Harvey
/skill-set Design (Ollama-Centric)
Directory Structure
agents/
├── skills/ # Skill implementations (e.g., `lint`, `test`)
└── skill-sets/ # Skill set YAML files (e.g., `code-review.yaml`)YAML Structure (Skill Set)
name: code-review
description: Tools for code quality
skills:
- trigger: /lint
prompt: "Analyze this {language} code for PEP8/compliance: {code}"
- trigger: /test
prompt: "Write tests for this {language} function: {code}"
metadata:
version: 1.0
dependencies: [pylint, flake8]Commands
| Command | Action |
|---|---|
/skill-set load <name> |
Loads <name>.yaml. Rejects if triggers are
missing/duplicate. Calculates token count via Ollama API. |
/skill-set unload |
Clears the current skill set from context. |
/skill-set list |
Lists all YAML files in agents/skill-sets/. |
/skill-set create <name> |
Creates a new empty YAML file for <name>. |
/skill-set delete <name> |
Deletes <name>.yaml. |
Validation Rules
- Missing Skills: Reject load if any
triggerin the YAML doesn’t exist inagents/skills/. - Duplicate Triggers: Reject load if any
triggeris duplicated in the YAML. - Token Count: Warn if the total tokens for the skill set exceed the model’s context window.
Tokenization
Use Ollama’s
/api/tokenizeendpoint for all models (pulled or local).Example API call:
curl http://localhost:11434/api/tokenize -d '{"model": "phi:2.7b", "content": "Your prompt here"}'
Go Implementation Notes
- Use
gopkg.in/yaml.v3for YAML parsing. - Use
net/httpto call Ollama’s API for tokenization. - Validate triggers against
agents/skills/before loading. - need to figure out how to handle llamafile and tokenization counte