---

name: cli-and-services description: "Use Agent Lightning agl CLI commands, store and Prometheus services, LLMProxy/vLLM service patterns, metrics, and safe endpoint checks." disable-model-invocation: true metadata: disco-role: operating license: MIT

CLI and services

Use this sub-skill when the user asks about agl commands, store servers, Prometheus metrics, LLM proxying, vLLM integration, OpenAI-compatible endpoint checks, or service startup failures.

Route by task

Request	Read/run
Check agl, agl store, or agl prometheus flags	references/cli-reference.md
Start or connect to a LightningStore service	references/service-workflows.md
Use LLMProxy with OpenAI-compatible/vLLM endpoints	references/service-workflows.md
Check a LiteLLM/OpenAI-compatible proxy safely	python scripts/check_litellm_proxy.py --base-url ... --model ...
Emit local demo metrics or check Prometheus dependencies	python scripts/check_prometheus_metrics.py --duration 1 --host 127.0.0.1
Debug CLI/service failures	references/troubleshooting.md

Key rules

• Always run agl --help and subcommand --help before relying on remembered flags.
• agl store supports in-memory and optional Mongo backends.
• agl prometheus requires Prometheus client support and PROMETHEUS_MULTIPROC_DIR when starting the exporter.
• agl vllm imports vLLM and is optional; expect it to fail in CPU/base environments without vLLM.
• LLMProxy wraps LiteLLM and can route /rollout/<id>/attempt/<id>/v1/... requests so spans are attributed to the correct attempt.
• Do not print API keys. Endpoint checkers should show model IDs and status, not secrets.

Safe checks

agl --help
agl store --help
agl prometheus --help
python scripts/check_litellm_proxy.py --help
python scripts/check_prometheus_metrics.py --duration 1 --host 127.0.0.1

The LiteLLM proxy checker makes network requests only when --base-url and --model are explicitly supplied. The metrics checker binds localhost by default in examples and exits quickly.

Boundary

This sub-skill owns command syntax and service surfaces. For rollout status semantics use runner-store-training. For interpreting spans or token IDs use tracing-and-instrumentation. For optional example selection use examples-and-recipes.