Struct AgentLoopConfig 

pub struct AgentLoopConfig {

    pub model_config: ModelConfig,
    pub provider_override: Option<Arc<dyn StreamProvider>>,
    pub thinking_level: ThinkingLevel,
    pub max_tokens: Option<u32>,
    pub temperature: Option<f32>,
    pub convert_to_llm: Option<ConvertToLlmFn>,
    pub transform_context: Option<TransformContextFn>,
    pub get_steering_messages: Option<GetMessagesFn>,
    pub get_follow_up_messages: Option<GetMessagesFn>,
    pub context_config: Option<ContextConfig>,
    pub execution_limits: Option<ExecutionLimits>,
    pub cache_config: CacheConfig,
    pub tool_execution: ToolExecutionStrategy,
    pub tool_timeout: Option<Duration>,
    pub retry_config: RetryConfig,
    pub before_turn: Option<BeforeTurnFn>,
    pub after_turn: Option<AfterTurnFn>,
    pub before_loop: Option<BeforeLoopFn>,
    pub after_loop: Option<AfterLoopFn>,
    pub before_tool_execution: Option<BeforeToolExecutionFn>,
    pub after_tool_execution: Option<AfterToolExecutionFn>,
    pub before_tool_execution_update: Option<BeforeToolExecutionUpdateFn>,
    pub after_tool_execution_update: Option<AfterToolExecutionUpdateFn>,
    pub on_error: Option<OnErrorFn>,
    pub before_compaction_start: Option<BeforeCompactionStartFn>,
    pub after_compaction_end: Option<AfterCompactionEndFn>,
    pub input_filters: Vec<Arc<dyn InputFilter>>,
    pub first_turn_trigger: TurnTrigger,
    pub config_id: Option<String>,
    pub context_translation: Option<Arc<dyn ContextTranslationStrategy>>,
    pub prun_pending: Option<Arc<Mutex<Vec<PrunRequest>>>>,
    pub response_format: ResponseFormat,
}

All static settings for a single [agent_loop] / [agent_loop_continue] call.

Build with the public fields directly or via [crate::agent::Agent]’s builder methods. The config is borrowed (&AgentLoopConfig) throughout the loop — it is never mutated.

Lifecycle hooks

All hook fields are Option<Arc<dyn Fn(...)>>. None means “no hook” (zero overhead). See the module-level doc for the guaranteed ordering relative to AgentEvents.

Fields

model_config: ModelConfig

Complete provider identity: model id, api_key, base_url, protocol, compat flags, cost rates. The agent loop resolves the concrete StreamProvider from model_config.api via ProviderRegistry. Set provider_override to bypass the registry for custom providers.

provider_override: Option<Arc<dyn StreamProvider>>

Custom provider override. When Some, bypasses ProviderRegistry dispatch and uses this provider directly. Useful for testing (MockProvider) or custom implementations. When None (the default), the provider is resolved from model_config.api.

thinking_level: ThinkingLevel

max_tokens: Option<u32>

temperature: Option<f32>

convert_to_llm: Option<ConvertToLlmFn>

Convert AgentMessage[] → Message[] before each LLM call. Default: keep only LLM-compatible messages.

transform_context: Option<TransformContextFn>

Transform context before convert_to_llm (for pruning, compaction).

get_steering_messages: Option<GetMessagesFn>

Get steering messages (user interruptions mid-run).

get_follow_up_messages: Option<GetMessagesFn>

Get follow-up messages (queued work after agent finishes).

context_config: Option<ContextConfig>

Context window configuration (auto-compaction). Compaction strategies are now part of ContextConfig.compaction (G5 consolidation).

execution_limits: Option<ExecutionLimits>

Execution limits (max turns, tokens, duration, cost). Cost is tracked automatically using model_config.cost rates after each turn. ExecutionLimits.max_cost enforcement is active whenever rates are non-zero.

cache_config: CacheConfig

Prompt caching configuration.

tool_execution: ToolExecutionStrategy

Tool execution strategy (sequential, parallel, or batched).

tool_timeout: Option<Duration>

Per-tool execution timeout.

When Some(d), each individual AgentTool::execute() call is bounded by d. On expiry, the tool’s child cancel token is signalled (cooperative cleanup) and a ToolError::Timeout is synthesised as the tool result — the LLM sees the failure and the agent loop continues. A per-tool override via AgentTool::timeout() takes precedence over this field. None (the default) means no per-tool timeout.

retry_config: RetryConfig

Retry configuration for transient provider errors.

before_turn: Option<BeforeTurnFn>

Called before each LLM turn. Return false to abort the turn.

after_turn: Option<AfterTurnFn>

Called after each LLM turn with the current messages and the turn’s usage.

before_loop: Option<BeforeLoopFn>

Called before each Agent loop. Return false to abort the loop.

after_loop: Option<AfterLoopFn>

Called after each Agent loop with the current messages and the loop’s usage.

before_tool_execution: Option<BeforeToolExecutionFn>

Called before each tool execution. Return false to skip the tool call.

after_tool_execution: Option<AfterToolExecutionFn>

Called after each tool execution.

before_tool_execution_update: Option<BeforeToolExecutionUpdateFn>

Called before each ToolExecutionUpdate event. Return false to suppress the event.

after_tool_execution_update: Option<AfterToolExecutionUpdateFn>

Called after each ToolExecutionUpdate event.

on_error: Option<OnErrorFn>

Called when the LLM returns a StopReason::Error.

before_compaction_start: Option<BeforeCompactionStartFn>

Called before compaction starts. Return false to skip compaction.

after_compaction_end: Option<AfterCompactionEndFn>

Called after compaction completes.

input_filters: Vec<Arc<dyn InputFilter>>

Input filters applied to user messages before the LLM call. Filters run in order; first Reject wins and discards any accumulated warnings. Warn messages accumulate and are appended to the user message.

first_turn_trigger: TurnTrigger

The trigger type for the first TurnStart event in this run. Defaults to TurnTrigger::User; set to SubAgent by sub-agent callers.

config_id: Option<String>

Stable identity for this config, used as the middle segment of loop_id: loop_id = "{session_id}.{config_id}.{N}"

When None and the Agent wrapper is used, the identity is auto-derived by Agent::next_loop_id() from the provider, model, and thinking level: "{provider_id}.{model_slug}[.thinking]"

For direct callers of agent_loop, set context.loop_id explicitly — this field is only read by Agent::next_loop_id() and has no effect inside agent_loop itself.

Set explicitly for human-readable or deterministic loop IDs, e.g.: config.config_id = Some("experiment-A".to_string()); → loop IDs: ses_xyz.experiment-A.1, ses_xyz.experiment-A.2, …

context_translation: Option<Arc<dyn ContextTranslationStrategy>>

G8 — Optional context translation strategy for cross-provider compatibility.

When set, messages are translated through this strategy before being sent to the LLM provider. This allows content types from one provider (e.g., Content::Thinking from Anthropic) to be translated or removed when targeting a different provider. The translation is read-only — originals are never modified.

prun_pending: Option<Arc<Mutex<Vec<PrunRequest>>>>

Shared state for PrunTool to communicate pruning requests to the loop.

response_format: ResponseFormat

Desired LLM output shape. Default Text preserves the historical free-form behaviour; JsonObject / JsonSchema request constrained structured output from providers that support it. See provider::ResponseFormat and the capability matrix in docs/specs/developer/provider.md.