@fileoverview Core conversation loop engine for open-multi-agent. {@link AgentRunner} is the heart of the framework. It handles: - Sending messages to the LLM adapter - Extracting

mnemon.moth ----------- Auto-instrumentation engine. On mnemon.init(), the moth wakes up, scans the environment for installed agent frameworks, and patches each one at the right ex

Mnemon -- execution caching and system learning for AI agents. Zero code changes. One import line. Gets smarter with every run. Saves tokens on every cache hit. Usage: import mnemo

Builds typed execution receipts with token usage, validating schema; guards used for receipt creation.

Builds a case-insensitive regex pattern matching @mentions of a name for triggering agent invocation.

waterfall.ts Core waterfall logic for the AI gateway. Given a standard OpenAI ChatCompletion request, tries each configured provider in priority order until one responds successful

Executes evaluation test cases against targets with configurable concurrency, scoring, progress tracking, and cost/token metrics collection.

Default session key used for chat conversations. This value is used when no custom session key has been configured. It follows the format: agent:<agentId>:<sessionName> /

Renders 3D scene component "game" = "frontend" "room" "objects"di "space graph", "displayed via 1. Perhaps a "3D" main score 0.

AI browser/agents. Manages web page state. Persisted at end.

Factory function that builds and configures a Flask app for AgentScope, including secret key generation.

Provides async browser automation using Playwright, supporting page interactions, authentication, and AI agent workflows.

claude_node.exceptions — Typed exception hierarchy for Claude CLI controller. All exceptions inherit from ClaudeError (which inherits from RuntimeError), allowing callers to catch

@fileoverview {@link AISdkAdapter} — bridge to the Vercel AI SDK (`ai` + `@ai-sdk/*`). When {@link AgentConfig.adapter} is set, {@link Agent} skips {@link createAdapter} and uses t

@fileoverview Shared OpenAI wire-format conversion helpers. Both the OpenAI and Copilot adapters use the OpenAI Chat Completions API format. This module contains the common convers

@fileoverview {@link AgentBackend} that drives an external coding agent over the Agent Client Protocol (ACP). OMA takes the ACP **client** role: it spawns the agent (a local CLI su

Normalizes agent skill IDs (dashes/underscores) → canonical form for consistent LLM tool dispatch & skill management.

Defines provider configs, model profiles, rate limits, tool-calling options for LLM routing.

Protocol bridge converting Anthropic API messages and tool calls to OpenAI format for cross-API compatibility.

Agent: unified agent class

Defines the Bento class for packaging ML models with metadata and artifacts.

Constructs a main coding prompt using dedented template with shadcn documentation context for frontend development guidance.

Manages] F: validates

Parses SSE stream data from multiple LLM providers (Anthropic, OpenAI), extracting and accumulating text deltas from streaming responses.

Abstract base class defining the interface and core attributes (role, goal, backstory, tools, cache, execution parameters) for all agents in the CrewAI framework.

Defines protocol constants and command enums for an AI assistant SDK, enabling communication between AI models and Android device capabilities like camera, SMS, location, and notif

eslint-disable @typescript-eslint/no-explicit-any */ * TypeScript wrapper for the GeminiNano native Expo module. At build time the native module is linked automatically by Expo. We

Shared helper for reading Claude Agent SDK message streams. Both the judge and Monte Carlo layers consume the same `query()` message stream (assistant text blocks + a terminal resu

Normalizes AI model configuration data for provider-specific usage.

compressors.py — Four compression strategies for the Engram Benchmark. Strategies: 1. EngramCompressor — LLM Observer + Reflector (Layer 6 Engram) 2. RuleCompressor — claw-compacto

Mnemon moth — LangChain integration. System 2 EME via per-step patching: RunnableSequence.invoke is patched to iterate each step individually. Each step is a segment. If a step's o

Lightweight OpenAI-compatible LLM client — stdlib only. Features: - Model fallback chain (gpt-5.2 → gpt-5.1 → gpt-4.1 → gpt-4o) - Auto-detect max_tokens vs max_completion_tokens pe

Resolves available agents via CLI discovery and path resolution; returns list for agent orchestration.

Renders a terminal dashboard for a captured OpenMultiAgent run.

StaticEncoder encodes text into semantic embeddings for vector search via model2vec_rs.

Builds LLM judge scorer: constructs prompts, parses verdicts, manages aborts/deadlines for eval.

Mnemon Execution Memory Engine (EME) — v2 Generalised plan cache for any expensive recurring computation. System 1: exact fingerprint match — zero LLM, sub-millisecond System 2: pa

Model-name → tokenizer dispatch. ! ! Mirrors `MODEL_PATTERNS` in `headroom/tokenizers/registry.py`. Three ! backends in priority order: ! ! 1. **HuggingFace** — anything the caller

AI Search Agent Executor Runs an autonomous agentic loop using Vercel AI SDK's generateText(). The agent can call search tools (search_web, read_url, parallel_read_url) to answer t

LLM execution engine orchestrating RAG pipelines and prompt flows via event bus for AI agents.

Intelligent Query Repair with Learning and Retries Strategies: 1. Check memory for known fix 2. Apply Python-level validation 3. Get LLM to fix with detailed context 4. Retry with

Mnemon Pre-Warmed Fragment Library 100 curated execution fragments for common LLM agent task patterns. Ships with Mnemon — provides immediate EME value on day one. 100 fragments ac

Computes Anthropic vision token usage for an image with given dimensions and tier, resizing via binary search to respect long-edge and token caps.

Manages in-memory conversation history with add, delete, query, save/load, token-aware truncation, and dynamic context windowing for agent interactions.

Defines text-chunking strategies for NLP pipelines, implementing regex splitting, sliding-window, and overlap segmentation to segment documents into manageable units.

Immutable provider registry that normalizes names, rejects duplicates, maps model aliases to provider candidates, and exposes deterministic model/provider metadata.

Parses malformed JSON strings via tolerant fallback, extracting literals/partials from LLM outputs.

MCTS search node; parses LLM state JSON, tracks value/visits/parent/children, enables selection, expansion, rollout, backup in tree search.

Builds AlphaPose model from config, assembling model, loss, dataset loaders, and training components.

Load and manage skill manifests for modular supervisor routing.