Claude API compaction block: keep Sume job ids past the summary

With the compact-2026-09-04 beta header, the Claude Messages API can compact a conversation on demand: you send the top-level compaction parameter and get a signed compaction block that summarizes the messages you sent. A summary may drop a Sume job_id, so store ids and keys in your own state and recover with the read tools.

Vendor facts are from the Claude API release notes (September 14 entry); Sume facts from Jobs and results and MCP tools and gates, read 2026-10-01.

What does the compaction block do?

Per the notes, on later requests you send the block first, in place of the messages it summarizes. You choose when to compact, the request can run in the background, and you can keep recent turns word for word after the summary.

What should survive outside the summary?

Three small values are enough to resume a Sume workflow.

How do I recover ids after compaction?

Call jobs_list to find jobs, jobs_status to see which are terminal, and one batch jobs_wait with job_ids for the rest. The docs say that after a slice expires you retry jobs_wait with the same ids and never resubmit the paid create. jobs_result returns the answer once a job is complete.

What if the summary lost a key?

Look the job up before creating anything. A new create with a new key is a new paid job. When the MCP connector is involved, the newer mcp-client-2026-09-15 beta pins the fetched tool list; see MCP tool listing pin and Sume.