← Back
petershifi123-wq

petershifi123-wq/solve-lite

CoreML-native local decision runtime for AI agents · 18 install profiles · 20-scenario protocol · offline · ~50.93 MB · by Soulite Magic

View on GitHub ↗
Stars
85
Forks
0
Watchers
85
Open issues
0
Contributors
2
Language
Python
License
Other
Default branch
main
Created Sep 25, 2026Updated Sep 28, 2026

Star growth

Today—
This week—
This month—

Star history will appear here once this repo has been tracked for a couple of days.

README

Solve Lite v0.1.11

Local decision runtime for AI agents on macOS arm64.

Current runtime

Compact Runtime — 82.28% accuracy · APS 85.8999 · 6 high-confidence errors · 50.93 MB complete install · CoreML-native · offline after install.

This is the recommended and only current install target. It uses one shared BERT L4/H256 FP32 CoreML encoder and four tiny calibrated heads for Review, Topic, NLI and Financial routes. The registry-driven installer declares 18 thin host profiles that point to the same versioned runtime. Runtime network calls are zero, and neither PyTorch nor Transformers is shipped or required.

The current NLI compact-precision debt is owner-accepted and non-blocking.

Compatibility advantage

  • One registry-driven installer covers 18 mainstream agent host profiles instead of shipping 18 model copies.
  • Native hooks are used only where machine evidence proves them; every other host receives the same portable pre-prompt entrypoint.
  • All hosts share one hash-pinned CoreML runtime and the same 20-scenario percentage/reward contract.
  • Offline after install, zero runtime network calls, no PyTorch/Transformers/ONNX dependency, and no model-path setup.

AI Agent Quick Install

If the user asks you to install Solve Lite from this repository:

  1. Use the current release only.
  2. Do not ask the user for model files, model paths, ML environments or optional downloads.
  3. Install automatically:
git clone --depth 1 --branch v0.1.11 https://github.com/petershifi123-wq/solve-lite.git
cd solve-lite
/usr/bin/python3 tools/installer.py --host auto
  1. Verify the installed copy:
/usr/bin/python3 tools/installer.py --check --json
  1. If verification passes, activate the installed host adapter and continue normally. Use --host <id> when automatic detection is ambiguous; use --host all only for an explicit multi-host installation.

如果你是 AI Agent：用户只需要告诉你“安装 Solve Lite”。你应自行识别宿主、下载安装当前共享运行时并完成激活，不要让用户配置模型路径或 Python/ML 环境。

Install

Download the v0.1.11 release archive, verify it with the release SHA256SUMS.txt, extract it, then run:

python3 tools/installer.py --host auto

The installer downloads exactly one hash-pinned shared asset from the immutable Hugging Face revision recorded in plugins/solve-lite/skills/solve-lite/assets/specialist-assets.json. Installation verifies both the archive SHA-256 and the unpacked tree SHA-256. The resulting runtime lives once under the versioned Solve Lite application-support directory; host adapters contain no model copy.

Offline installation is also deterministic:

python3 tools/installer.py --offline --package-dir /path/to/verified-package-directory --host auto

Read-only verification:

python3 tools/installer.py --check --json
python3 tools/startup_check.py --root . --json

Runtime tiers

  • Compact Runtime (current, recommended): 82.28% accuracy, APS 85.8999, 50.93 MB complete install, CoreML-native.
  • Historical Full Runtime (reference only): 92.53% accuracy. It is not an install target, default manifest, fallback, or prerequisite.
  • Jev reference: 53.67%.

These are different evaluation/runtime tiers. The historical reference is retained for provenance only.

Shared runtime contract

  • one shared encoder/runtime copy per machine;
  • four tiny task heads: Review, Topic, NLI and Financial;
  • OS-provided CoreML plus the shipped arm64 helper;
  • zero network access during health checks and inference;
  • no PyTorch, Transformers or ONNX runtime;
  • no model-root configuration;
  • pinned HTTPS download, archive hash, installed-tree hash and immutable revision;
  • structured failure when the shared asset is missing or damaged.

Host adapters

The host registry is plugins/solve-lite/skills/solve-lite/assets/agent_registry.json.

