Skip to content
Back to search

SkVM

The Language Virtual Machine for Agent Skills

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today
Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/sjtu-ipads--skvm/badge.svg)](/listings/sjtu-ipads--skvm)

HTML

<a href="/listings/sjtu-ipads--skvm"><img src="/api/listings/sjtu-ipads--skvm/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

Install instructions from the README

As written by the repo's author. findsafeskills has not verified these steps.

These steps run a script from the network that findsafeskills does not scan.

# curl one-liner (macOS / Linux)
curl -fsSL https://skillvm.ai/install.sh | sh

# or via npm (any platform with Node ≥ 18; postinstall fetches the matching binary)
npm i -g @ipads-skvm/skvm

The installer drops a standalone binary at `~/.local/share/skvm/bin/skvm` (symlinked into `~/.local/bin/skvm`) and bundles a private, isolated headless agent runtime used internally by `skvm jit-optimize` — it is fully self-contained and does not touch any agent or CLI you may have installed globally.

**Agent-facing skills** ship inside the install. Copy them into your agent harness's skills directory to teach it how to drive skvm:

# OpenClaw
cp -r ~/.local/share/skvm/skills/skvm-jit ~/.openclaw/workspace/skills/
cp -r ~/.local/share/skvm/skills/skvm-general ~/.openclaw/workspace/skills/

# Hermes Agent
cp -r ~/.local/share/skvm/skills/skvm-jit ~/.hermes/skills/
cp -r ~/.local/share/skvm/skills/skvm-general ~/.hermes/skills/

# pi Agent
mkdir -p ~/.pi/agent/skills
cp -r ~/.local/share/skvm/skills/skvm-jit ~/.pi/agent/skills/
cp -r ~/.local/share/skvm/skills/skvm-general ~/.pi/agent/skills/

- `skvm-jit` — post-task JIT optimization skill for submitting conversation logs to `skvm jit-optimize` - `skvm-general` — drives `profile` / `aot-compile` / `bench` / `proposals` on behalf of a user

Install SkVM for this machine with the official installer:
curl -fsSL https://skillvm.ai/install.sh | sh

Then install the two OpenClaw skills with these commands:
mkdir -p ~/.openclaw/workspace/skills

mkdir -p ~/.openclaw/workspace/skills/skvm-general
curl -fsSL https://skillvm.ai/skills/skvm-general/SKILL.md -o ~/.openclaw/workspace/skills/skvm-general/SKILL.md

mkdir -p ~/.openclaw/workspace/skills/skvm-jit
curl -fsSL https://skillvm.ai/skills/skvm-jit/SKILL.md -o ~/.openclaw/workspace/skills/skvm-jit/SKILL.md
curl -fsSL https://skillvm.ai/skills/skvm-jit/adapter-skill-paths.md -o ~/.openclaw/workspace/skills/skvm-jit/adapter-skill-paths.md

This should create:
~/.openclaw/workspace/skills/skvm-general/SKILL.md
~/.openclaw/workspace/skills/skvm-jit/SKILL.md
~/.openclaw/workspace/skills/skvm-jit/adapter-skill-paths.md

