini and a local heuristic otherwise) and is rejected before sending if it would exceed the model’s context window. Off by default: normal budgeting uses cheap local estimates, and this adds a network round-trip per inference for providers with a remote count endpoint.

§script_shell_timeout_secs: u64

Wall-clock timeout (seconds) for a Rhai script tool’s shell() host call, mirroring the built-in shell tool’s own 60-second cap so a script can’t hang an agent on a runaway command. Defaults to 60.

§stall_timeout_secs: u64

How long (seconds) a run may sit ready to work but unable to dispatch before it is failed instead of left running.

This only ever fires for something the runtime cannot resolve on its own - today, a stage whose provider is not configured. Waiting for a busy model’s inference pool is ordinary backpressure and is never failed, no matter how long it takes. Defaults to 60; 0 disables the watchdog and restores the old behaviour of waiting indefinitely.