CLI Plugins

OpenAPI mindset for CLIs: crawl real --help output, build structured CLI maps, and generate Claude Code plugins with reliable command knowledge.

---

TL;DR

• Input: any CLI binary available in your environment (docker, git, gh, uv, pnpm, etc.)
• Phase 1: cli-crawler parses help output into output/<cli>.json
• Phase 2: generate-plugin converts that JSON into plugins/cli-<name>/
• Result: Claude Code gets concrete, tool-specific command intelligence instead of guessing flags

---

Why This Exists

LLMs are strong at reasoning, but weak on precise, current CLI syntax unless the exact tool/version is in context.

Typical failure modes when relying on model memory alone:

• hallucinated flags
• wrong argument order
• outdated commands after CLI updates
• missing required options hidden in subcommand help

This project creates a deterministic bridge between real CLI help and AI plugin context.

---

Why Not Just --help?

--help is necessary, but not sufficient as a direct context source for agents:

Approach	Works for humans	Works for agents at scale	Main issue
Manual copy/paste from --help	Yes	No	token-heavy, inconsistent, stale quickly
man/online docs	Yes	Sometimes	fragmented formats, weak machine structure
Raw CLI help in prompt	Sometimes	Poorly	noisy, hard to navigate, expensive context
CLI Plugins (this project)	Yes	Yes	structured, compact, regenerable

---

Real Differentiators

1. Format-agnostic parsing Parses multiple help styles (Cobra, Click, Rich-Click, Git/manpage, POSIX-like patterns) with deterministic heuristics.

2. Progressive disclosure by default Generates compact SKILL.md plus on-demand references (commands.md, examples.md) to reduce context pressure.

3. Safety-first help detection Includes protections such as auth-required precedence and non-mutating fallback policy for subcommand probing.

4. Cross-platform resilience (Windows/WSL) Supports executable-name canonicalization (e.g., .exe identity handling) while preserving runnable invocation.

5. Global-only CLI quality fallback When a CLI has no subcommands, plugin output still provides meaningful usage examples and stable version labeling.

6. Inventory drift visibility config-audit compares config.yaml, output/*.json, and plugins/cli-* to highlight stale/missing artifacts.

7. Idempotent generation Re-runs overwrite generated outputs cleanly and predictably.

8. No runtime external dependencies Core runtime (cli-crawler, generate-plugin, config-audit) uses project code and stdlib-friendly implementation.

---

How It Works

CLI binary
   -> cli-crawler
   -> output/<cli>.json (+ optional output/<cli>.raw.json)
   -> generate-plugin
   -> plugins/cli-<slug>/

Main Commands

• cli-crawler: crawl and parse CLI help output
• generate-plugin: build Claude plugin from crawler JSON
• config-audit: detect inventory drift and suggest minimal config overrides

---

Quick Start

1. Clone and Install

git clone https://github.com/nsalvacao/cli-plugins.git
cd cli-plugins

# Option A (recommended)
uv sync

# Option B
# python3 -m venv .venv-wsl
# source .venv-wsl/bin/activate
# pip install -e ".[dev]"

2. Crawl a CLI

uv run cli-crawler docker -o output/docker.json --raw -v

3. Generate Plugin

uv run generate-plugin output/docker.json

Plugin output:

plugins/cli-docker/

4. Install/Use in Claude Code

cp -r plugins/cli-docker ~/.claude/plugins/cli-docker

Or test directly:

claude --plugin-dir plugins/cli-docker

---

Generated Plugin Layout

plugins/cli-my-tool/
├── .claude-plugin/
│   └── plugin.json
├── skills/
│   └── cli-my-tool/
│       ├── SKILL.md
│       └── references/
│           ├── commands.md
│           └── examples.md
├── commands/
│   └── scan-cli.md
└── scripts/
    └── rescan.sh

rescan.sh is generated per plugin and re-runs crawl + generation for that CLI.

---

Configuration Policy (config.yaml)

Use config.yaml for operational overrides, not as a full inventory catalog.

Example:

defaults:
  timeout: 5
  max_depth: 5
  max_concurrent: 5
  retry: 1
  environment: wsl
  raw_threshold: 10000

clis:
  docker:
    max_depth: 6
  kubectl:
    timeout: 8
    plugins:
      discovery_command: kubectl plugin list

Recommended minimal override fields:

• environment
• help_pattern
• max_depth
• max_concurrent
• plugins.discovery_command

---

Inventory Audit and Drift Detection

Run:

uv run config-audit \
  --config config.yaml \
  --output-dir output \
  --plugins-dir plugins \
  --report output/config-audit.json

The report highlights:

• missing_in_config
• stale_in_config
• missing_output
• missing_plugin
• suggested_minimal_overrides

---

Quality and Workflow

This repo follows a strict loop for each batch:

1. plan and define atomic tasks
2. tests first (RED)
3. implementation (GREEN)
4. full regression run (pytest tests/)
5. E2E with a new CLI
6. update tracking in specs/control files
7. atomic commit + PR + review

Common checks:

uv run ruff check .
uv run ruff format --check .
uv run pytest tests/

---

Roadmap

Planned: Grouped Plugins (Multi-CLI per Domain)

A major next step is to support grouping multiple related CLIs into a single plugin (for example, one domain plugin for several DevOps CLIs).

Expected impact:

• drastically fewer installed plugins
• simpler activation/maintenance model
• better domain-level discoverability
• lower operational overhead for large toolchains

This is planned through the group taxonomy and inference backlog (deterministic first, ML fallback), then future grouped-plugin generation.

Also Planned

• observability and confidence scoring expansion
• dashboard UI for inventory, operations, and plugin inspection
• PyPI release hardening and publish workflow

---

Contributing

See docs/CONTRIBUTING.md.

---

License

MIT

---

References

• Claude Code Plugins
• Claude Code Skills
• OpenAPI Specification