Sume MCP returned a tool descriptor, not a result: it did not run

Sume's hosted MCP tells agents that some clients register tools namespaced by server, as <server>-<tool>, and that if a call returns a tool descriptor (name, description, input schema) instead of a result, the tool did not run. Re-call it using the namespaced name your client lists.

Where the namespace comes from

The namespacing is client-side. A client that connects several servers prefixes each tool with the server name so two servers cannot collide. Sume's server never sees that form. It also accepts dotted names from some clients, which it converts to underscore names. So a single create can be named generate_image, sume-generate_image or something else, depending on the client, and only the name in your client's own tool list is right.

Result or descriptor

The two shapes differ in ways a harness can test.

A guard

Before waiting on a job, make sure there is one. This runs as written.

def ran(result: dict) -> bool:
    if 'inputSchema' in result or 'input_schema' in result:
        return False  # a descriptor, not a result
    return bool(result.get('response') or result.get('job_id'))

print(ran({'name': 'sume-generate_image', 'inputSchema': {}}))
print(ran({'job_id': 'job_1'}))

What to do next

Re-call with the namespaced name and the same arguments. For a paid create, keep the same idempotency_key: if the first call did not run there is nothing to deduplicate, and if it did, you get the same job and not a second one. If the name is wrong in a different way, the server answers tool_not_found with a did-you-mean suggestion; the Agents SDK post shows that shape.

Limits

This is guidance for client behavior, and the exact descriptor shape varies by client. Treat the check above as a starting point and adapt it to what your client returns.

Why this matters for money

A descriptor looks like a successful reply to a model: it is structured, it has a name, and it has a schema. If the agent treats it as a result, it may wait on a job id that was never created, or tell a user that a generation started. With a paid create the safe move is the cheapest: re-call with the right name and the same idempotency_key, and confirm a job id comes back.

Checklist before you ship

• Log the exact tool name your client registered, with its namespace.
• Re-call using the namespaced name the client lists, not the bare name.
• Never assume a paid create ran because the model produced a tool call.
• Confirm the job id exists before you start waiting.