Docs command

[IvanKiral]

Motivation

kontent docs search|endpoint|object <query> looks up Kontent.ai Learn docs and API reference from the terminal, so an agent (or a human) driving kontent mapi can find the right endpoint and payload shape without leaving the CLI. It talks to the Learn-MCP service (https://learn-mcp.kontent.ai, no auth, GET only) through a new src/lib/learn client on core-sdk's createFetchQuery. Each subcommand prints a JSON array of service-ranked candidates; --limit (1-10) and --compact on all three, --api filter on search. The zod/mini schemas declare only the fields the CLI reads, so unknown keys pass through to stdout. Empty query, no match and request failure are distinct error kinds, and no-match hints point at docs search or at dropping --api.

Agent evals harness (evals/, opt-in) runs the Agent SDK against the built CLI on one cloned environment, then grades the resulting state through the Management SDK with deterministic checks. The agent gets Bash (confined to a workspace dir) and WebFetch (kontent.ai only, key-in-URL denied); the policy is a pure reducer wired into the PreToolUse hook. Seven tasks run sequentially in dependency order; a task whose parent did not PASS is recorded BLOCKED. Each run writes run.json, a markdown report, per-task traces, raw SDK messages and a shim log of every kontent invocation to evals/results/ (git-ignored). The paid run is gated on EVALS_MAPI_KEY/EVALS_SOURCE_ENV_ID and never part of pnpm test or CI; the harness's own unit tests do run in pnpm test. Adds @anthropic-ai/claude-agent-sdk as a dev dependency. Playbook in evals/README.md.

kontent mapi fixes from the eval runs:

• An empty 2xx body (e.g. 204 on publish) now logs HTTP <status> <text> on stderr instead of leaving both streams silent.
• The not-logged-in error lists every credential source: kontent login, --mapiKey, Authorization header, KONTENT_MAPI_KEY.
• A blank --envId is rejected before any request goes out.
• Help clarifies that --envId is prepended to the path and points at kontent docs for endpoint lookup.

Docs and infra: each command's README now lives next to its module (project/sample/bootstrap/), and generateCommandDocs.ts resolves the page folder from the filesystem. test/helpers/environment.ts and requiredEnv.ts are extracted from the e2e helpers and shared with evals. CLAUDE.md, the root README table and .env.template are updated.

Checklist

• Code follows coding conventions held in this repo
• Automated tests have been added
• Tests are passing
• Docs have been updated (if applicable)
• Temporary settings (e.g. variables used during development and testing) have been reverted to defaults

How to test

pnpm build
node dist/index.js docs search "publish a variant" --limit 3
node dist/index.js docs endpoint "upsert language variant"
node dist/index.js docs object "rich text element" --compact
node dist/index.js docs search "webhook" --api delivery_api
node dist/index.js docs search ""            # exits 1, empty-query message on stderr

For mapi: run kontent mapi types --envId "" (rejected before any request) and, while logged out with no key set, kontent mapi types --envId <id> (credential hint on stderr). Publishing a variant with -X PUT now prints HTTP 204 No Content on stderr.

For the evals run, set EVALS_MAPI_KEY and EVALS_SOURCE_ENV_ID in .env, make sure ANTHROPIC_API_KEY is unset, and run pnpm evals:run. Results land in evals/results/<date>-<time>-<model>/.