Skip to content

Exercise 15 — Runtime Skills

Code-Along Part 3 · 10 minutes

Persona: Riley — SRE lead
Continue in: part3/lab/ (or the completed part3/solution/), Skills & guardrails workspace.
Files to edit:

  • src/main/resources/skills/incident-response/SKILL.md

Enterprise context and goal

Riley, the SRE lead, needs consistent response guidance for checkout errors and authentication failures without making the main agent prompt a growing collection of runbooks. Operators need recommendations that reflect the incident priority and name the checks they must perform, while keeping proposed work distinct from actions already completed. Riley wants that domain guidance maintained as a reusable runtime skill.

Extend the supplied incident-response skill with escalation rules and service-specific operator checks, then compare checkout and authentication recommendations in the same running application. A runtime skill advertises a name and description, then provides its instructions when the model calls activate_skill. Java approval and evidence checks remain separate from that model-selected guidance.

This is application guidance consumed by the running agent. Part 1's AGENTS.md governs your coding assistant. MCP supplies tools/data; A2A delegates to a remote agent. Skills provide reusable instructions.

Step 1 — Start the shared project (2 min)

From the repository root, choose the starter once:

cd part3/lab
export OPENAI_API_KEY=your-key
./dev.sh

Use part3/solution instead for the completed run-and-read route. Java 25 and a running Docker or Podman engine are required. Open http://localhost:8080. Stay in this folder for Exercises 15–20. Stop any previous workshop app using port 8080 first.

The supplied launcher creates a dedicated PostgreSQL container and named volume. Later code/resource edits use hot reload. Pre-download dependencies and the image with ./dev.sh --prepare before the event; downloads are outside the core timebox. No real API key is needed for that preparation.

Step 2 — Add domain guidance (4 min)

Open the skill file. Keep its YAML name and description. Extend the body with:

For P1 or a report of complete authentication failure, recommend ESCALATE to the on-call team.
For other reports recommend INVESTIGATE unless the facts explicitly justify escalation.
For checkout 503 errors, propose checking deployment timing and comparing error rate with baseline.
For authentication failures, propose checking identity provider health and a synthetic login.
Name checks for the operator to perform; never say these checks already passed.

Save. The supplied quarkus.live-reload.watched-resources setting watches this skill. Wait about three seconds for the change scanner, then make a new request to trigger reload; keep the terminal running.

Inspect the supplied ResponseAgent method:

@Skills("incident-response")
@OutputGuardrails(value = ResponseGuardrail.class, maxRetries = 2)
@Agent(description = "Recommends an evidence-grounded response requiring human approval.",
       outputKey = "recommendation")
Recommendation recommend(@MemoryId String invocationId, String priority, String report);

The supplied invocation service assigns a new memory ID for every recommendation and judge call, keeping unrelated incidents out of the same conversation. This temporary model context is separate from the durable response record used later.

The extension exposes the native skill tool. There is no shell execution or external skill download in this lab.

Step 3 — Compare two recommendations (4 min)

In Skills & guardrails, submit the default P2 checkout report. Inspect the exact report excerpts, suggested action, and operator checks. Then select P1 and use:

Complete authentication failure. Synthetic logins fail for all users. No remediation has run.

Expect an escalation recommendation and authentication checks. A model response may vary; a plausible answer alone does not prove tool activation. Exercise 19's nativeSkillToolRoundTrip verifies that the real skill instructions are returned to the model after its tool call. Runtime activation remains a model choice, while the approval and evidence checks are enforced in Java.

Done when

  • The skill file contains domain guidance for checkout and authentication incidents.
  • A new request after saving uses the updated skill without resetting the application.
  • You can distinguish the runtime skill from coding-assistant rules, MCP, and A2A.

Next: Exercise 16 — Runtime Output Guardrails.

Reference: Quarkus LangChain4j skills.