Retail Local Runner

Retail Local Runner

Use this skill when the user wants to run, stop, configure, inspect, or troubleshoot the Retail Shopping Assistant from this repository with app code running locally.

Core Workflow

The deterministic runner is:

python skills/retail-local-runner/scripts/local_runner.py <command>

Available commands:

• configure: create ignored config-local.yaml files that point app services to a remote NIM host.
• install-dev: create .local-run/dev-venv and install Python dev/test packages there.
• start: start local Milvus infra containers, then local memory, guardrails, catalog, chain-server, and UI processes.
• stop: stop only tracked local app/UI processes and local Milvus infra containers. Never stop remote NIMs.
• status: show tracked process, port, health, and Milvus infra status.
• logs: print recent logs from .local-run/logs.

Stop Workflow

When the user asks to stop, shut down, tear down, or restart the local Retail Shopping Assistant, run stop first. Do not ask for the NIM host and do not check or create config-local.yaml files for stop-only requests.

python skills/retail-local-runner/scripts/local_runner.py stop

The stop command only kills PID files tracked under .local-run/pids/ and stops local Milvus infra containers from docker-compose.yaml:

• milvus
• minio
• etcd

It must not stop or modify the remote NIM machine from docker-compose-nim-local.yaml. If ports are still occupied after stop, use status and lsof to report the untracked owner instead of killing unrelated processes.

Remote NIM Host

Before running configure or start, check whether all three local override files exist:

find shared/configs -name config-local.yaml -print

If any override file is missing, ask the user in chat for the remote NIM host URL before running the script. Do not treat the runner's missing-host error as the final answer. Ask exactly:

What is the remote NIM host URL? Use the base host without per-model ports, for example http://NIM_HOST.

Use one host URL such as http://NIM_HOST; the runner derives these endpoints:

• LLM: :8000/v1
• text embeddings: :8001/v1
• image embeddings: :8002/v1
• content safety: :8003/v1
• topic control: :8004/v1

Run:

python skills/retail-local-runner/scripts/local_runner.py configure --nim-host http://HOST

For a fresh start when the URL is already known, this is enough:

python skills/retail-local-runner/scripts/local_runner.py start --nim-host http://HOST

Local Services

The runner starts these local app processes:

• memory retriever: http://localhost:8011
• guardrails: http://localhost:8012
• catalog retriever: http://localhost:8010
• chain server: http://localhost:8009
• UI: http://localhost:3000

Only Milvus infra remains containerized:

docker compose -f docker-compose.yaml up -d etcd minio milvus

If another local Milvus is already healthy at localhost:19530 with health on http://localhost:9091/healthz, the runner reuses that local endpoint instead of creating conflicting Docker containers. stop still only stops this repo's tracked app processes and this repo's Compose infra; it does not stop Milvus containers owned by another project.

Operational Notes

• Runtime files live under .local-run/; service logs are in .local-run/logs/.
• Service virtual environments are created under each service's venv/ directory.
• Existing local Milvus on localhost:19530 is reused when healthy.
• Python dev/test packages must be installed into .local-run/dev-venv, never the global interpreter: python skills/retail-local-runner/scripts/local_runner.py install-dev.
• Use .local-run/dev-venv/bin/python -m pytest ... for local unit tests after install-dev.
• UI dependencies are installed into ui/node_modules when missing.
• The runner creates ui/public/images -> shared/images so React dev mode can serve catalog images from /images/..., matching the UI Dockerfile behavior.
• The runner sets CONFIG_OVERRIDE=config-local.yaml, SHARED_ROOT, SHARED_CONFIG_ROOT, REACT_APP_API_BASE_URL=http://localhost:8009, and BROWSER=none.
• Use status before deciding whether to start or stop.
• Use logs --service catalog-retriever --lines 120 when catalog startup is slow; first startup may populate Milvus embeddings through the remote NIMs.

Validation

After editing this skill, validate both folders:

uv run --with pyyaml python "$CODEX_HOME/skills/.system/skill-creator/scripts/quick_validate.py" skills/retail-local-runner
uv run --with pyyaml python "$CODEX_HOME/skills/.system/skill-creator/scripts/quick_validate.py" .agents/skills/retail-local-runner