repoforge-ai 0.3.0

Generate technical docs and AI agent skills (SKILL.md, AGENT.md) from any codebase — model-agnostic, monorepo-aware

⚒ RepoForge

AI-powered code analysis tool that generates technical documentation, AI agent skills, security scans, code graphs, and LLM-ready exports from any codebase — works with any LLM.

---

What it does

RepoForge analyzes your codebase and generates multiple types of output:

1. repoforge docs — Technical Documentation

Generates a complete Docsify-ready documentation site adapted to your project type:

Project type	Specific chapters
Web service	Data Models · API Reference
Frontend SPA	Components · State Management
CLI tool	Commands · Configuration
Data science	Data Pipeline · Models & Training · Experiments
Library/SDK	Public API · Integration Guide
Mobile app	Screens & Navigation · Native Integrations
Infra/DevOps	Resources · Variables · Deployment Guide
Monorepo	Global chapters + per-layer subdocs (each layer gets its own set)

All types share: Overview · Quick Start · Architecture · Core Mechanisms · Dev Guide

Output is a docs/ folder ready for GitHub Pages — zero extra config.

2. repoforge skills — AI Agent Skills

Generates SKILL.md and AGENT.md compatible with:

• Claude Code (.claude/skills/, .claude/agents/)
• OpenCode (mirrored to .opencode/)
• Cursor (.cursor/rules/*.mdc)
• Codex (AGENTS.md)
• Gemini CLI (GEMINI.md)
• GitHub Copilot (.github/copilot-instructions.md)
• agent-teams-lite (skill registry at .atl/skill-registry.md)
• Gentleman-Skills format (YAML frontmatter, Trigger:, Critical Patterns)

---

What's New in v0.3

#	Feature	Command / Flag	Description
1	LLM Export	repoforge export	Flatten repo into single LLM-optimized file (markdown or XML)
2	Complexity routing	--complexity auto|small|medium|large	Auto-routes generation depth by repo size
3	Hooks generation	--with-hooks	Generates HOOKS.md with recommended Claude Code hooks
4	Quality scorer	repoforge score	Standalone scorer with 7 dimensions (no API key)
5	Multi-tool targets	--targets claude,cursor,codex,...	Generate for 6 AI tools at once
6	Progressive disclosure	--disclosure tiered|full	L1/L2/L3 tier markers + DISCOVERY_INDEX.md
7	Token compressor	repoforge compress	Reduce token count by 50-75% (no API key)
8	Security scanner	repoforge scan	37 rules in 5 categories (no API key)
9	Plugin hierarchy	--plugin	Generates plugin.json + commands/ directory
10	Code graph	repoforge graph	Dependency graph with blast radius analysis

---

Installation

pip install repoforge-ai

Note: The CLI command is still repoforge after installation. The PyPI package name is repoforge-ai (the name repoforge was already taken).

Recommended: install ripgrep for 10-100x faster scanning:

brew install ripgrep          # macOS
sudo apt install ripgrep      # Ubuntu/Debian
scoop install ripgrep         # Windows

---

Quick start

# Generate docs (auto-detects language, Docsify-ready)
repoforge docs -w /path/to/repo --lang Spanish -o docs

# Preview locally (opens browser automatically)
repoforge docs -w . --serve

# Generate SKILL.md + AGENT.md for Claude Code
repoforge skills -w /path/to/repo

# Generate for ALL AI tools at once
repoforge skills -w /path/to/repo --targets all

# Generate + score + scan + compress in one shot
repoforge skills -w /path/to/repo --score --scan --compress

# Flatten repo for LLM context
repoforge export -w /path/to/repo -o context.md

# Build dependency graph
repoforge graph -w /path/to/repo --format mermaid

# Open skills browser
repoforge skills --serve-only

# See what would be generated (no LLM calls, free)
repoforge docs --dry-run
repoforge skills --dry-run

---

Model setup

RepoForge auto-detects your API key from env vars and picks the best available model.

