Module llm 

Resolve agent -> model -> provider; build Rig agent.

Modules

retry

Transient-error retry wrapper around a CompletionModel.

Structs

MistralrsWeights

Weight-source spec for a mistralrs-backed model. Lifted off [models.<name>] at resolve time. Only one of model_id / model_path is meaningful in any given instance; validation enforces that, but mistralrs::load is also defensive.

OutrigPromptHook

Per-request hook that traces every tool call to stderr and stops the agent loop after max calls. Cloned by rig per request; shared atomics keep a single turn’s calls counting against the same max.

ResolvedAgent

Fully-resolved view of one agent: every knob the agent loop needs to build a Rig client and run a turn.

Enums

LlmResolveError

Failures that surface while walking agents -> models -> providers or constructing the Rig client. Wrapped into crate::error::OutrigError at the top level via #[from].

ResolvedProvider

Runtime-shaped provider view – mirrors the config LlmProvider enum, but with the env-var-backed ApiKeyRef already resolved to a plain String for the OpenAi variant. Variants are kept in sync with LlmProvider’s.

RigAgent

Runtime-dispatched Rig agent. The OpenAi-backed and mistralrs-backed CompletionModel impls produce concretely different Agent<M> types (Rig’s trait carries associated types, so a single concrete RigAgent can’t carry both). Callers (the agent loop) match on the variant.

Constants

DEFAULT_REQUEST_TIMEOUT_SECS

Default per-request HTTP timeout for OpenAi-style providers when request-timeout-secs is unset (see doc/reference/config.md). Generous enough not to truncate long reasoning completions, and above typical proxy timeouts so a client-side timeout never races a still-in-flight server request.

DEFAULT_TOOL_RESULT_MAX_BYTES

Default byte ceiling applied to each individual MCP tool result before it is handed to Rig and appended to model-visible chat history.

MAX_TOOL_CALLS

Hard max on tool calls per turn. The hook below trips this; rig’s own max_turns is set to the same value as a defense in depth, so whichever fires first surfaces a controllable message.

Functions

build_agent

Build a Rig Agent ready to receive a turn. Preamble, sampling params, and the dynamic-tool list come from resolved plus the caller-supplied MCP-backed adapters. The cache_root argument is the directory into which the mistralrs HF-download path stages model files; it’s ignored for OpenAi providers.

resolve_agent

Walk cfg.agents -> models -> providers to resolve every knob the agent loop needs. Bails with a descriptive error if a reference is dangling or the api-key env var is unset.

resolve_agent_with_device_override

resolve_agent_with_overrides