登录

metis

WholiverMIT活跃维护

Metis is a coding agent that boosts AI/LLM coding performance by 50% QQ群: 801193315

v2.0.2发布于 2026-10-16 04:59 UTC438star37fork同步于 2026-11-07 23:37 UTC
对比 GitHub 官网agent-orchestrationagent-skillsagentic-workflowai-agentsai-coding

怎么上手

README 里没有找到可直接复制的安装命令

README 里没有找到可直接复制的安装命令，请到 README 或官方文档查看安装方式。

硬件要求

README 没有说明硬件要求。

下载：源码 zip·GitHub Releases（安装包）本站不镜像安装包

项目说明

这个项目的说明还在生成中，先参考下面「最近发版」和 README。

最近发版

5 个版本v2.0.2
全部 release
  1. v2.0.2Metis v2.0.2最新正式版2026-10-16 · 2026-10-16 04:59 UTC

    Metis v2.0.2

    🚀 Key Improvements & Fixes
    • Windows Large Repo Git Diff Protection (Fixes #45): Optimize Git diff parsing and branch comparisons in workspace-git.cjs using chunked asynchronous processing with timeout safeguards, avoiding process deadlocks and abort errors on large Windows repositories.
    • Session List Deduplication (Fixes #45): Standardize cross-platform session path comparisons and keying in src/core/session-manager.ts and sidebar state to completely eliminate duplicate conversation entries across Windows drive letters and relative paths.
    • Model Selector & Custom Model Polish (Fixes #47):
      • Expand model dropdown width and add full-name tooltips to prevent long model names from truncating.
      • Pin custom models to the top without redundant Custom prefix labels.
      • Unfold models by provider so multiple models from the same vendor are cleanly presented.
      • Add one-click deletion for custom models directly in Settings.
    • Remote Workspace & Server Mode Git Inspection (Fixes #44):
      • Add workspace service endpoints (/workspace/tree, /workspace/git-status, /workspace/git-diff, /workspace/file) to Metis Server mode.
      • Add RemotePathModal and remote workspace bridge for Desktop clients connecting to remote servers, enabling remote file picking, directory tree exploration, and Git review.
  2. v2.0.1Metis v2.0.12026-10-15 · 2026-10-15 16:02 UTC
    查看更新说明

    Global & Component Error Boundaries, Starter Prompts, Process Crash Recovery, Large File Diff Protection & Model Catalog 1.2.0

    • Global & Component Error Boundaries: Add top-level GlobalErrorBoundary and modular ComponentErrorBoundary layers to catch rendering faults in transcript, inspector, and chat views with localized error details, crash isolation, and one-click session reload and state reset affordances.
    • Desktop Process Lifecycle & Crash Recovery: Add Electron render-process-gone crash handler with interactive reload dialog, child process monitoring, hardware acceleration toggles (METIS_DISABLE_GPU, --disable-gpu), and standard developer shortcuts (Cmd+R / Ctrl+R / F5 for reload, Cmd+Alt+I / Ctrl+Shift+I / F12 for detached DevTools).
    • Interactive Quick Starter Prompts: Introduce StarterPrompts pills on empty conversation sessions to trigger one-click exploratory tasks (architecture analysis, bug finding, test design, refactoring) tailored to the active workspace.
    • Large & Binary File VCS Diff Protection: Filter out binary assets and files exceeding 300KB during working tree and branch diff scanning in workspace-git.cjs to eliminate UI stutter, high memory allocations, and diff panel freezes.
    • Model Catalog 1.2.0 & Enhanced Reasoning Protocols: Synchronize upstream model catalog to pi-ai 1.2.0; expand dynamic thinking/reasoning protocol detection to Step (step-3.5+), Codex, Astra/Luna/Sol, Qwen, DeepSeek, GLM-5.3 variants, and OpenAI-compatible aggregators with robust compatibility merging.
    • Desktop i18n & Add Model Polish: Update English, Chinese, and multi-language catalogs with localized error boundary prompts and refine model configuration dialogs.
  3. v2.0.0Metis v2.0.02026-10-14 · 2026-10-14 03:13 UTC
    查看更新说明

    Project Agent Harness Architecture, Visual Graph Editor, Sidebar Code Tree & Runtime Isolation

    • Project Agent Harness Architecture (.metis/harness.json): Persist and enforce project-level multi-agent architectures with immutable run-start snapshots, structural validation, and refusal on corruption (no silent fallback).
    • Interactive Agent Harness Graph Editor: Visual canvas for designing execution tiers, role assignments, dependencies, and declarative quality gates directly in Desktop.
    • Declarative Quality Gates: Run automated verification checks (command, expectExit, artifactPath) with configurable enforcement policies (block or loop-to:<nodeId> backtracking).
    • Sidebar Code Tree & Local Git Workspace: Interactive directory explorer integrated into the sidebar, featuring file Git status badges (M/A/D/U), unified diff preview, branch tracking, search filters, and editor shortcuts.
    • Universal Provider Session & Request Isolation: Inject X-Session-Id and unique X-Request-Id headers across all model providers and custom proxies (including Tierflow, LiteLLM, OneAPI) to prevent proxy-level cross-session KV cache contamination.
    • Uncapped Subagent Delegation: Set default subagent spawn capacity limits (DEFAULT_MAX_CHILDREN_PER_AGENT and DEFAULT_MAX_TOTAL_CHILDREN) to unlimited (0) with CLI --max-children 0 support.
    • Quota Message Sanitization: Prohibit models from surfacing internal quota exhaustion details and automatically filter historical quota notices in frontend rendering.
    • Self-Learning Fast Loop & Badges: Comprehensive runtime adaptation compiler, preference engines, and visual Self-Learning badges across the Desktop interface.
    • Multi-Language Parity & Onboarding: Complete 11-language catalog synchronization (1,820 keys each) and polished onboarding transition experiences.
  4. v1.4.9Metis v1.4.92026-10-09 · 2026-10-09 08:30 UTC
    查看更新说明