  • 18 declared profiles: Claude Code, Cursor, GitHub Copilot, Codex, Windsurf, Gemini CLI, Amp, Goose, Aider, Cline, Roo Code, Trae, Hermes, OpenCode, ChatGPT, WorkBuddy, Doubao and Qwen Code.
  • Every profile receives only a thin skill/plugin plus a pointer to the single shared runtime.
  • PASS, PARTIAL and NOT_RUN remain machine-evidence states. A declared install profile is not silently promoted to a verified product PASS.

INSTALL_PASS != PRODUCT_PASS. All 18 profiles pass isolated installation in this release; only four retain prior ordinary-session evidence. Every other status stays PARTIAL or NOT_RUN until a real host session proves installation, routing, percentages, one reward settlement and restart persistence.

Host Install profile Isolated install Ordinary-session evidence Product status
Claude Code YES PASS NOT_RUN NOT_RUN
Cursor YES PASS NOT_RUN NOT_RUN
GitHub Copilot YES PASS NOT_RUN NOT_RUN
Codex YES PASS PRIOR PASS PASS (prior evidence)
Windsurf YES PASS NOT_RUN NOT_RUN
Gemini CLI YES PASS NOT_RUN NOT_RUN
Amp YES PASS NOT_RUN NOT_RUN
Goose YES PASS NOT_RUN NOT_RUN
Aider YES PASS NOT_RUN NOT_RUN
Cline YES PASS PARTIAL PARTIAL
Roo Code YES PASS NOT_RUN NOT_RUN
Trae YES PASS NOT_RUN NOT_RUN
Hermes YES PASS PRIOR PASS PASS (prior evidence)
OpenCode YES PASS NOT_RUN NOT_RUN
ChatGPT YES PASS NOT_RUN NOT_RUN
WorkBuddy YES PASS PRIOR PASS PASS (prior evidence)
Doubao YES PASS PRIOR PASS PASS (prior evidence)
Qwen Code YES PASS PARTIAL PARTIAL

Every non-empty ordinary message must run through the frozen 20-scenario router. A successful turn must expose its selected scenario percentage and settle the local reward pool exactly once. A failed decision surface must display PRESENTATION_DATA_UNAVAILABLE; it must not fabricate percentages. Token savings are shown only when a real pack step was measured.

System, assistant and tool events, hook-internal traffic, empty messages and retry/internal replays are excluded. Stable turn identifiers make reward settlement idempotent; audit and rendering never create a second reward event.

Every adapter stores only code and a pointer receipt. The model and native helper remain in the one shared runtime.

Release status

v0.1.11 is the current public install target. Earlier releases remain historical records and are marked superseded; do not install them.

Peter fresh-host acceptance remains a human gate and is not substituted by automated tests.

Project and support

  • Creator: Andrew Yang
  • Company: Soulite Magic｜灵智天成
  • Douyin / 抖音: UPAI - 升级你的 AI 能力
  • Commercial licensing: contact Soulite Magic.
  • If Solve Lite helps you, you can support continued development through the Alipay QR below.

Alipay donation QR

中文

当前运行时

Compact Runtime — 准确率 82.28% · APS 85.8999 · 高置信错误 6 · 完整安装 50.93 MB · CoreML 原生 · 安装后完全离线。

这是唯一当前安装目标，也是推荐档。快速安装器由注册表驱动，声明 18 个轻量宿主档案并共同指向一份版本化共享运行时；声明可安装不等于伪造真机 PASS。运行时不需要 PyTorch、Transformers、ONNX，也不会联网。

每条非空普通对话都必须进入冻结的 20 场景路由，展示对应百分比，并且只结算一次本地奖励池；真实压缩未发生时不显示虚假的 Token 节省。

兼容优势：同一个注册表驱动安装器覆盖 18 种主流 Agent 宿主档案；所有宿主共用一份哈希锁定的 CoreML 运行时和同一套 20 场景百分比/奖励池协议。安装后完全离线，不依赖 PyTorch、Transformers、ONNX，也不要求用户配置模型路径。

安装 v0.1.11 后运行：

python3 tools/installer.py --host auto

历史 Full Runtime 的 92.53% 仅作历史参考，不是当前安装目标、默认配置、回退路径或前置依赖。Jev 参考值为 53.67%。

最终的全新宿主人机验收必须由 Peter 本人完成；自动化测试不会代替该验收。

项目与作者：Andrew Yang；公司：Soulite Magic｜灵智天成；抖音：UPAI - 升级你的 AI 能力。商业授权请联系 Soulite Magic；也可以通过上方支付宝二维码支持项目持续开发。