little-coder Docker Model Runner Sandbox Kit

Docker Sandboxes kit for running little-coder against a local Docker Model Runner endpoint.

Create and run a sandbox in the current directory:

sbx run --kit git+https://github.com/shelajev/little-coder-sbx-kit.git \
  little-coder-model-runner . -- --model llamacpp/qwen3.6-35b-a3b-64k

The kit does not guess which local model you have. You must pass a model explicitly:

-- --model llamacpp/<docker-model-tag>

For example:

-- --model llamacpp/qwen3.6-35b-a3b-64k

If you forget the model, the sandbox fails fast and prints the exact command shape to use.

Host Setup

Enable Docker Model Runner's TCP endpoint:

docker desktop enable model-runner --tcp

Check your installed local models:

docker model list

Use the model tag from that list after the llamacpp/ prefix. For example, a Docker Model Runner tag named qwen3.6-27b becomes:

--model llamacpp/qwen3.6-27b

Optional: Qwen 64k Variant

If you already have the Unsloth Qwen3.6-35B-A3B GGUF model under its Hugging Face tag, give it a short local tag:

docker model tag \
  huggingface.co/unsloth/qwen3.6-35b-a3b-gguf:UD-Q6_K_XL \
  qwen3.6-35b-a3b

Create a lightweight 64k-context variant:

docker model package --from qwen3.6-35b-a3b \
  --context-size 64000 \
  qwen3.6-35b-a3b-64k

Verify Docker Model Runner can resolve it:

curl http://localhost:12434/v1/models/qwen3.6-35b-a3b-64k

Then run the kit with:

-- --model llamacpp/qwen3.6-35b-a3b-64k

Run

Create and run a sandbox with another Docker Model Runner tag:

sbx run --kit git+https://github.com/shelajev/little-coder-sbx-kit.git \
  little-coder-model-runner . -- --model llamacpp/<docker-model-tag>

Create a named sandbox:

sbx create --name little-coder-current \
  --kit git+https://github.com/shelajev/little-coder-sbx-kit.git \
  little-coder-model-runner .

Run that named sandbox later:

sbx run --kit git+https://github.com/shelajev/little-coder-sbx-kit.git \
  little-coder-current -- --model llamacpp/qwen3.6-35b-a3b-64k

For custom agent kits, pass --kit again when running an existing sandbox. sbx records the custom agent name on the sandbox, but sbx run <sandbox> may not resolve a Git or local kit-defined agent unless the kit is supplied again.

Model Selection

The kit ships with a sentinel model value:

LITTLE_CODER_MODEL_ID=__REQUIRED_MODEL_ID__

That sentinel is intentional. Docker Sandboxes kits do not currently have an interactive "ask during install" parameter field, so the public, non-surprising path is to pass the model after the sbx run ... -- separator.

Any Docker Model Runner tag can be used as a llamacpp/ model:

sbx run --kit git+https://github.com/shelajev/little-coder-sbx-kit.git \
  little-coder-model-runner . -- --model llamacpp/<docker-model-tag>

The launcher registers unknown llamacpp/* model IDs with little-coder at runtime. The default local profile context is 64000, matching the Qwen 64k packaging example.

Smoke Test

For sbx exec, close or pipe stdin so pi does not wait for more input:

sbx exec little-coder-current -- sh -lc \
  'little-coder-dmr --model llamacpp/qwen3.6-35b-a3b-64k --thinking off --no-tools --no-session -p "Reply with exactly: sbx-ok" < /dev/null'

Expected output includes:

sbx-ok

Local Clone

If you clone this repo, run.sh runs a named sandbox with the local kit path. You still need to pass a model:

./run.sh little-coder-current -- --model llamacpp/qwen3.6-35b-a3b-64k

Use another sandbox name as the first argument:

./run.sh my-sandbox -- --model llamacpp/qwen3.6-27b

Network Policy

The kit allows only:

• registry.npmjs.org:443 to install little-coder
• host.docker.internal:12434 to reach Docker Model Runner

It sets:

LLAMACPP_BASE_URL=http://host.docker.internal:12434/v1
LLAMACPP_API_KEY=docker-model-runner
PI_OFFLINE=1

PI_OFFLINE=1 suppresses pi startup update checks. Inference still goes to Docker Model Runner.

Useful Model Candidates

• qwen3.6-35b-a3b-64k: Qwen3.6-35B-A3B 64k Docker Model Runner variant.
• qwen3.6-35b-a3b: Qwen3.6-35B-A3B short local tag.
• qwen3.6-27b: Qwen3.6 dense fallback.

Use whatever tag is present in docker model list.