The manual is a URL, and the agent reads one chapter at a time

The agent running in the container is a small model with a 32k window. It cannot hold Hoody's whole API in context. It doesn't have to.

https://hoody.com/SKILLS/SKILL-HTTP.md is a static file: what Hoody is, the auth model, the URL scheme, and an index of twenty service namespaces. No token, no login, one GET. When a task lands: snapshot this container, restore it if the tests fail: the agent reads the index, picks the namespace that owns the job, and fetches SKILL-HTTP/<ns>.md: the exact endpoints, payloads, and gotchas for that one surface. Links inside are absolute, so a chapter pasted into a prompt still resolves to the next one.

Three modes ship the same twenty namespaces: HTTP, SDK, CLI. A shell-only agent loads the CLI mode and never sees a JSON body.

The documentation is not in the weights. It is in the context window, one namespace at a time.