Mock arguments for agent execution

A workflow node.

YAML task parser

A typed workflow variable used in reads/writes/parameters.

Parses LLM responses to extract different types of tool calls. This parser handles three types of tool invocations: 1. Standard OpenAI tool_c

YAML-native Agent that executes YAML workflow definitions. This class is the main entry point for running YAML-defined workflows. It coordina

Thread-safe immutable wrapper for execution arguments with per-file config support. This class provides a way to pass execution configuratio

Thin wrapper around fastmcp Client for this codebase. Durable execution note: - Sandbox tools require `ctx.session_id` which is ultimate

Parsed LLM response with extracted tool calls. Attributes: content: The text content of the LLM response. tool_calls: List of Ope

Unified LLM client using LiteLLM.

A local tool available to the orchestrator LLM. These run in-process (not via MCP) and are presented to the LLM alongside MCP tools in the sa

Result of processing a single SWE-Bench instance.

A submodule invoked via call / parallel / gather. Records only enough about the submodule to produce human-readable diagnostics; the verifier

Result of plan verification.

Result of processing a single AIME problem.

Simple mock for tool function objects

Result of processing a single ELAIPBench question.

Result of a fuzzy match search.

Result of processing a single WritingBench query.

Reactive agentic loop with MCP tool access and task management. The core while(True) loop: call the LLM with MCP + orchestrator tools, execut

Computed effective tools and submodules for a specific context and step. This class encapsulates the logic for determining which tools and submod

Main interpreter for executing YAML workflow steps. This class provides the dispatch logic for executing workflow steps, delegating actu

Configuration for LLM client.

Executes LLM interactions with multi-step tool calling support. This class is the core execution engine for LLM-driven workflow steps. It man

Parses structured log events.

Configuration for agentic loop sessions.

Result of executing a tool. Encapsulates the outcome of a tool execution, including success/failure status, output data, and any error inform

Arguments for the YAML agent.

Represents an agent iteration.

MCP-compatible client that routes tool calls through `docker exec`. Subclasses MCPClient with connect=False to skip HTTP connection, then ove

Structured result from plan execution. Provides both the output text and metadata about execution health, enabling the orchestrator to detect

Parsed log event.

Represents a single change block in a patch.

Bridge between the agentic loop and the AgentSPEX workflow engine. Given a verified YAML plan file, loads it, creates a fresh Interpreter, an

Generate AgentSPEX YAML workflow plans for a given task. Uses a (typically cheap) LLM to produce YAML, validates it parses, and writes it to

Semantic verifier for AgentSPEX YAML workflow plans. Performs: 1. YAML syntax check (via yaml.safe_load inside YAMLTaskParser) 2. Require

Tracks one agent run.

Manages orchestrator conversation history and usage tracking.

Represents a workflow step.

In-memory task board with JSON persistence. Provides CRUD operations for tasks, a revision log for auditability, and atomic JSON persistence

Represents the entire workflow.

Visualize YAML workflows as flowcharts using Mermaid.js

YAML task parser

Highlights /command at the start of input in yellow.

Mixin providing basic step and task step execution. These are the most common step types, used for direct LLM interactions with optional tool

Standard keys used in workflow context dictionaries. These constants provide type-safe access to context dictionary entries, preventing typos

Mixin providing control flow steps: for_each, if/else, while, switch. These steps enable complex workflow patterns by allowing iteration, con