GitHub Models — free with GitHub Copilot ⭐

export GITHUB_TOKEN=$(gh auth token)
repoforge docs -w . --model github/gpt-4o-mini --lang Spanish

Groq — free tier, very fast

export GROQ_API_KEY=gsk_...
repoforge docs -w . --model groq/llama-3.3-70b-versatile

Ollama — 100% local, free

ollama pull qwen2.5-coder:14b
repoforge docs -w . --model ollama/qwen2.5-coder:14b

Claude Haiku — cheap, ~$0.05/run

export ANTHROPIC_API_KEY=sk-ant-...
repoforge docs -w . --model claude-haiku-3-5

OpenAI

export OPENAI_API_KEY=sk-...
repoforge docs -w . --model gpt-4o-mini

---

docs command

repoforge docs [OPTIONS]

  -w, --working-dir DIR     Repo to analyze  [default: .]
  -o, --output-dir DIR      Output directory  [default: docs]
  --model TEXT              LLM model (auto from env if not set)
  --lang LANGUAGE           Documentation language  [default: English]
                            English|Spanish|French|German|Portuguese|
                            Chinese|Japanese|Korean|Russian|Italian|Dutch
  --name TEXT               Project name (auto-detected by default)
  --complexity LEVEL        Override repo complexity  [default: auto]
                            auto|small|medium|large
  --theme [vue|dark|buble]  Docsify theme  [default: vue]
  --serve                   Serve docs after generating (opens browser)
  --serve-only              Skip generation, serve existing docs
  --port INT                Server port  [default: 8000]
  --dry-run                 Plan only, no LLM calls, no files written
  -q, --quiet               Suppress progress output

Complexity levels

Level	Files	Effect
auto	—	Auto-detected from file count + layer count
small	≤20	Fewer chapters, detailed per-file coverage
medium	≤200	Standard depth (default behavior)
large	>200	Concise chapters, more architectural focus

Publish to GitHub Pages

RepoForge can publish docs automatically with GitHub Actions.

Version pinning for workflow stability

When using the RepoForge GitHub Action, pin to a tagged release instead of @main:

uses: JNZader/repoforge@v0.2.0

This keeps consumer workflows stable and reproducible. Update the tag explicitly when you want new behavior.

Safe defaults (recommended):

1. Add .github/workflows/docs.yml to your repository.
2. Push to main.
3. By default, it runs in generate-only mode (deploy_mode=none) so it does not touch an existing Pages site.

To enable deploy (explicit opt-in):

1. Set repository variable REPOFORGE_DOCS_DEPLOY_MODE to one of:
   • auto (if no live site -> deploy root, if live site -> deploy subpath)
   • main (force root deploy; may replace existing site)
   • subpath (deploy to /repoforge/ on gh-pages, preserving existing files)
2. Set repository variable REPOFORGE_DOCS_CONFIRM_DEPLOY=true.
3. Optional: set REPOFORGE_DOCS_SUBPATH_PREFIX (default: repoforge).

You can also run it manually from Actions (workflow_dispatch) with deploy_mode, confirm_deploy, and subpath_prefix inputs.

Note: subpath preservation uses gh-pages branch deploy. If your repo uses Pages build_type=workflow, the workflow will fall back to generate-only for safety.

Required Pages configuration by deploy mode:

deploy_mode	Deployment mechanism	Required Pages setting
none	Generate only (no publish)	Any
main	actions/deploy-pages (artifact)	Build and deployment: GitHub Actions
subpath	peaceiris/actions-gh-pages (branch, keep_files)	Deploy from a branch → gh-pages + / (root)
auto	Chooses main or subpath	Must match chosen target (main => GitHub Actions, subpath => gh-pages)

If auto selects subpath but Pages is configured as GitHub Actions, the subpath branch publish may succeed but not be publicly visible.

After deploy, your docs are available at:

https://<your-user>.github.io/<your-repo>/

If deployed in subpath mode:

https://<your-user>.github.io/<your-repo>/<subpath-prefix>/

Example: adding docs to a repo with existing Pages

Your repo already has a live site at https://youruser.github.io/yourrepo/ and you want to add RepoForge docs without breaking it.

