Declarative template for creating nodes inside graphs. A `NodeTemplate` is a lightweight configuration object that can be reused across graphs.

Formatter for `KEY:\\nvalue` paragraph-style message payloads.

Top-level executable graph. RootGraph is the user-facing entry point for building and invoking workflows.

OpenAI model adapter using the Responses API only.

Raised when an external workflow document cannot be converted into a `Graph`.

Conversation history memory (list-of-dict message format).

Copy the raw LLM output to every required output field. This formatter intentionally does NOT attempt to parse, normalize, or extract any st

One directed edge between external node ids.

Semantic memory backed by embeddings and cosine similarity.

Strict JSON formatter with guarded fallbacks for malformed outputs.

Marker wrapper for NodeTemplate configs. Use this to explicitly share one dependency instance across node instances.

Raised when SKILL.md cannot be read or parsed into a valid skill definition.

One node discovered in an external workflow document.

Provider-agnostic text message block.

Best-effort JSON formatter tolerant to fenced blocks and noisy text.

One structured context unit to be injected into the user prompt.

Provider capability declaration used for early multimodal validation.

Directed message channel between two nodes. An `Edge` buffers at most one in-flight message at a time. It can be opened/closed via a gate, an

Normalized field declaration shared by Nodes, Edges and Agents.

Normalized DAG-ish description before materializing a MASFactory `Graph`.

Plain text formatter: clean prompts in, plain text out.

A hook stage with BEFORE/AFTER/ERROR keys derived from a stage name.

Factory-backed token usage accumulator.

Message formatter for CAMEL role-playing, supporting natural language output. This formatter is designed to handle flexible output formats fr

Adapter that exposes Python callables as LLM tool definitions.

Raised when the minimal WebSocket client encounters a protocol error.

Manages code files, including extraction, updates, rewriting, and version control.

Normalized query passed to context providers.

Fallback token counter using `cl100k_base`.

Manages document content, including parsing, updating, rewriting, and retrieving document content.

Manages document content, including parsing, updating, rewriting, and retrieving document content.

Tag-based stateful message formatter. Output format (no closing tags, no JSON wrapper): <key1>value1<key2>value2 Parsing rules:

Assign stable, collision-free tags to media assets within one request.

Manages code files, including extraction, updates, rewriting, and version control.

Compute concept coverage for text-generation solutions. This executor is primarily used for CommonGen-style tasks where the task description

Inject context as a single `CONTEXT` field in the user payload.

Composite graph node with internal entry and exit ports. Graph is both a node and a container. It can be nested inside another graph.

Simple in-process hook registry and dispatcher. Hooks are stored under an arbitrary `hook_key`. Callers register callbacks via `register()` a

Raised when a skill path exists but does not conform to the expected package layout.

OpenAI-compatible model adapter using Chat Completions only.

MCP-backed context provider. This adapter is intentionally lightweight: it delegates to a user-supplied callable that talks to an MCP server/

Provider-agnostic media message block recorded in history and encoded by adapters.

Embedder backed by `sentence-transformers`.

Very simple hashing-based embedder. This is intended for lightweight demos and tests, not for high-quality retrieval.

Raised when the target skill directory or its required SKILL.md file does not exist.

A small, stable facade for interacting with the `masfactory-visualizer` VS Code extension. Design goals: - Keep call-sites simple: `visu

LLM-driven agent node. The agent builds prompts from instructions, input fields, memories, retrievers, and output-field constraints, the

Anthropic token counter using Anthropic SDK.

A `TwinsFieldTextFormatter` that additionally extracts python code for selected fields. Default behavior extracts code for the `"solution"`

Collect context blocks from providers and inject them into the user payload.

Internal loop controller that decides continue vs terminate.

Default policy: provider order dominates; score sorts only within provider.

Options for compiling a Dify DSL export into an executable MASFactory graph. By default, `llm` nodes use a **real** `OpenAIModel` (see `dify_mode

Executable root graph for compiled Dify workflows.

Workflow edge definition.

Gemini token counter with SDK-first and local fallback behavior.

HuggingFace model inference adapter. This adapter handles inference calls to HuggingFace models through the Inference API, supporting mul

HuggingFace tokenizer-based counter for mapped model names.

Image media asset.

Local model server inference adapter. This adapter handles inference calls to locally deployed models through a local inference server en

A strongly-connected region compiled into one MASFactory `Loop` node.

Multi-agent mesh composite component with round-robin discussion. A composite component that creates a Loop with multiple agents taking turns

Model selector: checks model availability. This class manages checking model availability across different inference endpoints (HuggingFa

No-op executor that returns the solution unchanged.

OpenAI-family token counter based on `tiktoken`.