Skip to content

Latest commit

 

History

History
98 lines (75 loc) · 4.74 KB

File metadata and controls

98 lines (75 loc) · 4.74 KB

DisCo Examples

This directory contains self-contained HTML exports of complete DisCo sessions and reusable starter files for documented workflows. Clone or download the repository, open an HTML file in a browser, and expand the tool calls you want to inspect. No DisCo installation or network connection is required to read an export.

Creator

Creator examples are grouped by the workflow that produces the operating skills.

Repo-to-Skills

huggingface_hub Repository Skill

This session follows Creator as it scopes the huggingface_hub 1.29.0 repository, prepares and verifies an isolated inspection environment, uses a first workflow plus missing-only recovery to draft five sub-skills, integrates the verified graph, and records final review and native-test boundaries.

The accompanying artifact bundle contains the generated runtime skill, sanitized routing and review summaries, machine-readable verification results, and usability test cases. Private environment reports, raw native logs, source checkouts, and temporary runtime state are intentionally excluded.

Paper-to-Skills

The Paper2Skills Distiller turns an AI research paper into verified, module-level skills for paper replication. Start from the Distiller TOML configuration, which exposes the source-acquisition, recovery, runtime, iteration, and output settings accepted by DisCo Creator.

From the repository root, copy the starter and edit at least workspace_root, paper_slug, and paper_source. Set original_repo_source to a local path, Git URL, none, or unknown as appropriate:

cp examples/creator/paper-to-skills/distiller-run-config.toml \
  /absolute/path/to/distiller-run-config.toml

disco --creator -p "/skill:create-paper-skills Use Distiller to generate and verify paper-replication skills for each run in this config. config_path: /absolute/path/to/distiller-run-config.toml"

paper_source accepts a local PDF or text file, a direct PDF URL, an arXiv URL or identifier, or a paper title. With the starter's default path settings, generated skills are written under <workspace_root>/<paper_slug>/skills/ and final reports under <workspace_root>/<paper_slug>/distillation/reports/final/. The starter asks before expensive recovery work and before any final deployment that requires approval.

See DisCo Meta Skills for the common Creator entry points and DisCo Workflows for the complete source-resolution, recovery, validation, and deployment contract.

Researcher

vLLM vs SGLang: Qwen3.5-4B Serving Benchmark

This session shows how Researcher uses repository skills to run an auditable serving benchmark. repo-skills-router narrows the task to model deployment and inference serving, then selects only the relevant vLLM and SGLang skills.

The vLLM skill guides server configuration, metrics, and bounded runtime tuning; the SGLang skill guides startup, cache and scheduler settings, workload control, GPU measurement, and framework-specific troubleshooting. Together, they keep the model, workload, client, and hardware limits consistent across both frameworks, while separating correctness checks from performance measurement.

The report records the resulting commands, versions, and evidence. Its runtime comparison is an observation of this environment—not a performance claim about the skills themselves.

Sanitization

Some sensitive and environment-specific information has been sanitized from the public exports while preserving the task, skill usage, execution flow, and results needed to understand the examples.