# 1. Copy the docs workflow to your repo
cp .github/workflows/docs.yml <your-repo>/.github/workflows/docs.yml

# 2. Set repo variables for safe subpath deploy
gh variable set REPOFORGE_DOCS_DEPLOY_MODE --body "auto" --repo youruser/yourrepo
gh variable set REPOFORGE_DOCS_CONFIRM_DEPLOY --body "true" --repo youruser/yourrepo
gh variable set REPOFORGE_DOCS_SUBPATH_PREFIX --body "docs" --repo youruser/yourrepo

# 3. Set Pages source to gh-pages branch (required for subpath mode)
#    Settings → Pages → Build and deployment → Deploy from a branch → gh-pages / (root)

# 4. Add the GH_MODELS_TOKEN secret (PAT with models:read scope)
gh secret set GH_MODELS_TOKEN --repo youruser/yourrepo

# 5. Push and let the workflow run
git add .github/workflows/docs.yml && git commit -m "ci: add repoforge docs" && git push

Result:

• Your existing site stays at https://youruser.github.io/yourrepo/ (unchanged).
• RepoForge docs appear at https://youruser.github.io/yourrepo/docs/.

Manual flow

Still supported if you prefer not to use the workflow:

repoforge docs -w . -o docs --lang English
git add docs/ && git commit -m "docs: generate documentation"
git push
# Settings → Pages → Source: /docs on main branch

---

skills command

repoforge skills [OPTIONS]

  -w, --working-dir DIR     Repo to analyze  [default: .]
  -o, --output-dir DIR      Output directory  [default: .claude]
  --model TEXT              LLM model
  --complexity LEVEL        Override repo complexity  [default: auto]
                            auto|small|medium|large
  --targets TARGETS         Comma-separated output targets  [default: claude,opencode]
                            claude|opencode|cursor|codex|gemini|copilot|all
  --disclosure MODE         Skill output mode  [default: tiered]
                            tiered (L1/L2/L3 markers) | full (no markers)
  --with-hooks              Generate HOOKS.md with recommended Claude Code hooks
  --plugin                  Generate plugin.json + commands/ hierarchy
  --score                   After generation, score quality of SKILL.md files
  --compress                After generation, compress skills to reduce tokens
  --aggressive              Use aggressive compression (with --compress)
  --scan                    After generation, run security scanner
  --no-opencode             Skip mirroring to .opencode/
  --serve                   Open skills browser after generating
  --serve-only              Skip generation, open existing skills browser
  --port INT                Server port  [default: 8765]
  --dry-run                 Plan only, no LLM calls
  -q, --quiet               Suppress progress output

Multi-tool output

Target	Output location	Format
claude	.claude/skills/, .claude/agents/	SKILL.md + AGENT.md
opencode	.opencode/	Mirror of .claude/
cursor	.cursor/rules/*.mdc	Cursor rules format
codex	AGENTS.md (project root)	Single consolidated file
gemini	GEMINI.md (project root)	Gemini CLI instructions
copilot	.github/copilot-instructions.md	Copilot instructions

Output layout

.claude/
├── skills/
│   ├── backend/
│   │   ├── SKILL.md              ← layer-level skill
│   │   ├── auth/SKILL.md         ← per-module skill
│   │   └── reports/SKILL.md
│   └── frontend/
│       ├── SKILL.md
│       └── useGEELayers/SKILL.md
├── agents/
│   ├── orchestrator/AGENT.md     ← delegate-only orchestrator
│   ├── backend-agent/AGENT.md
│   └── frontend-agent/AGENT.md
├── commands/                     ← (with --plugin)
│   ├── review.md
│   └── deploy.md
├── plugin.json                   ← (with --plugin)
├── HOOKS.md                      ← (with --with-hooks)
├── DISCOVERY_INDEX.md            ← (with --disclosure tiered)
└── SKILLS_INDEX.md

.opencode/                        ← identical mirror
.atl/skill-registry.md            ← agent-teams-lite registry
.cursor/rules/*.mdc               ← (with --targets cursor)
AGENTS.md                         ← (with --targets codex)
GEMINI.md                         ← (with --targets gemini)
.github/copilot-instructions.md   ← (with --targets copilot)

---

export command

Flatten a repo into a single LLM-optimized file. No API key needed.

repoforge export [OPTIONS]

  -w, --working-dir DIR     Repo to analyze  [default: .]
  -o, --output FILE         Output file (default: stdout)
  --max-tokens INT          Token budget limit (prioritizes important files)
  --no-contents             Skip file contents — tree + definitions only
  --format FORMAT           Output format  [default: markdown]
                            markdown|xml (CXML-style)
  -q, --quiet               Suppress progress output

repoforge export -w .                         # print to stdout
repoforge export -w . -o context.md           # save to file
repoforge export -w . --max-tokens 100000     # limit output size
repoforge export -w . --no-contents           # tree + definitions only
repoforge export -w . --format xml            # XML output (CXML-style)

---

score command

Score quality of generated SKILL.md files across 7 dimensions. No API key needed.

