AI Tools for Developers

Agentic AI workflows enabled by Docker containers.

Source for many experiments in our LinkedIn newsletter

Running Prompts

To generate prompts for a project, clone a repo into $PWD and run the following command.

docker run --rm \
           -v /var/run/docker.sock:/var/run/docker.sock \
           --mount type=volume,source=docker-prompts,target=/prompts \
           vonwig/prompts:latest --host-dir $PWD \
                                 --user jimclark106 \
                                 --platform darwin \
                                 --prompts-dir "github:docker/labs-make-runbook?ref=main&path=prompts/lazy_docker"

The four arguments are

• project root dir on host,
• docker login username,
• platform,
• github ref for prompt files.

If you need to test prompts before you push, you can bind a local prompt directory instead of using a GitHub ref. For example, to test some local prompts in a directory named my_prompts, run:

docker run --rm \
           -v /var/run/docker.sock:/var/run/docker.sock \
           --mount type=bind,source=$PWD,target=/app/my_prompts \
           --workdir /app \
           vonwig/prompts:latest --host-dir $PWD \
                                 --user jimclark106 \
                                 --platform darwin \
                                 --prompts-dir my_prompts

Running a Conversation Loop

Set OpenAI key

echo $OPENAI_API_KEY > $HOME/.openai-api-key

Run

docker run --rm \
           -it \
           -v /var/run/docker.sock:/var/run/docker.sock \
           --mount type=volume,source=docker-prompts,target=/prompts \
           --mount type=bind,source=$HOME/.openai-api-key,target=/root/.openai-api-key \
           vonwig/prompts:latest \
                                 run \
                                 --host-dir $PWD \
                                 --user $USER \
                                 --platform "$(uname -o)" \
                                 --prompts "github:docker/labs-githooks?ref=main&path=prompts/git_hooks"

Running a Conversation Loop with Local Prompts

If you want to run a conversation loop with local prompts then you need to think about two different directories, the one that the root of your project ($PWD above), and the one that contains your prompts (let's call that $PROMPTS_DIR). Here's a command line for running the prompts when our $PWD is the project root and we've set the environment variable $PROMPTS_DIR to point at the directory containing our prompts.

Set OpenAI key

echo $OPENAI_API_KEY > $HOME/.openai-api-key

Run

docker run --rm \
           -it \
           -v /var/run/docker.sock:/var/run/docker.sock \
           --mount type=bind,source=$PROMPTS_DIR,target=/app/my_prompts \
           --workdir /app \
           --mount type=volume,source=docker-prompts,target=/prompts \
           --mount type=bind,source=$HOME/.openai-api-key,target=/root/.openai-api-key \
           vonwig/prompts:latest \
                                 run \
                                 --host-dir $PWD \
                                 --user $USER \
                                 --platform "$(uname -o)" \
                                 --prompts-dir my_prompts

GitHub refs

Prompts can be fetched from a GitHub repository when using the --prompts arg. The mandatory parts of the ref are github:{owner}/{repo} but optional path and ref can be added to pull prompts from branches, and to specify a subdirectory where the prompt files are located in the repo.

Function volumes

Every function container will have a shared volume mounted into the container at /thread. The volume is intended to be ephemeral and will be deleted at the end of the run. However, the volume can be saved for inspection by passing the argument --thread-id.

Output json-rpc notifications

Add the flag --jsonrpc to the list of arguments to switch the stdout stream to be a series of jsonrpc notifications.
This is useful if you are running the tool and streaming responses on to a canvas.

Try running the with the --jsonrpc to see a full example but the stdout stream will look something like this.

{"jsonrpc":"2.0","method":"message","params":{"content":" consistently"}}Content-Length: 65

{"jsonrpc":"2.0","method":"message","params":{"content":" high"}}Content-Length: 61

{"jsonrpc":"2.0","method":"message","params":{"content":"."}}Content-Length: 52

{"jsonrpc":"2.0","method":"functions","params":null}Content-Length: 57

{"jsonrpc":"2.0","method":"functions-done","params":null}Content-Length: 1703

Prompt file layout

Each prompt directory should contain a README.md describing the prompts and their purpose. Each prompt file is a markdown document that supports moustache templates for subsituting context extracted from the project.

prompt_dir/
├── 010_system_prompt.md
├── 020_user_prompt.md
└── README.md

• ordering of messages is determined by filename sorting
• the role is encoded in the name of the file

Moustache Templates

The prompt templates can contain expressions like {{dockerfiles}} to add information extracted from the current project. Examples of facts that can be added to the prompts are:

• {{platform}} - the platform of the current development environment.
• {{username}} - the DockerHub username (and default namespace for image pushes)
• {{languages}} - names of languages discovered in the project.
• {{project.dockerfiles}} - the relative paths to local DockerFiles
• {{project.composefiles}} - the relative paths to local Docker Compose files.

The entire project-facts map is also available using dot-syntax forms like {{project-facts.project-root-uri}}. All moustache template expressions documented here are supported.

Building

#docker:command=build

docker build -t vonwig/prompts:local -f Dockerfile .