    Universal Session & Request Isolation, Subagent Uncapped Scaling & Quota Sanitization

    • Universal Provider Session & Request Isolation: Inject X-Session-Id and X-Request-Id headers across all model providers and custom proxies (including Tierflow, LiteLLM, OneAPI) to prevent proxy-level cross-session KV cache contamination.
    • Uncapped Subagent Delegation: Set default subagent spawn capacity limits (DEFAULT_MAX_CHILDREN_PER_AGENT and DEFAULT_MAX_TOTAL_CHILDREN) to unlimited (0) and allow CLI --max-children 0, enabling complex multi-agent execution graphs without artificial limits.
    • Quota Message Sanitization: Prohibit models from surfacing internal subagent quota exhaustion details to end users, and filter out historical quota exhaustion complaints during frontend transcript rendering to keep the conversation experience clean and distraction-free.
  5. v1.4.7Metis v1.4.72026-10-07 · 2026-10-07 07:08 UTC
    查看更新说明

    Scheduled Routines, Plugin Contracts, Self-Learning Fast Loop, Length-Stop Compaction & Desktop UX

    • Context Window Length-Stop Auto-Compaction & Truncation Handling: Accurately detect context limit exhaustion (stopReason: "length" with >= 98% context filled and empty/minimal tokens) from reasoning models like DeepSeek-V4.1-Flash; trigger auto-compaction and pop trailing assistant messages before retrying so continuation proceeds seamlessly; treat empty length-stop turns as clear failure notices with retry affordances rather than displaying stale work text.
    • Desktop Conversation Turn Deduplication: Fix turn duplication in useMetisServer.ts during snapshot reconciliation (adoptSnapshotWithoutRegressing) when snapshot IDs differ from synthesized SSE IDs after abort or session reload.
    • Strict Language Consistency Enforcement: Enforce user language match across Plan/Build collaboration mode system prompts and milestone progress narration, strictly forbidding spontaneous switches to English when the user's latest prompt is Chinese.
    • Scheduled Routines System: Introduce the background RoutineHost scheduler and model-facing routine tool (list, create, update, delete, toggle, run), supporting standard 5-field cron expressions, interval scheduling, persistent execution status, and an interactive Routines management panel in Desktop Settings & Sidebar.
    • Plugin Architecture & Strict Contract Checking: Add the metis-plugin CLI (bin/metis-plugin.js) and contract validation suite (scripts/check-contracts.mjs, packages/) enforcing strict zero-drift contracts across BUILTIN roles, 16 evaluation frameworks, performance gates, and route providers.
    • Self-Learning Fast Loop & Adaptations Acceleration: Expand runtime adaptations with the Prompt Compiler (compiler.ts), Architecture Rule Engine (architecture-engine.ts), Preference Engine (preference-engine.ts), and Schedule Learner (schedule-learner.ts) to converge multi-turn adaptations rapidly while preserving clean baseline execution when inactive.
    • Multi-Agent Isolation & Subagent Experience: Strengthen recursive delegation boundaries, session barriers, and spawn guard protections; improve Desktop Inspector Subagents list and live transcript rendering with real-time status and graceful error handling.
    • Desktop UI & Localization Polish: Fully update and regenerate bilingual i18n catalogs (English & Chinese); refine Settings dialog layout and metric tiles; improve thinking block streaming lifecycle and plan progress point narration.

用户点评

登录用户的真实使用记录，可按版本和硬件筛选
共 0 条

    还没有人点评。第一个说出真实体验的人最有价值。

    相关项目

    同分类里 star 最高的项目
    对比全部 →
    ECC
    affaan-m·JavaScript

    为 Claude Code、Codex、Cursor 等编码智能体提供技能、直觉、记忆与安全防护的性能优化层，让 AI 编程助手更稳定高效地干活。

    智能体框架+4Agent 框架
    300641+191444kv2.2.32026-11-09 04:56 UTC
    deepseek-harness
    deepseek-ai·TypeScript

    DeepSeek 开源的 Agent 编排框架，把模型、工具、记忆等所有能力都做成可插拔插件，方便按需拼装和扩展智能体。

    Agent 框架+3Agent 框架
    278174+297434kdsh-v0.2.2-rc.12026-11-04 13:06 UTC
    hermes-agent
    NousResearch·Python

    一个可随使用不断成长的 AI Agent 框架，支持接入 Claude、ChatGPT 等多种大模型，用于搭建和编排自主智能体。

    Agent 框架+3Agent 框架
    262435+115559kv0.23.02026-11-10 23:17 UTC
    firecrawl
    firecrawl·TypeScript

    把任意网页抓取并转成干净 Markdown 或结构化数据，直接喂给 LLM 和 AI Agent，省去自己写爬虫和清洗的麻烦。

    网页抓取+4Agent 框架
    200517+117811kv2.11.02026-11-10 19:28 UTC