Dimensions: completeness, clarity, specificity, examples, format, safety, agent readiness.

repoforge score [OPTIONS]

  -w, --working-dir DIR     Repo to analyze  [default: .]
  -d, --skills-dir DIR      Skills directory  [default: .claude/skills/]
  --format FORMAT           Output format  [default: table]
                            table|json|markdown
  --min-score FLOAT         Minimum score (0.0-1.0). Exit 1 if below.
  -q, --quiet               Suppress progress output

repoforge score -w .                     # score with table output
repoforge score -w . --format json       # JSON output
repoforge score -w . --min-score 0.7     # fail if any skill < 70%

---

scan command

Security scanner with 37 rules in 5 categories. No API key needed.

Categories: prompt injection, hardcoded secrets, PII exposure, destructive commands, unsafe code patterns.

Context-aware: patterns inside Anti-Patterns sections are downgraded to INFO (not false positives).

repoforge scan [OPTIONS]

  -w, --workspace DIR       Repo root  [default: .]
  --target-dir DIR          Specific directory to scan (default: auto-detect)
  --format FORMAT           Output format  [default: table]
                            table|json|markdown
  --allowlist IDS           Comma-separated rule IDs to skip (e.g. SEC-020,SEC-022)
  --fail-on SEVERITY        Exit 1 if findings at or above this level
                            critical|high|medium|low
  -q, --quiet               Suppress progress output

repoforge scan -w .                              # scan with table output
repoforge scan -w . --fail-on critical           # CI gate: fail on critical
repoforge scan -w . --allowlist SEC-020,SEC-022  # skip specific rules

---

compress command

Token-optimize generated .md files with deterministic multi-pass compression (50-75% reduction). No API key needed.

Passes: whitespace normalization, filler phrase removal, table compaction, code block cleanup, bullet consolidation, abbreviations (aggressive only).

repoforge compress [OPTIONS]

  -w, --workspace DIR       Repo root  [default: .]
  --target-dir DIR          Directory to compress  [default: .claude/skills/]
  --aggressive              Use abbreviations (function→fn, configuration→config)
  --dry-run                 Show compression stats without modifying files
  -q, --quiet               Suppress progress output

repoforge compress -w .                       # compress .claude/skills/
repoforge compress -w . --aggressive          # also abbreviate words
repoforge compress -w . --dry-run             # show stats only

---

graph command

Build a code knowledge graph from scanner data. No API key needed.

Uses import/export name matching — no tree-sitter needed.

repoforge graph [OPTIONS]

  -w, --workspace DIR       Repo root  [default: .]
  -o, --output FILE         Output file (default: stdout)
  --format FORMAT           Output format  [default: summary]
                            mermaid|json|dot|summary
  --blast-radius MODULE     Show blast radius for a specific module
  -q, --quiet               Suppress progress output

repoforge graph -w .                              # summary to stdout
repoforge graph -w . --format mermaid             # Mermaid diagram
repoforge graph -w . --format json -o graph.json  # D3/Cytoscape-compatible
repoforge graph -w . --blast-radius src/auth.py   # who breaks if auth changes?

---

Monorepo support

Auto-detects layers from directory structure. Docs are generated hierarchically:

docs/
├── index.md              ← monorepo home + layer links
├── 01-overview.md        ← global tech stack, all layers
├── 03-architecture.md    ← how layers interact + Mermaid diagram
├── 06b-service-map.md    ← inter-service contracts
├── frontend/             ← classified as frontend_app
│   ├── index.md
│   ├── 05-components.md
│   └── 06-state.md
└── backend/              ← classified as web_service
    ├── index.md
    ├── 05-data-models.md
    └── 06-api-reference.md

---

repoforge.yaml — per-repo config

# repoforge.yaml (place in repo root)

# Override project name (default: from package.json / pyproject.toml)
project_name: "My App"

# Force project type (default: auto-detected)
# web_service | frontend_app | cli_tool | library_sdk | data_science
# mobile_app | desktop_app | infra_devops | monorepo | generic
project_type: web_service