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

README.md · README · 275 lines · 2 findings marked

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

  1. <div align="center">
  2. # SkVM
  3. **Compile and run LLM agent skills across heterogeneous models and harnesses**
  4. [English](./README.md) | [中文](./README.zh-CN.md)
  5. [Website](https://skillvm.ai) · [GitHub](https://github.com/SJTU-IPADS/SkVM) · [Paper](https://arxiv.org/abs/2604.03088)
  6. [![npm](https://img.shields.io/npm/v/@ipads-skvm/skvm?color=cb3837&logo=npm)](https://www.npmjs.com/package/@ipads-skvm/skvm)
  7. [![license](https://img.shields.io/github/license/SJTU-IPADS/SkVM)](./LICENSE)
  8. [![release](https://img.shields.io/github/v/release/SJTU-IPADS/SkVM)](https://github.com/SJTU-IPADS/SkVM/releases)
  9. [![last commit](https://img.shields.io/github/last-commit/SJTU-IPADS/SkVM)](https://github.com/SJTU-IPADS/SkVM/commits)
  10. </div>
  11. SkVM is a compilation and runtime system that makes LLM agent skills portable across heterogeneous models and harnesses. It has four major parts:
  12. - **Profiling** — measure a model+harness against pre-defined primitive capabilities
  13. - **AOT-Compilation** — compile a skill with multiple passes in AOT compiler
  14. - **JIT-Optimization** — improve runtime speed (JIT-boost) and skill content (JIT-optimize)
  15. - **Benchmark** — evaluate original, compiled, and optimized skills across tasks, conditions, and models
  16. <p align="center">
  17. <img src="./docs/skvm_arch.png" alt="SkVM architecture" width="66%" />
  18. </p>
  19. Reference: **SkVM: Revisiting Language VM for Skills across Heterogenous LLMs and Harnesses** — https://arxiv.org/abs/2604.03088
  20. ## News
  21. - **2026-05-18 — Headless pi driver, on by default.** `jit-optimize` and `jit-boost` now run their internal coding agent via [`@mariozechner/pi-coding-agent`](https://www.npmjs.com/package/@mariozechner/pi-coding-agent) in-process by default. Opencode remains a first-class peer — set `headlessAgent.driver: "opencode"` in `skvm.config.json` to keep using it. `ProposalMeta.optimizerDriver` records which driver produced each proposal for replay / audit.
  22. - **2026-05** — New `claude-code` adapter (drives the `claude -p` CLI). Note: heavy headless usage may hit account rate limits or usage-terms issues.
  23. - **2026-05** — Upload and optimize a skill with SkVM in your browser: [SkVM website](https://skillvm.ai/index.html#optimize-skill).
  24. ## Demo
  25. ### SkVM Optimization: Skill Quality Comparison
  26. <div align="center">
  27. <video src="https://github.com/user-attachments/assets/b21132b5-a697-41b9-a802-476389bdb466"/>
  28. </div>
  29. ### SkVM Accelerates Agent Execution
  30. <div align="center">
  31. <video src="https://github.com/user-attachments/assets/05636f9e-1d13-43c6-b82b-43af577fbf52"/>
  32. </div>
  33. ## Install
  34. ```bash
  35. # curl one-liner (macOS / Linux)
  36. curl -fsSL https://skillvm.ai/install.sh | sh
    • Note: Pipes a downloaded script directly into a shell (curl | sh / wget | bash)
    • Line 55 on GitHub
  37. # or via npm (any platform with Node ≥ 18; postinstall fetches the matching binary)
  38. npm i -g @ipads-skvm/skvm
  39. ```
  40. The installer drops a standalone binary at `~/.local/share/skvm/bin/skvm` (symlinked into `~/.local/bin/skvm`) and bundles a private, isolated headless agent runtime used internally by `skvm jit-optimize` — it is fully self-contained and does not touch any agent or CLI you may have installed globally.
  41. **Agent-facing skills** ship inside the install. Copy them into your agent harness's skills directory to teach it how to drive skvm:
  42. ```bash
  43. # OpenClaw
  44. cp -r ~/.local/share/skvm/skills/skvm-jit ~/.openclaw/workspace/skills/
  45. cp -r ~/.local/share/skvm/skills/skvm-general ~/.openclaw/workspace/skills/
  46. # Hermes Agent
  47. cp -r ~/.local/share/skvm/skills/skvm-jit ~/.hermes/skills/
  48. cp -r ~/.local/share/skvm/skills/skvm-general ~/.hermes/skills/
  49. # pi Agent
  50. mkdir -p ~/.pi/agent/skills
  51. cp -r ~/.local/share/skvm/skills/skvm-jit ~/.pi/agent/skills/
  52. cp -r ~/.local/share/skvm/skills/skvm-general ~/.pi/agent/skills/
  53. ```
  54. - `skvm-jit` — post-task JIT optimization skill for submitting conversation logs to `skvm jit-optimize`
  55. - `skvm-general` — drives `profile` / `aot-compile` / `bench` / `proposals` on behalf of a user
  56. <!-- ### Quick OpenClaw Integration
