Deterministic accept/reject decision for a candidate scoring result.

Singleton config loader that caches configuration files.

A collection of factories. The most general form of supported factory is an AsyncContextManager, meaning we support factories that are async

Normalized view of a scoring/results.json payload.

Generates research prompts from templates and idea specifications. Uses Jinja2 templating to inject idea-specific content into structure

Structured, persistent model of the research-in-progress (v3).

Manages research idea submissions and tracking. Handles validation, storage, and status updates for research ideas.

A ManagedEnv unifies Providers under a single unbrella with a cache of objects constructed by these providers. It also manages a resolver tha

Manages AutoResearch node checkpoints inside a workspace Git repository. If the workspace is not already a Git repository, this class initia

Manages GitHub operations for research projects. Requires GITHUB_TOKEN environment variable to be set.

A tool call parsed from the LLM response.

Stores AutoResearch attempt history under a NeuriCo logs directory.

Parsed response from the LLM.

Orchestrates multi-agent research pipeline. Pipeline stages: 1. resource_finder: Gather papers, datasets, code (CLI agent) 2. (optio

Original CLI behavior: print() / input() on the host terminal.

ApplicationContext is an extension of the ManagedEnv concept. In addition to the existing functionality on ManagedEnv, the Application contex

Result for one AutoResearch candidate attempt.

A Git-backed AutoResearch node.

Summary of a phase-1 AutoResearch initial-node construction.

A utility class to inherit from that exposes a number of convenience methods around the resolver. - Injection into an async function

Raised when a trim decision fails schema/cross/lint validation.