# Override layer detection (default: auto from directory names)
layers:
  frontend: apps/web
  backend: apps/api
  shared: packages/shared

# Default language for docs
language: Spanish

# Default model
model: github/gpt-4o-mini

# Override auto-detected complexity (affects generation depth)
# auto | small | medium | large
complexity: auto

# Output targets for multi-tool support
# Default: [claude, opencode]
# Valid: claude, opencode, cursor, codex, gemini, copilot
targets: [claude, opencode]

# Generate HOOKS.md with recommended Claude Code hooks
generate_hooks: false

# Generate plugin.json + commands/ hierarchy
generate_plugin: false

---

Python API

from repoforge import (
    # Skills + agents
    generate_artifacts,
    # Documentation
    generate_docs,
    # LLM export
    export_llm_view,
    # Quality scoring
    SkillScorer, SkillScore,
    # Multi-tool adapters
    adapt_for_cursor, adapt_for_codex, adapt_for_gemini, adapt_for_copilot,
    resolve_targets, ALL_TARGETS,
    # Progressive disclosure
    extract_tier, build_discovery_index, estimate_tokens,
    # Token compression
    SkillCompressor, CompressionResult, compress_file, compress_directory,
    # Security scanning
    SecurityScanner, ScanResult, Finding, Severity, scan_generated_output,
    # Plugin hierarchy
    Command, PluginManifest, build_commands, build_plugin_manifest,
    manifest_to_json, manifest_to_markdown, write_plugin,
    # Code graph
    CodeGraph, Node, Edge, build_graph, build_graph_from_workspace,
)

# Skills + agents
generate_artifacts(
    working_dir="/path/to/repo",
    output_dir=".claude",
    model="github/gpt-4o-mini",
    also_opencode=True,
    complexity="auto",
    with_hooks=True,
    with_plugin=True,
    targets="claude,cursor,codex",
    disclosure="tiered",
    compress=True,
)

# Documentation
generate_docs(
    working_dir="/path/to/repo",
    output_dir="docs",
    model="claude-haiku-3-5",
    language="Spanish",
    complexity="auto",
)

# LLM export (no API key needed)
result = export_llm_view(
    workspace="/path/to/repo",
    output_path="context.md",
    max_tokens=100000,
    fmt="markdown",
)

# Quality scoring (no API key needed)
scorer = SkillScorer()
scores = scorer.score_directory(".claude/skills")
print(scorer.report(scores, fmt="table"))

# Security scanning (no API key needed)
scan_result = scan_generated_output("/path/to/repo")
scanner = SecurityScanner()
print(scanner.report(scan_result, fmt="table"))

# Code graph (no API key needed)
graph = build_graph_from_workspace("/path/to/repo")
print(graph.to_mermaid())
print(graph.summary())
affected = graph.get_blast_radius("src/auth.py")

---

How it works

1. SCAN     (free, no LLM) — detect layers, extract exports/imports, detect stack
2. PLAN     (free, no LLM) — select chapters by project type, rank modules, route by complexity
3. GENERATE (LLM calls)    — one call per chapter or skill
4. ADAPT    (free, no LLM) — convert to target formats (Cursor, Codex, Gemini, Copilot)
5. ENRICH   (free, no LLM) — hooks, plugin manifest, disclosure index, compression, security scan
6. WRITE                   — Docsify-ready docs/ or .claude/ skills + multi-tool output

The LLM only generates text. All structural analysis, scoring, compression, scanning, and graph building is deterministic.

---

Cost estimate

Model	~cost for medium repo
GitHub Models (Copilot)	free
Groq	free (rate-limited)
Ollama (local)	free
Claude Haiku 3.5	~$0.05
GPT-4o-mini	~$0.04
Claude Sonnet	~$0.50

---

Supported stacks

Language-agnostic — tested with Python, TypeScript, JavaScript, Go, Java, Kotlin, Rust, Ruby, PHP, and any monorepo combination.

---

License

MIT

---

v0.3.0 Inspirations

• CodeViewX — original export concept
• Gentleman-Skills — skill format spec
• agent-teams-lite — agent orchestration pattern
• repomix / rendergit — LLM export and XML format inspiration
• aider — RepoMap-based code graph approach
• semgrep — security scanning rule categories