  57. If you want OpenClaw to bootstrap SkVM for itself, you can give OpenClaw the following instruction block. It tells OpenClaw to install SkVM with the official installer, download `skvm-general` and `skvm-jit` from their concrete SkillVM URLs, and place them into OpenClaw's global skills directory at `~/.openclaw/workspace/skills/`.
  58. ```text
  59. Install SkVM for this machine with the official installer:
  60. curl -fsSL https://skillvm.ai/install.sh | sh
    • Note: Pipes a downloaded script directly into a shell (curl | sh / wget | bash)
    • Line 89 on GitHub
  61. Then install the two OpenClaw skills with these commands:
  62. mkdir -p ~/.openclaw/workspace/skills
  63. mkdir -p ~/.openclaw/workspace/skills/skvm-general
  64. curl -fsSL https://skillvm.ai/skills/skvm-general/SKILL.md -o ~/.openclaw/workspace/skills/skvm-general/SKILL.md
  65. mkdir -p ~/.openclaw/workspace/skills/skvm-jit
  66. curl -fsSL https://skillvm.ai/skills/skvm-jit/SKILL.md -o ~/.openclaw/workspace/skills/skvm-jit/SKILL.md
  67. curl -fsSL https://skillvm.ai/skills/skvm-jit/adapter-skill-paths.md -o ~/.openclaw/workspace/skills/skvm-jit/adapter-skill-paths.md
  68. This should create:
  69. ~/.openclaw/workspace/skills/skvm-general/SKILL.md
  70. ~/.openclaw/workspace/skills/skvm-jit/SKILL.md
  71. ~/.openclaw/workspace/skills/skvm-jit/adapter-skill-paths.md
  72. ```
  73. **After installation, when executing any task that uses a skill, append this instruction to the task: "If this task uses a skill, use SkVM to optimize that skill during or after execution."**
  74. This is the quickest way to make an OpenClaw environment self-host the SkVM toolchain and the two integration skills without requiring a separate manual setup flow. -->
  75. ## Quick Start
  76. Configure your agent harness, provider, and API key via the interactive wizard:
  77. ```bash
  78. skvm config init
  79. ```
  80. This writes `$SKVM_CACHE/skvm.config.json` (default `~/.skvm/skvm.config.json`). For non-interactive setups, see [docs/providers.md](docs/providers.md).
  81. **Adapter config mode.** External harnesses run in one of two modes: `managed|native` (default `managed`):
  82. - **managed** — skvm provisions a fresh, minimal config inside the sandbox (e.g. a new openclaw agent, or a generated `hermes` `config.yaml`). Clean baseline, no reliance on host state.
  83. - **native** — skvm clones the user's existing harness config (e.g. an openclaw source agent under `~/.openclaw/agents/<name>`, or a hermes active profile). Requires the harness to be set up locally first.
  84. **Model id format.** Model parameters on the CLI take the form `<provider>/<model-id>` — the leading `<provider>` selects the model service provider, while `<model-id>` is how that provider refers to the model. For OpenRouter that's three segments, e.g. `openrouter/qwen/qwen3.5-35b-a3b`; for the Anthropic native API it's two, e.g. `anthropic/claude-sonnet-4.6`.
  85. ### 1. Profile a model's primitive capabilities
  86. Writes a target capability profile to `~/.skvm/profiles/`.
  87. If your target model + adapter pair is already covered by the pre-built profiles shipped in `skvm-data/profiles/`, you can copy the cached result into your local profile cache and skip `skvm profile` entirely:
  88. ```bash
  89. mkdir -p ~/.skvm/profiles
  90. cp -R skvm-data/profiles/. ~/.skvm/profiles/
  91. ```
  92. ```bash
  93. skvm profile \
  94. --adapter=bare-agent \
  95. --model=<provider>/<model-id>
  96. # e.g. --model=anthropic/claude-sonnet-4.6
  97. # e.g. --model=openrouter/qwen/qwen3.5-35b-a3b
  98. ```
  99. With the default `--concurrency=1`, this example typically takes about 20 minutes for one full run. If you want it to finish faster, increase `--concurrency` to profile more primitives in parallel.
  100. ### 2. Compile a skill against that profile
  101. The compiler rewrites the skill to match the target's capabilities. A cached profile for the same `--model` + `--adapter` pair must exist when a selected pass consumes it — only pass 1 (`rewrite-skill`) reads the profile (run `skvm profile` first, or use `skvm pipeline` which profiles automatically). A compile of only pass 2/3 (e.g. `--pass=bind-env` for `env-setup.sh` generation) needs no profile; see the `tcp` column of `--list-passes`.
  102. ```bash
  103. skvm aot-compile \
  104. --skill=path/to/skill-dir \
  105. --model=<provider>/<model-id> \
  106. --adapter=bare-agent \
  107. --pass=1 \
  108. --compiler-model=<provider>/<model-id>
  109. ```
  110. Compiled variants are written under `~/.skvm/proposals/aot-compile/<adapter>/<safeModel>/<skillName>/<passTag>/` by default.
  111. ### 3. Autotune the skill with synthetic tasks
