codectx

A CLI tool that compiles repository context for AI agents — ranking files by importance, compressing them into structured summaries, and emitting a single markdown document optimized for LLM reasoning.

The problem

Feeding a raw repository to an AI agent wastes context. Most files are noise — tests, docs, boilerplate, lockfiles. The signal (core architecture, key abstractions, dependency structure) is buried.

Naive approaches either blow past context limits or arbitrarily truncate. Neither helps an agent reason about a codebase.

How codectx works

codectx treats context generation as a compilation step:

1. Scans the repository and builds a dependency graph
2. Scores every file by fan-in centrality, git commit frequency, and entry-point proximity
3. Assigns tiers — top 15% get structured summaries, next 30% get signatures, rest get one-liners
4. Enforces a token budget, highest-signal files first
5. Emits a structured CONTEXT.md an agent can reason from immediately

The output is not a source dump. Core files get AST-derived structured summaries: purpose, internal dependencies, public types, and function signatures — at roughly 10% of the token cost of the raw source.

Benchmark

Tested against five well-known Python repos. Naive baseline counts all source files excluding tests, docs, and examples — the same file set codectx analyzes.

Average reduction: 76%. Every repo fits within a 128k context window. Naive baseline exceeds it for fastapi and rich.

Repos with lower reduction (fastapi, typer) contain large entry point files that codectx includes as full source by design — the CLI surface is exactly what an agent needs to see in full.

Installation

Requires Python 3.10+.

pip install codectx

pip install codectx[llm]

uv add codectx

pipx install codectx

Docker

Run codectx without installing Python locally:

docker run --rm -v /path/to/repo:/repo ghcr.io/hey-granth/codectx analyze /repo

This writes CONTEXT.md into your host repository because /repo is a bind mount.

On Linux, if you see permission errors on mounted files, run with your host UID/GID:

docker run --rm --user $(id -u):$(id -g) -v /path/to/repo:/repo ghcr.io/hey-granth/codectx analyze /repo

See the full Docker guide.

Quick start

# analyze the current repository
codectx analyze .

# adjust token budget
codectx analyze . --tokens 60000

# custom output path
codectx analyze . --output my-context.md

# machine-readable output
codectx analyze . --format json

# watch mode — regenerate on file changes
codectx watch .

# tune watch debounce window
codectx watch . --debounce 1.5

# semantic search ranking
codectx analyze . --query "authentication flows"

# LLM summaries (requires codectx[llm])
codectx analyze . --llm --llm-provider openai --llm-model gpt-4.1-mini
codectx analyze . --llm --llm-provider ollama --llm-model llama3.1 --llm-base-url http://localhost:11434/v1

# bypass manifest up-to-date checks
codectx analyze . --force

# cache management
codectx cache info .
codectx cache clear . --force
codectx cache export . --output .codectx_cache.tar.gz
codectx cache import . --input .codectx_cache.tar.gz

# task-specific ranking profiles
codectx analyze . --task architecture
codectx analyze . --task refactor
codectx analyze . --task debug
codectx analyze . --task feature

Output

CONTEXT.md is structured with fixed sections in a consistent order:

A structured summary for a core module looks like this:

### `src/myapp/core/engine.py`

**Purpose:** Main execution engine for request processing.

**Depends on:** `core.models`, `core.cache`, `utils.retry`

**Types:**
- `Engine` — methods: process, shutdown, health_check

**Functions:**
- `def process(request: Request, timeout: float = 30.0) -> Response`
  Handle a single request through the full processing pipeline.
- `def shutdown(graceful: bool = True) -> None`
  Stop accepting new requests and drain the queue.

Configuration

Create .codectx.toml in your project root:

[codectx]
token_budget = 120000
output_file = "CONTEXT.md"
extra_ignore = ["**/generated/**", "*.draft.py"]

CLI flags override config file values. Supported task profiles for --task: default, debug, feature, architecture, refactor.

Supported languages

Python, TypeScript, JavaScript, Go, Rust, Java, C, C++, Ruby. Language detection is automatic from file extension via tree-sitter.

How it works

Repository
    ↓
[Walker]      → scan files, apply .gitignore + .ctxignore
    ↓
[Parser]      → extract imports and symbols via tree-sitter
    ↓
[Graph]       → build dependency graph with rustworkx
    ↓
[Ranker]      → score files by fan-in, git frequency, entry proximity
    ↓
[Compressor]  → assign tiers, enforce token budget
    ↓
[Summarizer]  → generate AST-driven structured summaries
    ↓
[Formatter]   → emit structured markdown
    ↓
CONTEXT.md

See ARCHITECTURE.md for detail on each stage. See DECISIONS.md for reasoning behind key design choices.

Caching

codectx persists cache data under your OS cache directory (default: ~/.cache/codectx, or $XDG_CACHE_HOME/codectx when set), scoped by repository hash.

• Parse/cache artifacts are reused across runs.
• A run-level manifest tracks options and file hashes to skip no-op analyze runs.
• Semantic embedding cache is stored under the repo-scoped cache path and invalidated by file-content hash changes.

Development setup

git clone https://github.com/hey-granth/codectx.git
cd codectx
uv sync

Or with pip:

pip install -e ".[dev]"

Tests

pytest
pytest --cov=src/codectx   # with coverage

Type checking and linting

mypy src/codectx --strict
ruff check src tests
ruff format src tests

Contributing

Issues and pull requests are welcome. The project prioritizes correctness, performance, and maintainability. File an issue before starting significant work.

License

MIT. See LICENSE.