Sume MCP 256 KiB output budget: oversize error is not a failed job

A Sume hosted MCP tool result is limited to 256 KiB once serialized. When a result is bigger, the server returns an explicit error rather than a silently truncated value, and the error does not mean the job failed. The figure comes from Sume's operations note on API and MCP work bounds on origin/main, read on 2026-10-04. Read the job directly before you resubmit anything.

Why does the budget exist?

A tool result goes back into a model's context. A very large list or log would crowd out the conversation, and a half-cut JSON value would be worse than an error because a model could act on it. Sume chooses to fail loudly. The budget applies to the serialized result of one tool call, including a jobs_result, a jobs_events page or a list.

How should I react?

How do I get less data back?

Use the narrow calls. jobs_status returns the state without the payload, jobs_events returns sanitized lifecycle events with a default limit of 50, and usage_get takes an optional limit that defaults to 20. The Jobs and results page lists the read tools.

Inside script_run, filter before you return. The script sees full inner results, and only the value you return, plus the calls[] journal, goes back to the conversation. Return ids and the one field you need, not whole objects. That keeps the response small and the model's context clean.

When is a failure real?

Only trust a failure from the job. Call jobs_get and read error.public_reason, error.message and error.retryable; if the status looks odd, jobs_events shows the lifecycle. A jobs_cancel on a failed job returns 409 with no error fields, so do not use it to inspect. Never recreate a job on an oversize error, because that would pay twice for work that already succeeded. Reusing the same idempotency_key protects you if you do re-send. The MCP overview covers the rest of the tools.