Claude Code v2.1.296: earlier subagent compaction, one model for workflows
What is it?
A subagent definition can now set autoCompactWindow so it compacts (summarizes) its context earlier than the main conversation. CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL runs all workflow agents on a single model. There is also a new env var that lengthens the max backoff for overloaded (529) retries, and an allow_large option on the Read tool.
Why does it matter?
Long-running subagents may hit context limits less often, and one model makes workflow cost easier to control. Teams running managed policies should know that a managed PreToolUse hook that denies a call no longer ends the turn. The notes also show Sonnet 5.5 cache-read pricing in cost displays changing from $0.20 to $0.10 per million tokens. It is not confirmed whether that reflects an API price change.
Maturity
GAStable release with about 90 items; only a subset is covered here. Behavior is as stated in the release notes, not tested.
How to track it
- Follow-up fixes to
autoCompactWindowandCLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL: watch GitHubanthropics/claude-codereleases (Watch → Releases). - Whether the Sonnet 5.5 cache-read figure reflects a real API price change: check the Anthropic pricing page and the Claude platform release notes.
- The managed
PreToolUsehook change: teams running policies should look for further changes in the next release notes.