Auto-ML-Skills

DisCo: A skill-powered research agent with task-related operating knowledge

English | 简体中文

Auto-ML-Skills specializes what a research agent knows to consider without prescribing how it must research. Relevant methods, procedures, checks, and recovery actions are scattered across an LLM's parametric prior, repositories, papers, benchmark resources, and other knowledge sources. DisCo reorganizes source-grounded evidence into operating-knowledge skill graphs. In Creator mode it constructs and verifies those graphs; in Researcher mode it loads the task-relevant portion through progressive disclosure and uses code, tools, and experiments to complete the research task.

The Research Skills Library is the growing collection of these reusable skill graphs. Its scope includes repository skills, paper-derived skills, task- or benchmark-specific skills, and future domain collections. This public checkout currently ships more than 1,000 operating skills organized into graphs for 170 widely used ML and research repositories across 14 workflow-oriented families, plus a library-level router. The repository collection is therefore part of the library, not its boundary.

Naming note: Auto-ML-Skills is the repository name. DisCo is the user-facing CLI/runtime, and the Research Skills Library is the published reusable skill collection.

🧭 Table Of Contents

• 📣 News
• 💡 Why Auto-ML-Skills
• 🧰 What Is Included
• 🗂️ Library Coverage
• ⚙️ Installation
• 🚀 Quick Start
• 🛠️ Creator Meta Skills
• 🤝 Contributing
• 📚 Documentation
• 🙏 Acknowledgement
• 📄 License
• 📝 Citation

📣 News

• 2026-08-03: Auto-ML-Skills brings together DisCo's Creator and Researcher workflows and a Research Skills Library with 1,000+ operating skills for 170 widely used repositories.

💡 Why Auto-ML-Skills

Research agents can propose methods, write code, run experiments, and iterate on solutions. On a concrete ML research task, however, their success also depends on whether they can identify and apply the right existing knowledge within the same limited execution budget used to produce the final result.

• Task-relevant knowledge is dispersed. Applicable methods, tools, procedures, validation checks, and recovery actions may be split across the model's prior, source repositories, papers, task assets, and web resources.
• Raw sources are not operating context. An agent still needs to determine when a source applies, which evidence matters, how to execute it, what it should observe, and how to detect and recover from failure.
• Reconstruction competes with research. Re-reading and reconciling sources during every task consumes context, tool calls, compute, and experiment budget that could otherwise improve the research outcome.
• Harnesses and skills specialize different things. A harness governs how an agent plans, searches, critiques, and executes. Skills organize what the agent knows to consider, without forcing a fixed research trajectory.

We call this task-related body of methods, procedures, checks, and recovery actions operating knowledge. DisCo turns source-oriented knowledge into an operating-knowledge skill graph whose skills declare use conditions, execution behavior, supporting evidence, validation steps, and failure handling. Creator mode performs this ML knowledge distillation through evidence-backed exploration, skill-graph generation, and verification with refinement. Researcher mode then loads only the relevant graph fragment as its operating context and applies it to the task.

🧰 What Is Included

Component	Purpose
DisCo	The TypeScript research-agent CLI. Creator constructs and maintains skill graphs; Researcher uses operating skills with code, tools, and experiments.
Research Skills Library	The reusable operating-knowledge layer. This checkout publishes more than 1,000 skills organized into graphs for 170 widely used repositories, plus their sibling repo-skills-router; the library model also supports paper-, task-, benchmark-, and domain-specific collections.
Examples	Sanitized end-to-end sessions showing Creator building a FlagEmbedding skill graph and Researcher applying Gymnasium and Stable-Baselines3 skills to an auditable battery-dispatch experiment.
Documentation	Workflow, architecture, catalog, deployment, and portability guides. The 15 bundled Creator meta skills remain part of DisCo rather than a separate manual installation.

Each session exposes a mode-specific skill inventory:

Mode	Visible skills	Responsibility
Researcher (default)	operating skills and user skills without metadata.disco-role	Use routed operating knowledge, code, tools, and experiments to complete an ML research task.
Creator	Skills explicitly marked metadata.disco-role: meta	Start with distill-ml-knowledge, choose direct, reuse-existing (one workflow or a bounded composition), or design-reusable, and send only a verified recurring construction gap to design-meta-skill.

Use --agent-mode creator|researcher in non-interactive sessions, or /creator and /researcher in the interactive UI. Switching opens a new session with a clean context. See DisCo Workflows for session/export behavior, detailed workflows, and deployment rules.

🗂️ Library Coverage

The public repository collection spans four broad areas and 14 workflow families. It is built from a curated set of 170 widely used ML and research repositories, producing 1,060 root and focused skills with broad coverage of training, data, evaluation, agents, retrieval, vision, generation, ML operations, and scientific computing.

