Babashka AI Coding Tools

A Babashka task for running tests via nREPL connections.

Overview

Provides utilities for connecting to nREPL servers and running tests.

Runs in Babashka, making it fast & easy for Agents to run tests from the command line.

Can be used with Aider which allows for focused context and minimizes token consumption.

Features

• Connect to nREPL servers
• Reload namespaces before running tests
• Capture reload errors
• Run tests in specified namespaces
• Capture test output (stdout/stderr)

Unique value proposition

Why use this instead of other options?

• if you prefer a test driven workflow
• all changes are written to source before being sent to the REPL
• fresh nREPL connection for every test run ensures clean state
• can be used with AI agents that are cheaper e.g. Aider
• agent flexibility allows use of any coding model

Installation

The REPL where your tests will run must have tools.namespace installed

Add the following dependencies to your bb.edn (check for latest commit)

babashka/nrepl-client {:git/url "https://github.com/babashka/nrepl-client"
                       :git/sha "19fbef2525e47d80b9278c49a545de58f48ee7cf"}
nextdoc/ai-tools {:git/url "https://github.com/nextdoc/ai-tools.git"
                  :git/sha "f30c664e96af56e8392ff4bd1fca6147d11f589a"}

Add this task (from the bb.edn in this project) to your bb.edn

nrepl:test {:requires [[io.nextdoc.tools :as tools]]
            :doc      "Run a test in the JVM using an nrepl connection i.e. fast test runner from cli"
            :task     (System/exit (tools/run-tests-task *command-line-args*))}

Run the task to confirm installation and see the options

bb nrepl:test

Command Line Options

• -n, --namespaces: Comma-separated list of test namespaces to run (required)
• -d, --directories: Comma-separated list of directories to scan for changes & reload before running tests (optional)
• -p, --port-file: Path to the file containing the nREPL port (default: ".nrepl-port")

The --directories option is useful if some of your sources fail to reload cleanly using tools.namespace.

Usage

The task is intended to be used with AI coding agents as a test runner.

If using a coding agent that uses markdown to know about the test runner, add this text to your agent instructions...

Run tests using this command `bb nrepl:test -n <fully qualified test namespace>`

Ensure the REPL in your project is started.

Test runner

Run the task with valid options to run your test(s)

The task will return a zero exit code if the tests pass. The return code is used by most coding agents to determine if a test passed or not.

If standard out or standard error is present this will be echoed to the terminal. This allows coding agents to see logs and exceptions from test runs.

Workflows

The Test Runner task is designed for a TDD style development workflow.

Tests can be in their own namespace or it can be useful to create a fiddle namespace and have sample code and a test to drive that code in the same file.

There are many AI development agents that can be used with this task. What follows is the setup and workflow for the agents we have tested...

Aider

Aider is a very precise AI development agent. Precision means that you manage all the files in its context and which files are readable and writable. This provides a high level of control but requires more manual work to manage the context.

In your .aider.conf.yaml file you will want...

auto-commits: false
watch-files: true
test-cmd: bb nrepl:test -n your.test.namespace
auto-test: true
yes-always: true

This will allow you to add comments in any namespace. Aider will:

• detect via its file watcher
• respond with answers and/or changes
• reload some or all of the changed namespaces
• fix any errors it notices if the updated source doesn't compile
• re-run the tests using the test runner task
• add the output from the test runner to its context
• fix any errors it notices if the test runner returns a non-zero exit code
• iterate on the change test loop until the test runner returns a zero exit code

Tips:

• use /web to load a API references and example code into the context
• or use /load to load many local and remote resources
• keep the context size under 30k. Iteration tends to slow down around this point. /clear and reload init context

The biggest downside of working with this level of automation is that there is no human in the loop interruption unless Aider iterates until it makes the tests pass or it hits its maximum number of reflections. You can inject yourself back into the loop by using CTRL-C at any time in the terminal.

The upside is this is a very lean and focused automated loop which you can control with a great deal of precision. This provides the full agentic power of the model, but with the smallest amount of token consumption.

Claude Code

Claude code represents a higher level of agentic coding compared to Aider. When using this TDD style workflow, the following features provide increased benefit:

• Context is automatically managed and compacted when required
• More tools are available
• The agent can adjust the Babashka task invocation on its own or with your instruction

Add the following instruction to your CLAUDE.md

# Run single test via nREPL (fast iteration). Use this for "TDD mode" iteration.

# The developer will indicate when to operate in "TDD mode".

# More information on how to use this task is available at https://raw.githubusercontent.com/nextdoc/ai-tools/refs/heads/master/README.md

bb nrepl:test -n <Fully qualified test namespace>

This allows you to instruct the agent to use TDD mode with a specific test or fiddle namespace. It will run this task and benefit from the speed and other values described above.

Tips:

• imbalanced parentheses happens less and less with higher power models. When you notice the agent in a loop with this it's often faster to interrupt and fix it manually. Then ask it to resume.
• To enjoy the same context specific benefit of an Aider AI comment, add a comment anywhere with as much detail as you need. Then instruct Claude to implement it using the specific line number. This provides the location and file context which improves the knowledge of the agent when it starts.

Gotchas

It can be useful to instruct the agent to add logging or print lines to see what is happening when running tests. If you have any core async and logging inside those go blocks then this can suppress standard out capture by the test runner. In this case you will see the standard out in your host REPL instead. You can paste it into the conversation with the agent.

Alternatives

This tool is designed to provide fast test execution and feedback for coding agents without need for any extra infrastructure. Just a simple command line task for a coding agent to invoke.

It has a single tool for interacting with your REPL and that is a code-reloading test invocation. A test can be used as a proxy for eval, so it is not limited to test workflows only.

If you want finer grained tools and more control then you need a more sophisticated integration.

For the next level of sophistication you probably want to look at an MCP server.

Is it safe?

This is buyer beware. Any coding agent that can write and evaluate code on your computer has associated risks.

The responsibility for checking the code being run is on you.

You might think that providing access to a REPL raises the risk of unwanted side effects. This is partially true but most coding agents can also write bash scripts and pose the same level of danger from running those scripts.

License

Copyright © 2025

Distributed under the Eclipse Public License version 1.0.