The MCP tool schemas exposed by glab mcp serve have very sparse parameter metadata after !2390 optimized token usage by stripping flag descriptions. This causes AI agents to frequently misuse tools because they're basically guessing what flags expect. Examples of failures I keep hitting * --output flag — no enum, no description: The schema just says {"type": "string"}. Agents reasonably guess it's a file path (like curl -o) and pass something like /tmp/output.zip. The actual valid values are json or ndjson, but there's nothing in the schema to indicate that. * glab_api(args=["projects/123/jobs/456/artifacts"], flags={"output": "/tmp/artifacts.zip"}) → Error: must be one of [json ndjson] * args — always just "Positional arguments": For glab_api, there's no hint that the value should be an API endpoint path like projects/:id/jobs. Agents have to guess the format. * --field — no format guidance: No indication that it expects key=value pairs. I am using claude at the moment, and it keeps hitting the same issues again and again. I would have to create local memories to "teach" it how to use the mcp which feel like it should not be needed. Is this to be expected? At the moment it seem to work better to just point the ai to the cli tool instead.