DisCo's Creator CLI and bundled meta skills provide repeatable workflows for creating, verifying, maintaining, and importing new skill graphs. A library-level router and progressive disclosure let the library scale across more repositories, papers, benchmarks, tasks, and domain collections while loading only the skills relevant to the current task.

The Imported Repo Skills Catalog lists every published graph with its upstream repository, update date, package version, source commit, and intended workflow coverage.

⚙️ Installation

Using DisCo with the collection published in this repository requires both of the following installation steps, in order:

1. Install the disco CLI.
2. Install the public repository-skill collection into DisCo's managed skill directory.

Installing portable Creator meta skills into another agent is optional. DisCo already bundles them.

Install DisCo

Install the DisCo CLI from npm:

npm install -g @auto-ml-skills/disco
disco

DisCo requires Node.js >=22.19.0 and builds on Pi's multi-provider model layer. The npm package includes its own DisCo-modified coding-agent source and uses pinned @earendil-works/pi-agent-core, @earendil-works/pi-ai, and @earendil-works/pi-tui packages as dependencies. It does not depend on @earendil-works/pi-coding-agent, discover .pi resources, or share a globally installed Pi dependency tree.

Configure at least one provider in the startup flow with /login, or use environment variables such as OPENAI_API_KEY, ANTHROPIC_API_KEY, GEMINI_API_KEY, OPENROUTER_API_KEY, or MISTRAL_API_KEY.

Build from source for local development

git clone https://github.com/VectorSpaceLab/Auto-ML-Skills.git
cd Auto-ML-Skills
# If model catalog fetching fails behind an HTTP(S) proxy, use: NODE_USE_ENV_PROXY=1 bash scripts/build-from-source-link.sh
bash scripts/build-from-source-link.sh

The script installs the standalone package dependencies from the checked-in shrinkwrap, builds DisCo, and links the disco command globally for local use.

Install The Published Repository Collection

Install the official collection and its router with DisCo:

disco repo-skills install

The command uses a shallow checkout of the official repository, installs only the published runtime collection, and records its source commit. Git must be available locally. Later, inspect or update the managed collection with:

disco repo-skills status
disco repo-skills update

status is local-only: it checks managed digests, router presence and current skill coverage without contacting GitHub. Run update when you want to check and apply the latest official commit.

Updates replace only official managed skill IDs. Repo skills created or imported locally by Creator are preserved. If an official skill was modified locally or collides with an unmanaged skill, the command stops; an explicit --force update first keeps a recoverable backup.

DisCo registers the managed collection, but its repository roots and focused sub-skills use disable-model-invocation: true and are omitted from the initial model context. By default, repo-skills-router remains visible, routes to one practical scenario, and then points DisCo to the selected skill under its sibling repo-skills/ collection. Automatic router selection can be disabled and restored without uninstalling the collection:

disco repo-skills router disable
disco repo-skills router enable

When disabled, the router is omitted from automatic model selection but remains registered for explicit /skill:repo-skills-router invocation. Start a new Researcher session after an install, update, or router setting change.

Manual installation fallback

git clone https://github.com/VectorSpaceLab/Auto-ML-Skills.git
cd Auto-ML-Skills
mkdir -p ~/.disco/agent/skills
cp -R \
  research-skills-library/repo-skills \
  research-skills-library/repo-skills-router \
  ~/.disco/agent/skills/

Running disco repo-skills install later can adopt an unchanged manual copy and preserve additional local skill IDs.

For router behavior, third-party skill packages, and deployment-scope details, see DisCo Workflows, the Research Skills Library guide, and the DisCo CLI README.

Portable Meta Skills For Another Agent (Optional)

DisCo already bundles its Creator workflows. To run them in another compatible agent, follow Meta Skills For Other Agents.

🚀 Quick Start

Researcher is the default. After completing both installation steps, ask for a concrete research outcome; relevant repository skills are selected through the router and opened progressively:

disco --agent-mode researcher -p "Benchmark vLLM and SGLang with the same model and workload on this machine. Tune each server under identical hardware and memory constraints, report the best verified throughput for each, and preserve the commands and measurements needed to reproduce the comparison."

Use Creator when the task needs a new or updated operating-skill graph. It starts with distill-ml-knowledge, which normalizes the task and selects direct, reuse-existing, or design-reusable. Direct distillation builds a task-conditioned graph; reuse-existing invokes one adequate workflow or a bounded composition; design-reusable hands an evidence-backed recurring construction gap to design-meta-skill.

Create skills for a specific repository. The bundled repository workflow already covers source inspection, environment preparation, skill generation, verification, and import. distill-ml-knowledge therefore selects reuse-existing with reuse mode: single and hands this request to create-repo-skill:

git clone https://github.com/FlagOpen/FlagEmbedding.git
disco --agent-mode creator -p \
  "/skill:distill-ml-knowledge Create and verify a repository skill graph for the local FlagEmbedding checkout at ./FlagEmbedding. Use its documentation, examples, tests, and runtime APIs to cover embedding inference and evaluation. Prepare the CPU/GPU environments required by the confirmed scope, then ask before importing the verified graph."

