RepoCanon

Generate repo-specific AI context for Codex, Claude Code, Copilot, and Cursor.

Turn any repository into canonical AI-readable project context.

RepoCanon is a Python CLI that analyzes a local codebase and generates project-specific instruction files for AI coding tools from a single internal repo model.

Instead of manually maintaining separate context for different tools, RepoCanon infers your repo’s structure, commands, conventions, and boundaries, then generates outputs such as:

• AGENTS.md
• CLAUDE.md
• Copilot repository instructions
• Cursor project rules

The goal is simple: make AI coding tools behave like they already understand your repo.

Why RepoCanon

AI coding tools are useful, but they usually guess:

• where things live
• how the repo is structured
• which commands to run
• what patterns are preferred
• what boundaries should not be crossed

RepoCanon reduces that guesswork by turning repo-specific knowledge into maintainable instruction files.

What it does

RepoCanon:

• analyzes a local repository
• detects languages, frameworks, commands, and topology
• infers conventions and architectural boundaries
• builds a normalized project model
• generates tool-specific AI context files from that model

RepoCanon is deterministic-first. It does not require an LLM to work.

Supported targets

• Codex via AGENTS.md
• Claude Code via CLAUDE.md
• GitHub Copilot via .github/copilot-instructions.md (and optional path-scoped files)
• Cursor via .cursor/rules/*.mdc

Installation

pip install repocanon

Requires Python 3.11+.

Quickstart

# 1. Analyze the current repo and persist a normalized model.
repocanon analyze .

# 2. Inspect what was inferred and how confident RepoCanon is.
repocanon audit .

# 3. Preview generated outputs without touching the filesystem.
repocanon preview .

# 4. Write the generated files into the repo (defaults to all targets).
repocanon generate .

You can also generate one or more specific targets:

repocanon generate . -t agents
repocanon generate . -t claude -t copilot
repocanon list-targets         # see what's available

Need machine-readable output? Append --json to analyze, audit, or diff. Need to undo? repocanon clean . (or scope it with -t copilot) removes only RepoCanon-authored files (those that carry the generator header marker).

Example outputs

A real run produces files like:

AGENTS.md
CLAUDE.md
.cursor/rules/project-overview.mdc
.cursor/rules/commands-and-validation.mdc
.cursor/rules/code-style-and-conventions.mdc
.cursor/rules/architecture-boundaries.mdc
.github/copilot-instructions.md
.github/instructions/tests.instructions.md

See docs/samples/ for sample generated files from the bundled fixture repos.

How it works

RepoCanon has three layers:

1. Repo analysis

It scans the local repo and extracts:

• languages
• frameworks
• package managers
• commands
• configs
• directory structure
• file patterns

2. Convention inference

It infers patterns such as:

• test layout (centralized vs colocated)
• frontend/backend split
• monorepo structure (apps/packages/libs/services)
• architectural boundaries
• naming conventions
• preferred libraries
• common anti-pattern risks (e.g. editing existing migrations)

3. Target generation

It maps one normalized project model into tool-specific outputs.

That means the same repo understanding can be reused across multiple AI coding tools.

Design principles

• deterministic first
• local-first (no telemetry, no network calls)
• tool-agnostic core
• small, readable outputs
• no generic filler — every section is grounded in repo facts
• explicit uncertainty when confidence is low
• human-editable generated files (sections between <!-- repocanon:manual:* --> markers survive regeneration)

Commands

repocanon analyze [PATH]

Analyze the repository and write a normalized model to:

.repocanon/project-model.json

repocanon generate [PATH]

Generate output for the selected targets (defaults to all).

Useful flags:

• -t, --target — repeat to pick targets (agents, claude, copilot, cursor, or all)
• --dry-run — render and report without writing or persisting the model
• --output-dir — write into a sibling directory (path-traversal-safe)
• --force — replace files even if RepoCanon would otherwise preserve manual edits

repocanon preview [PATH]

Print generated output to the terminal without writing files. Same -t flag as generate.

repocanon list-targets

List every target the current build supports.

repocanon clean [PATH]

Delete RepoCanon-authored files for the selected targets. Files without the RepoCanon header marker (i.e. anything you wrote yourself) are skipped. Pair with --dry-run to see what would happen.

repocanon audit [PATH]

Show inferred conventions, rationale, and confidence levels. Pass --json to emit the audit as JSON for piping into other tools.

repocanon diff [PATH]

Compare the current repo scan with the saved model and report meaningful changes (specific commands added/removed, package list deltas, structural fingerprint changes). Pass --json for a machine-readable diff.

repocanon init [PATH]

Create a local RepoCanon config file at .repocanon/config.toml.

Configuration

RepoCanon stores project config in:

.repocanon/config.toml

Example:

[project]
name = "my-repo"

[scan]
include = ["src/**", "app/**", "packages/**"]
exclude = ["node_modules/**", ".next/**", "dist/**", "build/**"]

[generate]
targets = ["agents", "claude", "copilot", "cursor"]
safe_overwrite = true

Architecture overview

repocanon/
├── analyzer/    # deterministic repo scanning + inference
├── models/      # Pydantic v2 project model
├── generators/  # one module per AI target
├── output/      # writers, preview, diff
├── report/      # audit + summary tables
└── cli.py       # Typer entry point

The analyzer is a straight pipeline: file inventory → manifest parsing → framework/package-manager detection → command extraction → topology + conventions → final ProjectModel. Generators only consume that model — they never touch the filesystem.

Limitations

RepoCanon is inference-based. It can detect a lot, but not everything.

It may be less accurate when:

• the repo is highly unconventional
• conventions are implicit rather than visible in files
• commands live outside standard manifests
• architecture is unclear from structure alone

When confidence is low, RepoCanon says so rather than inventing detail.

Roadmap

• more framework detectors (Django, Rails, .NET, Spring, etc.)
• stronger monorepo inference (Bazel, Pants, Nx graph)
• better path-scoped output generation
• safer merge/update behavior for edited generated files
• optional LLM-assisted summarization (off by default)
• additional target formats

Why not just write these files manually?

You can. But in practice:

• they drift out of date
• they are inconsistent across tools
• they are often generic
• they rarely reflect the actual repo structure

RepoCanon keeps those files grounded in the codebase.

How RepoCanon maps one repo model to multiple AI coding tools

RepoCanon is intentionally a many-to-one-to-many pipeline:

repo files ─┐                              ┌─► AGENTS.md            (Codex)
            ├─► analyzer ─► ProjectModel ──┼─► CLAUDE.md            (Claude Code)
manifests  ─┘                              ├─► copilot-instructions (Copilot)
                                           └─► .cursor/rules/*.mdc  (Cursor)

The analyzer collapses everything it sees into a single normalized ProjectModel (Pydantic v2). That model is the only thing target generators read; they never touch the filesystem. This gives RepoCanon two important properties:

1. One source of truth. Languages, frameworks, commands, conventions, anti-patterns, and architecture boundaries all live in one place. Adding a new target means writing a new generator that consumes the same model — not re-implementing detection.
2. Idiomatic outputs per tool. Each generator picks the parts of the model that make sense for its target and renders them in that tool's idiom: a verbose AGENTS.md for Codex, a terse CLAUDE.md for Claude Code, a repo-wide instructions file (plus optional path-scoped ones) for Copilot, and a small set of focused .mdc rule files for Cursor.

The same model also powers audit, diff, and preview, so you can verify what RepoCanon inferred before any file is written.

Contributing

Contributions are welcome. See CONTRIBUTING.md for local setup, tests, and development workflow.

License

MIT