Catch broken environment, import, and route connections with Weftgate

A developer maintains Weftgate, an open-source verification gate for coding agents that checks environment-variable reads, dependency imports, and FastAPI route references against repository contracts. The tool runs locally via CLI, MCP, hooks, and CI, with no runtime dependencies and no network calls on its default path, and returns accept, review, or reject results with suggested corrections. The maintainer cautions that a passing gate does not prove an application works and that fixture scores are not claims about arbitrary repositories.

I maintain Weftgate, an open-source verification gate for coding agents. It checks environment-variable reads, dependency imports, and FastAPI route references against contracts in the repository. It runs locally through a CLI, MCP, hooks, and CI. The core has no runtime dependencies and makes no network calls on its default path. It is an early static analyzer with explicit limits. A passing gate does not prove that an application works. Use Python 3.10 or later in a fresh directory: mkdir weftgate-example cd weftgate-example python3 -m venv .venv . .venv/bin/activate pip install weftgate printf 'DATABASE URL=\n' .env.example printf 'import os\nurl = os.environ "DATABSE URL" \n' app.py weftgate check app.py The environment oracle compares the read with .env.example . The misspelled hard claim rejects, exits 1, and suggests DATABASE URL . Correct the reference and check again: python printf 'import os\nurl = os.environ "DATABASE URL" \n' app.py weftgate check app.py This time the reference resolves. Now try a computed key: python printf 'import os\nkey = "DATABASE URL"\nurl = os.environ key \n' app.py weftgate check app.py That returns review , with exit code 0. The environment parser does not establish the computed key's value, so the finding stays advisory. These are constructed demonstrations, not field accuracy benchmarks. The 58-second demo https://github.com/Avinash-Amudala/weftgate/releases/download/v0.1.1/weftgate-demo-motion.mp4 traces three intentional mistakes in a fixture: | Reference | Repository evidence | Suggested correction | |---|---|---| | import requestz | Dependency declarations | requests | | os.environ "DATABSE URL" | .env.example | DATABASE URL | | Route handler helth | Defined/imported handlers | health | You can reproduce the broader mutation fixture locally: weftgate eval mutate --fixture --seed 13 The harness measures detection, blocking, and usable suggestions separately. Its fixture score is not a claim about arbitrary repositories. An import absent from a manifest can be transitive. A router produced by a factory can register paths that a static scan cannot enumerate. A missing index cannot establish absence. These cases must review or be reported as unverifiable. An accept result only describes the extracted claims. Continue running your tests, type checker, security checks, and normal review. weftgate doctor weftgate audit weftgate check app/main.py doctor explains which contracts are available. Environment reads supplied externally should be documented in your declaration source. Python and Node dependency resolution can depend on workspace and runtime configuration, so review findings in context. For a pull request: name: Verify connections on: pull request permissions: contents: read jobs: verify: runs-on: ubuntu-latest steps: - uses: actions/checkout@v5 with: fetch-depth: 0 - uses: Avinash-Amudala/weftgate@v0.1.1 MCP clients can start weftgate mcp from the target repository. The tools and CLI call the same gate functions; no separate model API or server account is required. The most useful feedback is a minimal example of a false block or a missing contract. There are also contribution issues for workspace fixtures and a first-project walkthrough. Disclosure: this walkthrough was generated with AI and checked against the project's reproducible fixtures. It does not claim independent benchmark results.