Create one task-related skill directly. Use direct when an approved knowledge-source directory can support the current task immediately and no new reusable construction workflow is needed. This example distills guidance for the classic Kaggle Home Credit Default Risk problem into one runtime SKILL.md:

disco --agent-mode creator -p \
  "/skill:distill-ml-knowledge Use the direct path to create a task-related skill for Kaggle Home Credit Default Risk. The task description is at /path/to/home-credit-task/description.md, and the deep-research materials are in /path/to/home-credit-deep-research/. Output a single SKILL.md, verify it, and ask before importing it to .agents/skills/home-credit-default-risk/."

Creator validates and stages every output for approval. The repository graph uses DisCo's managed repo collection, while the task-bound competition skill defaults to the current project's .agents/skills/; each is imported only after its own verification and approval, then used in a new Researcher session. The full DisCo Workflows guide covers explicit skill invocation, task-specific graphs, repository and paper construction, verification, maintenance, export, deployment scopes, and mode-switch behavior.

🛠️ Creator Meta Skills

DisCo bundles 15 Creator-only meta skills for adequacy assessment, workflow design, repository and paper distillation, verification, maintenance, and cross-agent export. The canonical distill-ml-knowledge entry point owns task normalization, adequacy/composition assessment, and path selection; design-meta-skill only designs the reusable bundle after receiving a verified design-reusable handoff. Each declares metadata.disco-role: meta; Researcher does not see them. The Bundled Skills Reference is the source of truth for the full inventory and contracts. To run these workflows outside DisCo, see Meta Skills For Other Agents.

🤝 Contributing

We welcome contributions in three main areas:

1. Contribute generated repo skills. Add a publishable runtime skill under research-skills-library/repo-skills/<skill-id>/, include provenance and routing metadata, and update the sibling research-skills-library/repo-skills-router/ so agents can discover it.
2. Extend or refresh existing repo skills. Improve stale, incomplete, or unclear skills with source-grounded changes. Update provenance or routing metadata when the upstream baseline or coverage changes.
3. Improve the DisCo CLI source. Changes to the TypeScript CLI under src/ are welcome, including package/repo and paper-to-skill workflows. Run focused checks and document behavior changes. Repo-skill workflow changes should preserve the create/verify split, review/test artifact layout, import-readiness gates, and locked router-update transaction. Updates to the integrated Paper2Skills workflow should preserve its source-resolution, modularization, generated-skill validation, recovery, analysis, and final-report contracts.

For repo-skill PRs, list the model, provider, reasoning or thinking level, source repository commit, and verification steps used to produce or revise the skill. For DisCo CLI changes that touch paper-to-skill behavior, include the paper source, run config, recovery mode, validation artifacts, and final report path when applicable. See CONTRIBUTING.md for the full checklist.

📚 Documentation

Page	Description
DisCo Workflows	Mode and session behavior, Researcher execution, Creator construction and maintenance, deployment scopes, and cross-agent export.
Examples	Sanitized Creator and Researcher session exports, including FlagEmbedding skill construction and Gymnasium/Stable-Baselines3 battery dispatch.
Imported Repo Skills Catalog	Public catalog of included runtime repo skills, organized by repository-skill family with upstream baselines.
Research Skills Library	Broader library model, current repository-skill collection, canonical layout, and DisCo installation command.
Architecture	Repository layers, mode boundaries, runtime routing, source layout, authoring pipelines, project/managed deployment scopes, and the managed repository library.
Bundled Skills Reference	Role metadata, Creator meta-skill contracts, Researcher routing, and construction artifact layouts.
Meta Skills For Other Agents	Portable Creator-only workflow installation for Codex, Claude Code, and project-local agents.
DisCo CLI README	General task execution, runtime skill routing, package installation, and skill authoring workflows.
Contributing	Contribution rules for generated repo skills, router/catalog updates, workflow skills, documentation, and CLI source.

🙏 Acknowledgement

DisCo's CLI and agent runtime are built on the foundation of earendil-works/pi, an open-source AI agent toolkit with a unified LLM API, agent loop, terminal UI, and coding-agent CLI.

Auto-ML-Skills is also made possible by the GitHub open-source community. The repo skills in this library exist because many researchers and engineers have released high-quality ML, agent, data, bio/chem, vision, and infrastructure projects for the community to build on.

📄 License

The repository-level Auto-ML-Skills materials and the open-sourced runtime repo skills under research-skills-library/ are released under the Apache License 2.0 unless a file states otherwise. The standalone DisCo npm package under src/ is distributed under its own MIT License, with upstream attribution and third-party terms in src/THIRD_PARTY_NOTICES.md.

See LICENSE for the repository-level Apache-2.0 text.

📝 Citation

TBA