Claude Code subagent autoCompactWindow: keep Sume job ids
Claude Code 2.1.296 lets a subagent auto-compact earlier than the main chat. Keep Sume job ids out of the summary and re-read state with jobs_status.

Claude Code 2.1.296 (October 9, 2026) added autoCompactWindow to subagent frontmatter and to --agents definitions, so one subagent can auto-compact earlier than the main conversation's window. If that subagent is driving paid Sume jobs, a job id that only lives in its conversation can be summarised away. Do not rely on the transcript: write each job id and idempotency_key to a file or the parent's task list, and recover state with the read-only jobs_list and jobs_status tools.
The changelog says only what the setting is for. It does not state a unit or default in the entry, and the sub-agents page I fetched did not describe the field yet, so check the value your build accepts before you set it (changelog read 2026-10-13).
What changed in Claude Code 2.1.296?
The relevant entries from the Claude Code changelog, read 2026-10-13:
| Changelog entry | What it means | For a Sume subagent |
|---|---|---|
autoCompactWindow in subagent frontmatter and --agents | A subagent can auto-compact earlier than the main window | Earlier summaries, so earlier risk of losing ids held only in context |
CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL | Runs every workflow agent on one model | Not a compaction setting; it only picks the model for workflow agents |
| Default limit on MCP instructions and tool descriptions sent up front raised from 2,048 to 4,096 characters | More of a server's text is loaded at start | No change to job state |
Why do Sume job ids need to live outside the context?
A Sume paid call returns a job. The job belongs to the workspace and to the member whose key or Agent turn created it, and it has a stable id you can fetch later with GET /v1/jobs/:id or the jobs_get MCP tool. The statuses are queued and processing (not finished) and completed, failed and canceled (terminal), per the Jobs and results page (read 2026-10-13).
So losing an id from a summary is recoverable but annoying, and losing the idempotency_key is worse: Sume requires that key on write and paid tools, and resubmitting with the same key returns the original job instead of a second charge. A fresh key after compaction would be a second charge.
What should a compacting subagent do instead?
- Generate the
idempotency_keyin the parent or in a file before the subagent starts, and pass it in the task text. - Have the subagent append every returned job id to a file such as
jobs.txtas soon as the call returns. - After a compaction, or at the start of every phase, call
jobs_listandjobs_statusinstead of trusting memory. They are read-only and need no gate. - Use
dry_run=truefor previews andmax_spend_creditson paid calls, so an early summary cannot widen what a retry may spend.
---
name: sume-renderer
description: Submits Sume jobs and records ids in jobs.txt
mcpServers:
- sume
autoCompactWindow: <value your build accepts>
---
After every Sume call, append the job id and its idempotency_key to jobs.txt.
After any context compaction, read jobs.txt and call jobs_status for each id.What does this not solve?
Sume does not control how Claude Code compacts, and it cannot tell a subagent which ids it forgot. The only guarantee on the Sume side is that a job id stays fetchable and that a repeated key does not create a second job. A subagent that never wrote the key down can still re-submit with a new one.
This post covers the subagent setting only. For compaction on the Claude API itself, see the related post on compaction blocks.
Sources
Related posts
More in Developers
- CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL: model choice vs Sume spend gates
Claude Code 2.1.296 can run every workflow agent on one model. Sume's idempotency_key, dry_run and max_spend_credits gates do not change with the model.
- Claude MCP connector denylist: switch off Sume's paid tools
Enable all Sume tools in an mcp_toolset, then disable the paid ones by name. Anthropic recommends denylisting write tools for read-only assistants.
- Claude MCP connector supports tool calls only: how Sume results arrive
The Messages API connector supports only MCP tool calls, not resources or prompts. Fetch Sume job results with the jobs_result tool, not a resource read.
- Clef takes 1 to 64 questions per request: check a brief in one call
Cloudflare's Clef page allows 1 to 64 typed questions on one state. Put the brief checks in one request, then submit to Sume once.
Written by Sume