GitHub - pwnaxe/squick - 惯性聚合

Pre-computed, LLM-targeted code context for AI coding agents.

Squick scans a codebase, extracts structural facts (call graph, imports, symbols, framework markers, HTTP endpoints, content-type schemas, route patterns, dependency manifests), and emits a small set of artifacts that AI agents read instead of re-scanning the repository on every prompt.

Why

AI coding agents currently spend tokens "looking around" the repository before answering even simple questions. Squick inverts that cost: do the analysis once at file save, save tokens on every prompt thereafter.

Install

# npm (recommended for AI-agent users - works with `npx -y` too)
npm i -g @hubhorizonllc/squick

# PyPI
pip install squick

# crates.io
cargo install squick-cli

# Direct binary (Unix)
curl --proto '=https' --tlsv1.2 -LsSf \
  https://github.com/pwnaxe/squick/releases/latest/download/squick-cli-installer.sh | sh

# Direct binary (Windows)
irm https://github.com/pwnaxe/squick/releases/latest/download/squick-cli-installer.ps1 | iex

All channels install the same binary, exposed on PATH as squick.

Quick start

squick scan ./your-project

Writes a small set of artifacts to .squick/:

conventions.md - detected stack, library choices, repository layout, API surface. Attach this to your AI chat when asking about architecture or library usage.
schemas.md - extracted data schemas (Strapi content types) and HTTP endpoints. Attach this to your AI chat for backend, data, or API questions.
context.md - tiny index pointing at the two files above.

For programmatic consumers (MCP servers, scripts) add --full:

squick scan ./your-project --full

This additionally writes context.ndjson (one JSON fact per line) and graph.txt (subject-predicate-object triples).

What gets extracted

Structure (Tree-sitter): symbols, imports, JSX components, doc comments, references.
Heuristics: function-name verbs, suffixes, Python dunders, framework markers.
Dictionaries (YAML): conventional routes, file roles, framework affinity.
Manifests: package.json, pyproject.toml - identity, dependencies, scripts, framework detection.
Endpoints: FastAPI/Flask decorators, Django urlpatterns, Express member-calls, Next.js App Router file layout.
Data schemas: Strapi content types (kind, names, attributes, relations).

Supported languages

TypeScript / TSX / JavaScript / JSX / Python.

Supported frameworks (out of the box)

Backend: Strapi, Django, Django REST Framework, FastAPI, Flask, Express, Koa, Fastify, NestJS, Sanity, Payload CMS, WordPress (file roles).

Frontend: Next.js (App Router + Pages Router), React, Tailwind.

Add a YAML file under dictionaries/frameworks/ to teach Squick a new framework. No Rust changes required for most additions.

CLI

squick scan [root]                One-shot scan into .squick/
  --format markdown|json          Output format (default: markdown)
  --out PATH                      Override output path
  --dict-dir PATH                 Override dictionary directory
  --include GLOB                  Repeatable. Only scan matching paths
  --exclude GLOB                  Repeatable. Skip matching paths
  --no-schemas                    Skip .squick/schemas.md
  --full                          Also emit context.ndjson + graph.txt

squick watch [root]               Re-scan on file save (same flags)
squick init [root]                Create empty .squick/ directory
squick mcp                        Start an MCP server on stdio
  --dict-dir PATH                 Override dictionary directory

MCP server (for AI agents)

Squick speaks the Model Context Protocol on stdio. Any MCP-aware host can invoke its tools to pull project context on demand rather than re-reading source files.

Tools exposed:

squick_scan(root) - the conventions summary (most useful default).
squick_get_conventions(root) - explicit conventions content.
squick_get_schemas(root) - data schemas as JSON.
squick_get_endpoints(root) - HTTP endpoints as JSON.
squick_get_file_context(root, file) - context for one file only.
squick_get_ndjson(root) - full project context as NDJSON.
squick_get_graph(root) - RDF-style triples for graph traversal.

Configure Claude Code

{
  "mcpServers": {
    "squick": {
      "command": "squick",
      "args": ["mcp"]
    }
  }
}

For zero-install invocation (no global package needed):

{
  "mcpServers": {
    "squick": {
      "command": "npx",
      "args": ["-y", "@hubhorizonllc/squick", "mcp"]
    }
  }
}

The same shape works for Cursor (.cursor/mcp.json), Cline, Continue, and any other MCP-aware host.

Dictionary format

Dictionaries are YAML files under dictionaries/<category>/<name>.yaml:

name: frameworks/example
description: One-line description of what this dictionary recognises.
entries:
  - pattern: "models.py"
    match: filename
    tag: data-models
    confidence: high
    kind: literal
    note: "Optional context for reviewers."

Globs accept * and ?. Regex uses Rust syntax. Literal matches are case-insensitive.

Workspace layout

squick/
  crates/
    squick-core/       Types, scanner, AST extraction, resolver, manifests
    squick-dict/       YAML dictionary engine
    squick-format/     Output emitters (markdown / JSON / NDJSON / triples / conventions)
    squick-watch/      Debounced file watcher
    squick-cli/        `squick` binary
  bindings/
    node/              napi-rs bindings (npm distribution)
    python/            PyO3 bindings (PyPI distribution)
  extensions/
    vscode/            VS Code extension
  dictionaries/        YAML pattern catalogues

Built by Horizon LLC

Squick is built and maintained by Horizon LLC - we design and build custom AI developer tooling, MCP integrations, and agent infrastructure for engineering teams.

Need custom AI tooling for your team? Get in touch via pixelhorizon.dev.

License

Trademarks

"Squick" and the Squick logo are trademarks of Horizon LLC. The Apache License 2.0 grants no rights in the trademarks. See TRADEMARKS.md.

Contributing

Contributions to source code, dictionaries, and documentation are welcome under the terms of the Apache License 2.0. By submitting a contribution, you agree that it is licensed under the same terms as the project itself.

Run the test suite before submitting:

推荐订阅源

Hacker News - Newest: "AI"

Why