xiaohou521/fusion-moa

README 摘要

Fusion MoA 简体中文 Fusion MoA is a model-, GPU-, and harness-independent Mixture-of-Agents runtime for coding agents. Bring a main model and an optional pool of read-only experts, configure them in one recipe, and expose one stable API to Codex, Claude Code, OpenCode, DeepSeek Harness, or another client. Status The v0.1 contract includes: - strict fusion/v1 recipes with cross-reference and secret validation; - declared model capabilities and per-model concurrency limits; - OpenAI-compatible, llama.cpp, and Anthropic-compatible providers; - direct , main-critic , and parallel review-board policies; - OpenAI Chat, OpenAI Responses, and Anthropic Messages endpoints; - portable function/tool-call round trips and /v1/models discovery; - native final-model SSE for all three public protocols; - Python entry points for third-party providers and policies; - an evaluation-gated promotion command for controlled RSI; - a version-pinned DeepSeek Harness profile bundle. Native final-model streaming means expert orchestration completes first, then text and tool-call deltas from the one authoritative main-model call are forwarded without buffering the full answer. Expert output is never exposed as th…

分类