PeytonT/openclaw-hermes-repo-skills

OpenClaw and Hermes Agent Repository Skills

This dataset contains repository-skill records mined from two agent/harness repositories:

• —https://github.com/openclaw/openclaw
• —https://github.com/NousResearch/hermes-agent

It was generated by repo-skills-miner: https://github.com/peytontolbert/repository-skill-miner

The dataset is intended for retrieval, routing, classification, and analysis of reusable repository skills. It is not a raw source-code dump. Each row normalizes a mined unit of repository knowledge with provenance, static metadata, a bounded source excerpt, and LLM-generated annotations.

Dataset Files

• —data/train.parquet
• —dataset_summary.json

The Parquet file is compressed with Zstandard and is approximately 66 MB locally.

Row Count

Total rows: 93,039

What Each Row Contains

Each row includes:

• —Stable row id
• —Dataset source label
• —Source repository and revision
• —Source path and line range
• —Language
• —Primitive type and subtype
• —Skill kind, module, qualname, and signature
• —License detected from the source repository
• —Static side-effect hints
• —Required permission hints
• —Failure-mode hints
• —Bounded source excerpt, capped at 12,000 characters
• —LLM annotation model name
• —Annotation summary
• —Annotation primitive labels
• —Suggested use cases
• —Patch-relevance hints
• —Risk notes
• —Verification hints
• —Annotation confidence

Schema

Columns:

• —id
• —dataset_repo_label
• —source_repo
• —source_revision
• —source_path
• —language
• —primitive_type
• —primitive_subtype
• —license
• —skill_kind
• —module
• —qualname
• —signature
• —line_start
• —line_end
• —side_effects
• —required_permissions
• —failure_modes
• —source_excerpt
• —source_excerpt_truncated
• —annotation_model
• —annotation_summary
• —annotation_primitive_labels
• —annotation_use_when
• —annotation_patch_relevance
• —annotation_risks
• —annotation_verification_hints
• —annotation_confidence

Skill Types

Overall distribution:

Primitive-type distribution:

Language and Surface Coverage

Additional smaller categories include unknown text artifacts, build configs, config files, PowerShell, example files, properties files, patches, and batch scripts.

OpenClaw Contents

OpenClaw contributes 27,746 rows.

Breakdown:

OpenClaw rows emphasize:

• —TypeScript runtime and plugin surfaces
• —Gateway and agent instructions
• —AGENTS.md guardrails
• —SKILL.md skill manifests
• —Plugin manifests
• —CI workflows
• —Scripts and operational configs
• —Swift, Kotlin/JVM, Go, shell, JSON, YAML, and web assets
• —QA scenarios and documentation sections

Representative examples include:

• —Gateway guardrails from src/gateway/AGENTS.md
• —TypeScript symbol indexes for runtime, gateway, plugin, and UI files
• —Script artifacts such as SQLite/vector smoke tests
• —Skill packaging and skill-creator helper scripts
• —CI and release workflow artifacts

Hermes Agent Contents

Hermes Agent contributes 65,293 rows.

Breakdown:

Hermes rows emphasize:

• —Python tests and runtime code symbols
• —Agent, tool, gateway, CLI, and plugin code
• —Optional skills
• —Skill documentation and references
• —Website and developer documentation
• —Configuration and setup artifacts
• —Integration and unit-test surfaces

Representative examples include:

• —Python test methods/classes for tools, plugins, providers, CLI, and agent runtime
• —Skill reference sections under skills/ and optional-skills/
• —Hermes CLI/gateway code symbols
• —Website and user-guide documentation sections
• —Plugin and tool configuration artifacts

Static Hints

The miner adds conservative static hints for side effects, required permissions, and failure modes. These are routing and risk hints, not formal verification claims.

Common side-effect hints:

Common required-permission hints:

Common failure-mode hints:

LLM Annotations

Every row has an annotation record. The annotation fields are designed to support retrieval and downstream model training.

Important fields:

• —annotation_summary
• —annotation_primitive_labels
• —annotation_use_when
• —annotation_patch_relevance
• —annotation_risks
• —annotation_verification_hints
• —annotation_confidence

Annotation labels are useful but noisy. For example, many Hermes rows are labeled as unit_test because the repository contains a large number of test methods and test classes. Treat these labels as weak supervision, not gold labels.

Known annotation/card notes:

• —Missing annotation rows: 0
• —Rows with empty annotation_summary: 8
• —annotation_confidence is not calibrated and should not be interpreted as a probability.
• —Some generated labels differ by casing or phrasing, such as Documentation and documentation.

Recommended Uses

This dataset is suitable for:

• —Skill retrieval experiments
• —Repository-skill indexing
• —Code/documentation routing
• —Primitive classification
• —Patch-relevance retrieval
• —Verification-hint retrieval
• —Training or evaluating hierarchical retrievers such as ToLBERT-style models
• —Building retrieval-time operator libraries for patch agents

The strongest retrieval fields are usually:

• —source_path
• —primitive_type
• —primitive_subtype
• —skill_kind
• —source_excerpt
• —annotation_summary
• —annotation_use_when
• —annotation_verification_hints
• —annotation_patch_relevance

Limitations

• —Source excerpts are bounded to 12,000 characters and may be truncated.
• —Static side-effect, permission, and failure-mode fields are heuristic hints.
• —LLM annotations are weak supervision and may contain noisy labels.
• —The dataset reflects the local repository snapshots used during mining.
• —Licensing is inherited from source repositories and represented per row via the detected source license field. Review upstream licenses before redistribution or commercial use.

Generation

Generated by:

https://github.com/peytontolbert/repo-skill-miner

High-level generation flow:

1. 1.Clone source repositories locally.
2. 2.Mine repository skills with generic extractors.
3. 3.Add code-symbol, symbol-index, artifact, document-section, and fallback text-artifact records.
4. 4.Run distributed LLM annotation workers over the mined Parquet cards.
5. 5.Join full skill cards with annotation shards.
6. 6.Reconstruct bounded source excerpts from the local checkouts.
7. 7.Write Hugging Face-style compressed Parquet.

The dataset was generated as Parquet to avoid the size and memory overhead of JSONL for this workload.