Stackbone evals skill

This skill tells you how to drive the eval loop: what to build first, which surface can build it, and how to read the verdict. It holds no criterion catalog and no flag tables: the criteria, their settings and every exit code live in the docs, and the live catalog lives on the box itself.

Where the facts live

The docs are served over MCP as the stackbone-docs server (https://docs.stackbone.ai/mcp, public, tools search_docs, get_doc, list_docs). stackbone docs prints the connection details.

• Start with list_docs. Evaluation (Home › Features) is the concept page and eval (CLI › Reference) is the command page. Take their paths from that list and get_doc them. Never type a path from memory: pages move, titles stay.
• The criterion catalog is not in the docs, it is on the box. Ask the box for it before you write a suite: the same catalog carries the judging models and the judge workflows this workspace actually has, painted in live. A catalog read from a doc is a catalog from another release.
• --help on the binary beats the docs for the installed version. Do not explain an exit code from memory.

No stackbone-docs tools in your session? Fetch https://docs.stackbone.ai/llms.txt: the same index, one line per page with its title, its description and a link to its raw markdown. Pick the page by title and fetch that link. https://docs.stackbone.ai/llms-full.txt is every page in one file (~700 KB).

Reach the box

Most of this loop is not in the CLI. Point your client at the box once: