Lab 8 — Agent skills and tool scaling

This standalone snapshot extends harness ask in two ways:

• a skill: a packaged procedure (SKILL.md) that the harness loads only when it’s relevant.
• an MCP server: an external process that offers new tools. Here it is a fake order system the app can’t otherwise see.

The exercise matches the Claude Code Lab 8.

What changes

• harness/project_skills.py:
  • Skills live in .harness/skills/<name>/SKILL.md in the project, or ~/.harness/skills/<name>/SKILL.md for you alone. Front matter gives the name, a description and optional allowed-tools.
  • Only each skill’s name and description go in the system prompt. The body is loaded on demand: by the model, with the use_skill tool, when a task matches the description; or by you, with /<skill-name> [extra instructions].
  • allowed-tools lists permission rules (Lab 6 syntax) that are pre-approved once the skill is in use, so its routine steps run without prompts. They never override a deny or an explicit ask rule, and /permissions shows them.
• harness/mcp_client.py:
  • A minimal MCP client over stdio: start the server, initialize, tools/list, tools/call. Each server tool becomes mcp__<server>__<tool>, so it can’t collide with a local tool.
  • MCP tools go through the same hooks, permissions and trace as local tools. A tool the rules don’t mention asks (Lab 6 default).
  • harness mcp add NAME -- COMMAND... registers a server for this project only, in ~/.harness/projects/<project>/mcp.json (outside the repository, like Claude Code’s local scope). harness mcp list starts each server and lists its tools; harness mcp remove NAME unregisters it.
• orders_mcp.py is a ~100-line, dependency-free MCP server with synthetic orders: list_orders(status) and get_order(order_id).
• The CLI: /skills and harness skills list skills; /mcp lists servers and their tools.

harness/skills.py keeps the earlier in-process models of a skill registry with approval state, tool discovery and MCP namespacing.

Learner steps

Start from: labs/app/ at tag lab07-done (see the track guide).

1. Install this lab and configure Foundry as before:

   python -m venv .venv
   . .venv/bin/activate          # Windows: .\.venv\Scripts\Activate.ps1
   pip install -e '.[dev]'
   cp .env.example .env
   az login
   cd ../app
   

Part A: a release-notes skill

1. Read assets/.harness/skills/release-notes/SKILL.md, then install and commit it:

   cp -R ../lab08-skills-tools/assets/. .
   git add .harness && git commit -m "chore: add release-notes skill"
   harness skills
   

2. Invoke it explicitly:

   harness ask
   

   /release-notes
   

   Then /exit and, in a new session, invoke it implicitly:

   harness ask "What changed since tag lab04-done? Write it up as release notes."
   

   Observe: in the second run, a use_skill tool call picked from the description. Then git log runs and CHANGELOG.md is written, although a one-shot run has no one to answer a permission prompt: the skill’s allowed-tools pre-approved write_file(CHANGELOG.md). Commits are grouped by Conventional Commit type (your Lab 5 convention pays off here). Nothing is committed. Review CHANGELOG.md and commit it yourself.

Part B: an MCP server for orders

1. Register the server for this project, with an absolute path:

   harness mcp add orders -- python "$(cd ../lab08-skills-tools && pwd)/orders_mcp.py"
   harness mcp list
   

   harness mcp add orders -- python (Resolve-Path ..\lab08-skills-tools\orders_mcp.py).Path
   harness mcp list
   

2. Use it:

   harness ask -n restock
   

   Type /mcp to see the server and its tools, then:

   Using the orders tools, find pending orders that we cannot fulfil with the current stock in catalog.csv. Write RESTOCK.md listing each SKU, quantity needed, stock on hand and shortfall.
   

   Observe: tool calls named mcp__orders__list_orders / mcp__orders__get_order, the permission prompt for each new tool (answer a to allow the same call for the session), and how the model joins MCP data with local files. Check RESTOCK.md against catalog.csv yourself. Did it correctly ignore the large cancelled order?

3. Unregister the server so it doesn’t follow you into later labs:

   harness mcp remove orders
   

4. Record
   • How the skill was triggered (slash command vs description match), and which calls its allowed-tools let through without a prompt.
   • The MCP tool names and whether you were asked to approve them.
   • Anything in RESTOCK.md that was wrong.

5. Checkpoint. Commit CHANGELOG.md and RESTOCK.md if you’re keeping them, then in app/:

   git tag lab08-done
   

6. Back in the lab folder, run the offline checks and inspect the declared capabilities:

   pytest checks/
   harness lab-info
   

harness ask keeps tracing (Lab 7), permissions (Lab 6), file memory (Lab 5), plan mode and todos (Lab 4), sessions (Lab 3) and the Lab 2B tools, built-in policy, project hooks and rules. This is a teaching harness, not a sandbox: only use it on the practice app.