Skip to content

Contributing

Jump to bottom
github-actions[bot] edited this page Aug 5, 2026 · 2 revisions

Contributing Guide

Thank you for your interest in contributing to agent-toolkit. This document covers everything you need to get started — from dev setup to the full walkthrough for adding skills, agents, and loop templates.

Dev Setup

Prerequisites

Tool Minimum version Purpose
git 2.30 Source control
bash 5.0 Install and validation scripts. macOS ships bash 3 — install via brew install bash
Python 3.10 Validation scripts and CLI
jq 1.6 Catalog generation scripts
node / npm 18 npx skills and MCP server install

Verify your setup:

git --version && bash --version && python3 --version && jq --version

Clone and install

git clone https://github.com/ulises-jeremias/agent-toolkit.git
cd agent-toolkit

# Install the CLI in development mode
pip install -e ".[dev,yaml]"

# Verify
agent-toolkit --help

Python development dependencies

pip install pytest ruff pyyaml jsonschema

Run the test suite

pytest

Validation commands

Always run all validation commands before opening a PR. All must exit 0.

# Validate all skill manifests (schema, frontmatter, no secrets)
bash scripts/validate-skills.sh

# Validate all loop template manifests
bash scripts/validate-loops.sh

# Regenerate catalogs and verify no drift with source files
bash scripts/build-catalog.sh

# Validate agent manifests
python3 scripts/validate-agents.py

# Validate skills with Python validator
python3 scripts/validate-skills.py

# Verify plugin bundles are in sync with canonical sources
python3 scripts/gen-surfaces.py --check

Branch Naming

Prefix When to use
feat/ New skill, loop template, agent persona, or major profile addition
fix/ Bug fix in an existing skill, loop, or profile
docs/ Documentation-only change
chore/ Maintenance: dependency update, script fix, catalog regeneration
schema/ Changes to JSON schemas

Examples:

feat/delivery-gh-address-comments
fix/oss-triage-missing-deny-list
docs/readme-mcp-section
chore/regenerate-catalogs

How to Add a Skill

Full guide: docs/HOW_TO_ADD_SKILL.md

Complete walkthrough

Step 1 — Choose the domain

Domain Purpose Examples
core Orchestration, session management assistant, output-handshake
delivery Software-delivery lifecycle adr, planning, development-workflow
design UI/UX and Figma integration ui-ux-pro-max, figma-implement-design
forge Version-control forge automation github-cli-workflow, gh-fix-ci
integrations Third-party platform connectors slack-cli, linear, clickup-cli
data Data platform validation dbt-validation, snowflake-validation
tooling Developer tooling jupyter-notebook, playwright-cli
ops Operational and health-check utilities triage, llm-cost-advisor
loops Recurring automation loop management loop-runner

Step 2 — Create the directory

mkdir -p skills/<domain>/<skill-name>

Skill names must be kebab-case and match the name field in the frontmatter exactly.

Step 3 — Write SKILL.md

---
name: my-skill
description: >-
  HOW — Clear description of what this skill does and when to use it.
  Include trigger keywords so the AI knows when to invoke this skill.
metadata:
  author: your-github-username
  version: "1.0"
compatibility: Optional — any tool requirements (e.g. "requires gh CLI >= 2.40")
---

# my-skill (HOW)

One-line purpose statement.

## Capabilities

| Can do | Cannot do |
|--------|-----------|
| ... | ... |

## Step-by-step procedure

1. Step one
2. Step two
...

## Output format

Describe what the skill outputs and how it is structured.

## Safety rules

- Never do X
- Always confirm before Y

## Checklist

- [ ] Check one
- [ ] Check two

Use the HOW — prefix for tool skills (operate a CLI/API) and WHAT — for workflow skills (define what to do during a project phase).

Step 4 — Add optional references

For skills with complex domain knowledge, create a references/ subdirectory:

skills/forge/my-skill/
├── SKILL.md
└── references/
    └── DOMAIN_REFERENCE.md

Keep reference files under 200 lines. Link to them explicitly from the skill body.

Step 5 — Run validation

python3 scripts/validate-skills.py

Step 6 — Register in catalogs/skills-layout.json

Open catalogs/skills-layout.json and add your skill's name to the appropriate group array. Then regenerate:

python3 scripts/gen-surfaces.py
python3 scripts/gen-surfaces.py --check  # verify no drift

Step 7 — Register in catalogs/skill-catalog.yaml

- name: my-skill
  domain: forge
  responsibility: HOW
  role: my_skill_role
  triggers:
    - trigger phrase one
    - trigger phrase two
  depends_on: []

Step 8 — Open a PR

PR description checklist:

## Skill Checklist
- [ ] `skills/<domain>/<skill-name>/SKILL.md` created
- [ ] Frontmatter has `name` and `description`
- [ ] `name` in frontmatter matches directory name
- [ ] `python3 scripts/validate-skills.py` passes with no errors
- [ ] Registered in `catalogs/skills-layout.json` (correct group)
- [ ] Registered in `catalogs/skill-catalog.yaml` (with triggers)
- [ ] `python3 scripts/gen-surfaces.py --check` passes
- [ ] No secrets or hardcoded tokens in skill body
- [ ] `references/` documents linked from skill body (if present)

How to Add an Agent

Full guide: docs/HOW_TO_ADD_AGENT.md

Complete walkthrough

Step 1 — Create the directory

mkdir -p agents/<agent-name>

Agent names must be kebab-case and should work as an @mention.

Step 2 — Write AGENT.md

---
name: my-agent
description: >-
  Specialist description. Use when: [trigger keywords].
  The description is shown in the agent picker — make it unambiguous.
tools: Read, Grep, Glob, Bash
---

# My Agent

You are a [role] specialist.