  112. The optimizer LLM derives tasks from the skill itself, then loops edit → rerun → score.
  113. By default, synthetic mode generates 2 training tasks and 1 held-out test task.
  114. ```bash
  115. skvm jit-optimize \
  116. --skill=path/to/skill-dir \
  117. --task-source=synthetic \
  118. --task-concurrency=3 \
  119. --target-adapter=bare-agent \
  120. --optimizer-model=<provider>/<model-id> \
  121. --rounds=1 \
  122. --target-model=<provider>/<model-id>
  123. ```
  124. Results are written under `~/.skvm/proposals/jit-optimize/<adapter>/<safeTargetModel>/<skillName>/<timestamp>/` by default.
  125. ### 4. Optimize from an existing conversation log
  126. No rerun, just diagnose and edit. Good for post-mortems and for the `skvm-jit` post-task optimization hook.
  127. ```bash
  128. skvm jit-optimize \
  129. --skill=path/to/skill-dir \
  130. --task-source=log \
  131. --target-adapter=bare-agent \
  132. --logs=path/to/session.jsonl \
  133. --optimizer-model=<provider>/<model-id> \
  134. --target-model=<provider>/<model-id>
  135. ```
  136. ### Review, accept, or reject the proposal
  137. ```bash
  138. skvm proposals list # CLI listing
  139. skvm proposals show <id> # CLI detail view
  140. skvm proposals accept <id> # CLI accept
  141. skvm ui # Web review UI
  142. ```
  143. ## Configuration
  144. SkVM keeps all runtime artifacts — cached profiles, proposal trees, bench and compile logs — under a single cache root:
  145. ```
  146. ~/.skvm/
  147. ├── profiles/ # Cached target capability profiles
  148. ├── log/ # Profile, compile, bench, and runtime logs
  149. └── proposals/ # AOT-compile, jit-boost, jit-optimize outputs
  150. ```
  151. The cache is user-global and shared across every directory you invoke `skvm` from, so profiles cached in one project are reused everywhere. Override the location via:
  152. - `--skvm-cache=<path>` flag (one-off)
  153. - `SKVM_CACHE` env var (persistent), e.g. `export SKVM_CACHE=/mnt/fast/skvm`
  154. Individual subdirectories can also be pointed elsewhere with `SKVM_PROFILES_DIR`, `SKVM_LOGS_DIR`, and `SKVM_PROPOSALS_DIR`.
  155. ## Dataset: skvm-data
  156. The benchmark skills, tasks, and pre-built profiles live in a separate Git submodule ([SJTU-IPADS/SkVM-data](https://github.com/SJTU-IPADS/SkVM-data)). Clone it if you plan to run `skvm bench` and want to use the bundled skills/tasks directly:
  157. ```bash
  158. git submodule update --init # or: git clone --recurse-submodules
  159. ```
  160. This populates the `skvm-data/` directory:
  161. ```
  162. skvm-data/
  163. ├── skills/ # 108 skill directories (each contains a SKILL.md)
  164. ├── tasks/ # 216 task directories (each contains a task.json)
  165. └── profiles/ # Pre-built target capability profiles
  166. ├── bare-agent/
  167. ├── hermes/
  168. └── openclaw/
  169. ```
  170. `skvm bench` resolves skills and tasks from `skvm-data/` by default. Override the location via:
  171. - `--skvm-data-dir=<path>` flag (one-off)
  172. - `SKVM_DATA_DIR` env var (persistent)
  173. Commands that take an explicit `--skill=<path>` or `--task=<path>` do not need the submodule — they work with any directory on disk.
  174. `skvm-data/profiles/` already includes pre-built profiles for some model + adapter combinations. If the pair you need is already there, copy `skvm-data/profiles/` into your profile cache directory (default: `~/.skvm/profiles/`, or `SKVM_PROFILES_DIR` if set) and you can skip running `skvm profile` for that target. See the profile list in the profiling section above for the currently bundled combinations.
  175. ```bash
  176. mkdir -p ~/.skvm/profiles
  177. cp -R skvm-data/profiles/. ~/.skvm/profiles/
  178. ```
  179. ## Learn more
  180. - **[docs/usage.md](docs/usage.md)** — full command reference: `profile`, `aot-compile`, `run`, `bench`, `jit-optimize`, `proposals`, and more
  181. - **[docs/architecture.md](docs/architecture.md)** — subsystem map, data flow, and on-disk layout
  182. - **[docs/grade-protocols.md](docs/grade-protocols.md)** — grader protocol reference for custom `grade.py` task graders
  183. - **Paper**: https://arxiv.org/abs/2604.03088
  184. ## Citation
  185. If you use SkVM in your research, please cite:
  186. ```bibtex
  187. @article{chen2026skvm,
  188. title={SkVM: Revisiting Language VM for Skills across Heterogenous LLMs and Harnesses},
  189. author={Chen, Le and Feng, Erhu and Xia, Yubin and Chen, Haibo},
  190. journal={arXiv preprint arXiv:2604.03088},
  191. year={2026}
  192. }
  193. ```

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.