Gemini CLI tool output bound vs Sume's 256 KiB MCP result cap
Gemini CLI v0.63.0-preview.0 bounds tool output size. Sume refuses MCP results above 256 KiB (mcp_output_too_large); not a job failure, so read narrower.
Two different limits can shorten a tool result. Gemini CLI v0.63.0-preview.0 lists a change to "bound tool output size", and Sume's MCP server returns mcp_output_too_large for a result above 256 KiB. The Sume error is an output limit, not a job failure: read a narrower or paginated slice.
The Gemini CLI line is from its release list; the Sume limit from packages/mcp-server in the Sume repository and the hosted MCP docs, read 2026-10-01.
What did Gemini CLI change?
The 2026-09-29 preview notes include "bound tool output size and optimize memory lifecycle in long-running agent loops". The release list does not give the numeric limit, so check your version's settings before relying on one.
What does Sume do with a large result?
The server's constant is MCP_OUTPUT_MAX_BYTES = 256 * 1024. Past that it returns a text result with code mcp_output_too_large and max_bytes, saying the tool returned more data than MCP can display, to use a narrower or paginated read, and never to resubmit a paid create.
| Signal | Cut by |
|---|---|
Result contains mcp_output_too_large and max_bytes | Sume server |
| Result ends abruptly with no Sume error code | Look at the client's output bound |
| Job status is terminal and healthy | Not a job failure either way |
How do I read the rest?
Use the narrowest read that answers the question: jobs_status or jobs_get for one job, jobs_result for artifacts, and list tools with their paging inputs. Report media.sume.com URLs and public ids instead of pasting bulk output. Timeouts are a separate topic, covered in Gemini CLI MCP timeouts and `jobs_wait`.
Should I resubmit when I see the error?
No. The message says it is not a job failure, and a resubmitted paid create would start another job. Re-read the existing job by id.
Sources
Related posts
More in Integrations
- GPT-6 Sol tool calls: use the Responses API to drive Sume
GPT-6 Sol only calls functions in Chat Completions when reasoning effort is none. For Sume's MCP tools, use the Responses API and keep reasoning on.
- Make OpenAI Sora modules deprecated: use HTTP with Sume
Make's notes list the OpenAI Sora and Video modules as deprecated by Sept 24. Replace them with an HTTP module that submits a Sume video job and a webhook.
- Make webhook 300 requests per 10 seconds: Sume bulk callbacks
Make returns 429 above 300 webhook requests per 10 seconds. Sume treats a non-2xx callback as a failed attempt and retries up to 10 times, 30 seconds apart.
- Make webhook logs kept 3 days: recover a Sume result later
Make keeps webhook logs for 3 days (30 on Enterprise). A Sume result does not depend on them: store the job id, read job events, and redeliver the callback.
Written by Sume