Configuration for a single MCP server

Raised when batch processing fails

Manages MCP configuration loading and validation

Represents a connection to an MCP server

A ClientSession that logs all communication

Automatic request batching for OptILLM Collects incoming requests into batches and processes them together for improved throughput.

Represents a solution attempt by an agent

Logger for OptiLLM conversations including all provider interactions and metadata. Logs are saved in JSONL format (one JSON object per line)

Manages MCP servers and capabilities

Enhanced mock OpenAI client for MARS testing

Represents a problem-solving strategy learned by the system.

Object for logging the number of LLM calls and tokens used in the pipeline

Individual agent for mathematical reasoning with OpenRouter reasoning API

MLX-based inference pipeline that mirrors PyTorch pipeline interface

Dynamically handles both approaches and plugins

Calculates various confidence metrics for token-level assessment.

Calibrates confidence thresholds based on warmup traces.

Configuration for MLX models

Enhanced mock OpenAI client for IMO25 testing

Mock OpenAI API response

OpenAI-compatible client that proxies to multiple providers

Represents a verification attempt result

Universal answer extractor using math-verify with fallback patterns

Main AutoThink processor class for external use. Wraps the internal processor implementation.

Container for a single request in a batch

Manages a single browser session across multiple searches. Implements context manager for automatic cleanup.

Classifies queries as HIGH or LOW complexity for token budget allocation. Uses the adaptive-classifier model for classification.

Represents a single conversation entry being logged

Main DeepConf processor implementing online mode with early termination.

Wrapper that adds extended timeout support for different client types

Simplified implementation of Test-Time Diffusion Deep Researcher (TTD-DR) algorithm This class implements the core concepts from the TTD-DR

Background health checker for providers

OpenAI SDK Compatible client for local inference with dynamic model support

LoRA manager with enhanced error handling and caching

Handles calculation of log probabilities for generated tokens

Container for logprobs calculation results

RSA-inspired aggregation system for combining solutions Key features: - Population management (N > K for diversity) - Recursive aggr

Multi-pass verification system inspired by IMO25 solver

Shared workspace for agent collaboration and solution tracking

Manager for MLX models and operations

Memory-efficient attention using linear attention mechanism. Supports automatic fallback to optimized implementations when available.

Mock client that simulates degraded provider responses. mode: - "empty_choices": response with an empty choices list - "none_content"

Mock OpenAI client for testing

Mock OpenAI client for testing

Wrapper for a provider configuration and client

Represents an extracted reasoning strategy from an agent solution

Thread-safe session state manager for deep research. Ensures only one browser session is active per research query.

Implementation of the SELF-DISCOVER framework. The framework operates in two stages: 1. Stage 1: Discover task-specific reasoning struct

Main evaluator class for SimpleQA benchmark

Hook for applying steering vectors during generation.

Manager for loading and applying steering vectors from a dataset.