Struct PipelineConfig 

pub struct PipelineConfig {

    pub max_context_tokens: usize,
    pub max_output_tokens: usize,
    pub enable_tools: bool,
    pub iteration_budget_enabled: bool,
    pub max_iterations: usize,
    pub tracked_task_max_iterations: usize,
    pub enable_context_reduction: bool,
    pub enable_fallback: bool,
    pub always_include_categories: Vec<ContextCategory>,
    pub max_history_messages: usize,
    pub log_token_usage: bool,
    pub auto_compact_threshold: f64,
    pub compaction_strategy: CompactionStrategy,
    pub tool_result_max_chars: usize,
    pub tool_routing_strategy: ToolRoutingStrategy,
    pub reflection: ReflectionConfig,
}

Configuration for the pipeline

Fields

max_context_tokens: usize

Maximum tokens for context (model-dependent)

max_output_tokens: usize

Maximum output tokens

enable_tools: bool

Enable tool execution

iteration_budget_enabled: bool

Whether to enforce an iteration budget for agentic loops.

max_iterations: usize

Maximum agentic loop iterations for general requests when budgeting is enabled.

tracked_task_max_iterations: usize

Maximum agentic loop iterations for tracked-task requests when budgeting is enabled.

enable_context_reduction: bool

Enable context reduction

enable_fallback: bool

Enable fallback to secondary provider

always_include_categories: Vec<ContextCategory>

Categories to always include

max_history_messages: usize

Maximum number of history messages to include in prompt Older messages are dropped to save tokens

log_token_usage: bool

Log token usage before/after optimization

auto_compact_threshold: f64

Auto-compaction threshold (0.0-1.0) When estimated tokens exceed this percentage of the limit, automatically trigger context summarization. Default: 0.8 (80% of limit)

compaction_strategy: CompactionStrategy

Strategy to use when auto-compaction is triggered Default: CompactionStrategy::Summarize

tool_result_max_chars: usize

Maximum characters to include from a single tool result in the continuation prompt. Results larger than this are truncated with an indicator so the LLM knows content was omitted. Default: 8000 (up from the previous hard-coded 2000)

tool_routing_strategy: ToolRoutingStrategy

Strategy for selecting tools to expose to the LLM for each request.

• ToolRoutingStrategy::Keyword (default): uses the keyword-based RequestAnalyzer — fast, deterministic, zero extra latency.
• ToolRoutingStrategy::Llm: always fires a pre-flight LLM call to semantically select tools — highest accuracy, one extra round-trip.
• ToolRoutingStrategy::Hybrid: uses keywords when confidence is high; falls back to LLM routing when confidence is below the threshold.

reflection: ReflectionConfig

Configuration for ERL-inspired experiential reflection.

When enabled, the agent generates structured reflections on suboptimal turns and stores them for retrieval in future context injection.

Implementations

impl PipelineConfig

pub fn context_tokens_for_provider(provider: &str) -> usize

Get recommended context tokens for a specific provider (static fallback).

Prefer [for_model] or [for_model_with_cache] for dynamic, model-specific limits. This method only provides provider-level defaults and may be inaccurate for models with different context sizes (e.g., gpt-3.5-turbo vs gpt-4o).

pub fn for_provider(provider: &str) -> PipelineConfig

Create config optimized for a specific provider (static).

Prefer [for_model] for model-specific limits.

pub fn for_model(provider: &str, model_id: &str) -> PipelineConfig

Create config optimized for a specific provider and model using dynamic capabilities.

This uses heuristic-based model capabilities. For learned/cached capabilities, use [for_model_with_cache] instead.

pub fn for_model_with_cache( provider: &str, model_id: &str, cache: &ModelCapabilitiesCache, ) -> PipelineConfig

Create config for a model using a capabilities cache (for learned/API-discovered limits).

This method uses cached capabilities if available, falling back to heuristics. The cache can contain:

• API-discovered limits (most accurate)
• Error-learned limits (from context_length_exceeded errors)
• User-configured overrides

pub fn update_from_cache( &mut self, provider: &str, model_id: &str, cache: &ModelCapabilitiesCache, )

Update this config’s context limits from a capabilities cache.

Useful when you want to reconfigure an existing pipeline after learning new model limits (e.g., from a context_length_exceeded error).

Trait Implementations

impl Clone for PipelineConfig

fn clone(&self) -> PipelineConfig

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for PipelineConfig

fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error>

Formats the value using the given formatter.

impl Default for PipelineConfig

fn default() -> PipelineConfig

Returns the “default value” for a type.

impl PipelineConfigExt for PipelineConfig

fn with_user_settings(self, settings: &PipelineSettings) -> Self

Apply user settings from AppConfig.pipeline to this configuration.