跳到主要内容
    ↑↓ 选择↵ 打开esc 关闭
    中文English
    vocino

    Autonomy

    v0.6.3模型接入
    opencode-autonomy

    Autonomy-first opencode suite — ships features end-to-end without babysitting. Agents, commands, and opinionated config for long-horizon coding.

    GitHub 星标

    4

    近 30 天 +1

    月装机量

    1,672

    近 7 天 205

    综合评分SCORE

    47.5

    生态多维模型

    最近提交

    7 天前

    2026-08-12

    快速安装与配置

    opencode.json

    写入当前项目的 opencode.json,只对这个仓库生效。

    opencode.json

    {
      "$schema": "https://opencode.ai/config.json",
      "plugin": ["opencode-autonomy@0.6.3"]
    }

    opencode 启动时会通过内嵌运行时自动加载 npm 依赖并缓存至本地目录,无需手动在全局环境执行安装。

    Actually ships. One command, no babysitting — concept to verified code.

    Problem

    Most opencode configs are:

    • too chatty — asks to run npm install
    • too fragile — no verification, breaks tests
    • too single-model — same blind spots
    • too complex — 15 agents, no one knows what runs

    You end up cleaning up after it.

    How to install

    Copy-paste prompt — keeps your models

    Paste this into opencode (preferred — it understands your current config):

    Install the opencode-autonomy plugin from https://github.com/vocino/opencode-autonomy — keep my existing models, providers, and apiKeys in opencode.json. Only set model/small_model/agent model if I don't already have one. Merge autonomy agents (build, build-worker, fixer, explore, plan, council-critic, council-creative) without overwriting my custom agent models. Add parallel build-worker lane support (same model as build). Preserve my provider baseURL/apiKey. Verify with opencode debug config.
    

    You still get the power of different models (build + workers share 1M, fixer is sonar-ish, explore is qwen), but if you already pinned build to claude-sonnet or your own meta key, it stays yours. New users get the 5-model defaults.

    Shell install — also preserves models

    Same merge logic now lives in the CLI:

    opencode plugin opencode-autonomy --global
    npx opencode-autonomy@latest --clean   # copies agents/ + merges opencode.json without nuking your models
    

    Verify:

    opencode debug config
    agents list   # should show build, build-worker, fixer, explore, plan
    

    If you want our defaults, just delete ~/.config/opencode/opencode.json and re-run npx — you'll get the full 5-model suite.

    How to use

    Ship a feature

    /ship Add dark mode toggle, persist in localStorage
    

    Scans repo, plans 3-5 file batches, runs lint → typecheck → test → build, hands failures to @fixer, loops until green.

    Fix

    /fix Settings crashes when email empty
    

    Parallel search with @explore, patches, reruns checks.

    Explore

    @explore Find all API endpoints touching user creation
    

    Tab to plan = read-only, asks first. Good for dry runs.

    Autonomy & Permissions — READ THIS

    This is allow-all autonomy. Be aware.

    permission: {"*":"allow", "external_directory":"allow", "doom_loop":"allow"}
    batch_tool, 300 step build, 150 step fixer, 3x same-error stop
    5000 lines / 200KB logs, tail 12, subagent_depth 3
    

    ALLOW = allow all with "*":"allow" — full disk + bash + external dirs. Adds plugin: ["opencode-autonomy"] at runtime, keeps your model/provider.

    What happens:

    • edits, npm install, git, rm without asking
    • batches 3-5 files then verifies
    • no "should I continue?" — that's the point

    Undo: npx opencode-autonomy --disable restores backup of opencode.json.

    If that's not you, use plan agent or don't install.

    What's inside

    • opencode.json.example — full config, readable in 5 minutes
    • commands/ship.md — the whole loop, parallel-aware
    • agents/build.md — orchestrator, parallel lane detection + fan-out
    • agents/build-worker.md — same-model workers, one per lane
    • agents/fixer.md, agents/explore.md, council agents — minimal set
    • src/autonomy.ts — single source of truth for forced keys
    • src/plugin.ts — v1 config hook, preserves your model/provider
    • bin/cli.mjs — zero-dep npx installer

    Why 5 models, 5 families (build + workers share meta):

    • meta/muse-spark-1.2-contributor — build orchestrator + build-worker lanes, 1M, 80% of work
    • openrouter/google/gemini-flash-latest — titles
    • openrouter/anthropic/claude-sonnet-4-5 — fixer
    • openrouter/qwen/qwen3-coder — explore
    • openrouter/openai/gpt-4o-mini — plan

    Two keys, different blind spots. Built on CachyOS / Arch gaming box — small and verifiable.

    Install modes

    # everywhere (recommended)
    opencode plugin opencode-autonomy --global
    
    # this repo only
    opencode plugin opencode-autonomy
    
    # pin version
    opencode plugin opencode-autonomy@0.4.0 --global -f
    
    # local markdown you can edit
    npx opencode-autonomy@latest --clean
    

    Updates: opencode plugin opencode-autonomy@latest --global -f

    How it works

    /ship "goal"
      -> Concept: read repo + git status
      -> Plan: TodoWrite if 3+ steps + annotate parallel lanes (disjoint file sets)
      -> Implement: parallel via @build-worker (same model, 2-3x faster) OR sequential batch 3-5 files
      -> Merge: collect lane summaries (no file overlap)
      -> Verify: detect-oracle.sh → lint/type/test/build
      -> Fix: @fixer per failing lane, parallel if disjoint
      -> Ship: report + parallelism used + commit msg
    

    detect-oracle.sh finds your checks from package.json.

    Ecosystem

    • opencode-ambient — see tokens and money as light in your room
    • Install both: autonomy drives, ambient glows

    Using opencode on Arch Linux.

    Development

    git clone https://github.com/vocino/opencode-autonomy.git
    npm install && npm run build && npm test
    

    License

    MIT