Cline oversized MCP result cache: keep Sume job results small

Cline's SDK v0.0.89 caches oversized MCP results per session, up to 16 MiB, and returns a bounded preview with a cline://cache/... URI that read_files can page; an entry expires after five model iterations without a read (release notes, read 2026-10-04). With Sume you rarely need that path. Ask jobs_wait for ids without include_results, then fetch only the results you need.

How Sume keeps a wait small

jobs_wait holds at most 55 seconds per call and accepts a batch of job ids. By default it returns statuses. With include_results it adds the results, and when they are too big for one answer it names the ones it left out in results_omitted.job_ids (MCP tools and gates).

A pattern that avoids the cache

Start the jobs, wait on all ids with statuses only, and read results for the finished ones in a second call. If a wait with include_results lists ids under results_omitted.job_ids, read those in a batch jobs_result call. A model then sees a small payload each turn, and nothing depends on a client-side cache expiring after five iterations.

• Wait loop: repeat jobs_wait while any job is still running; a call that returns early is not a failure.
• Read step: one batch call for finished ids.
• A client timeout does not cancel a job, so a dropped connection never loses the render.

What the cache is good for

Cline's cache helps with results you cannot shrink, such as a large file listing. Use it there. For Sume job output, the cheaper fix is to ask for less in the first place.