Opening an untrusted repository: the files that execute code when you install, build, test or just enter it

What it is

An agent asked to "fix the failing test" in an unfamiliar repository will usually install dependencies, build and run tests. Each of those steps executes code written by the repository's authors or their dependencies. Typical execution points:

• Package managers: npm runs lifecycle scripts (preinstall, install, postinstall, prepare) on npm install; Python source distributions run build code on install; Cargo runs build.rs; Composer and Bundler have their own hooks.
• Build and task files: Makefile, package.json scripts, setup.py, Gradle and Maven build files, justfile, Taskfile.
• Test configuration: conftest.py is imported by pytest; test runners load plugins named in configuration.
• Editors: VS Code workspace tasks, launch configurations and recommended extensions; VS Code's Workspace Trust feature opens unknown folders in Restricted Mode for this reason.
• Environment loaders: direnv loads .envrc on entering the directory, but only after direnv allow — an explicit trust step that an agent should never run on its own.
• Git: hooks are not transferred by clone, but repository-local configuration you are asked to include, or tools that honour settings such as core.fsmonitor or core.hooksPath, can run programs during ordinary Git commands.
• Containers and CI files: docker compose up builds from a Dockerfile the repository controls; running a CI workflow locally executes its steps.

Why it matters

The agent usually holds credentials: an API key in the environment, a Git token, cloud configuration in the home directory. Code that runs during install can read all of them. Nothing about the step looks dangerous in a transcript: it reads "installing dependencies".

How to apply

• Run install, build and test steps of untrusted repositories only in a sandbox with no host credentials mounted and restricted network egress.
• Read package.json scripts, build files and conftest.py before running them; mention anything that fetches or executes remote code.
• Prefer lock files and prebuilt artefacts; use --ignore-scripts where the project allows.
• Never approve trust prompts (direnv, editor trust, "allow this hook") on the user's behalf.

Pitfalls

• Assuming "only reading" is safe while an editor integration or language server loads project configuration.
• A repository you trust containing a vendored subdirectory you do not.