## When invoked

1. First action to take
2. Second action to take
...

## Domain expertise

[Tables, checklists, or structured knowledge the agent uses]

## Operating rules

**Always:**
- ...

**Never:**
- ...

**Escalate when:**
- ...

## Output format

[How the agent structures its responses]

Only grant Write or Edit to agents that explicitly need to mutate files. Review agents should have Read, Grep, Glob, Bash only.

Step 3 — Run validation

python3 scripts/validate-agents.py

Step 4 — Register in catalogs/agent-catalog.yaml

- name: my-agent
  role: my_agent_role
  domain: review
  triggers:
    - trigger phrase
    - another trigger
  handoffs:
    - code-reviewer

Step 5 — Sync into plugin bundles

python3 scripts/gen-surfaces.py
python3 scripts/gen-surfaces.py --check

Step 6 — Open a PR

PR description checklist:

## Agent Checklist
- [ ] `agents/<agent-name>/AGENT.md` created
- [ ] Frontmatter has `name`, `description`, and `tools`
- [ ] `name` in frontmatter matches directory name
- [ ] `python3 scripts/validate-agents.py` passes with no errors
- [ ] Registered in `catalogs/agent-catalog.yaml` with triggers
- [ ] `python3 scripts/gen-surfaces.py --check` passes
- [ ] `tools` list is minimal — only what the agent genuinely needs
- [ ] No secrets or hardcoded tokens in agent body or references

How to Create a Loop Template

Full guide: docs/HOW_TO_CREATE_LOOP.md

Graduated testing: always start at L1

Never deploy a new loop at L2 or L3. Follow this sequence:

  1. Create at L1 (read-only, empty allowlist, full deny list)
  2. Run at L1 for 3+ clean days (no unexpected mutations, report is accurate and stable)
  3. Evaluate for upgrade to L2 (are the proposed actions correct? Do you trust it?)
  4. Upgrade to L2: change tier, populate allowlist, update request prompt to act (not just propose)

Complete walkthrough

Step 1 — Create the directory

mkdir -p loops/<loop-name>

Step 2 — Write loop.yaml

name: my-loop
description: "Daily L1 read-only report of X (no mutations)"
tier: L1
cadence: 1d
resumable: true  # set true if processing 5+ items

goal: |
  What success looks like. The loop stops when this is achieved.

allowlist: []    # empty for L1
deny:
  - comment
  - label
  - assign
  - merge
  - close
  - push
  - approve
  - force-push

exit_conditions:
  - goal_met
  - budget_exhausted
  - human_escalation

budget:
  max_tokens: 30000
  max_runs_per_day: 1
  max_wall_seconds: 600

verifier: null

request: |
  You are running the My Loop (L1 — read-only daily report).

  [Resumability instructions if resumable: true]

  Step 1 — [What to gather]
  Step 2 — [How to process]
  Step 3 — Write report to loops/my-loop/report.md

  SAFETY: [Re-state what is forbidden]

Step 3 — Run validation

python3 scripts/validate-manifests.py

Step 4 — Open a PR

PR description checklist:

## Loop Checklist
- [ ] `loops/<loop-name>/loop.yaml` created
- [ ] `name`, `goal`, and `request` fields present
- [ ] `tier` set to L1 (all new loops start at L1)
- [ ] `budget.max_tokens` set to a conservative estimate
- [ ] `exit_conditions` includes `budget_exhausted` and `goal_met`
- [ ] `deny` list is comprehensive (L1: deny all mutations)
- [ ] `resumable: true` for any loop processing 5+ items
- [ ] Validation passes: `python3 scripts/validate-manifests.py`
- [ ] `STATE.md` and `report.md` are in `.gitignore`
- [ ] No secrets or hardcoded tokens in `loop.yaml`
- [ ] Request prompt includes explicit safety constraint statements

Validation Commands Reference

Command What it checks
bash scripts/validate-skills.sh All skill directories have SKILL.md and skill.json; frontmatter is valid; no secrets
python3 scripts/validate-skills.py Same as above with more detailed error output
bash scripts/validate-loops.sh All loop.yaml files pass the JSON schema in schemas/loop.schema.json
python3 scripts/validate-manifests.py All loop manifests
python3 scripts/validate-agents.py All agent AGENT.md files have required frontmatter
bash scripts/build-catalog.sh Regenerates catalogs and verifies they match source files
python3 scripts/gen-surfaces.py --check Verifies plugin bundles match canonical sources
pytest Python CLI test suite

PR Process

  1. Fork the repository
  2. Create a branch: feat/my-skill-name
  3. Make changes following the guidelines above
  4. Run all validation commands (all must exit 0)
  5. Open a PR with the appropriate checklist from this guide in the description
  6. A maintainer will review and request changes or merge

Commit message format

Follow Conventional Commits:

feat(delivery): add gh-address-comments skill
fix(oss-triage): correct deny list in request.md
docs(readme): expand MCP templates section
chore(catalogs): regenerate after adding security-sweep loop

Format: <type>(<scope>): <short imperative description>

Types: feat, fix, docs, chore, schema, refactor, test

Code of Conduct

This project follows the Contributor Covenant Code of Conduct. Report unacceptable behavior by opening a private GitHub security advisory or contacting the maintainer directly.

Getting help

  • GitHub Discussions — for questions about skill design, compatibility, or architecture
  • GitHub Issues — for bugs or feature requests
  • SECURITY.md — for security issues

agent-toolkit

Getting Started

Reference

Compiler

Clone this wiki locally