Claude Code release notes: 2.1.296 adds subagent compaction control

Claude Code 2.1.296 lets a subagent compact its conversation earlier than the main conversation. Released on October 9, the version also adds a separate model override for workflow agents.

Source: Anthropic's Claude Code 2.1.296 release notes.

Claude Code adds separate subagent compaction

A subagent is a specialized AI assistant that handles a task in its own context window, the material the model can use in a request. It then returns results to the main conversation.

The new autoCompactWindow field can be set in a subagent's frontmatter, the settings block at the top of its Markdown file, or in definitions passed through --agents. It lets the subagent automatically compact its history earlier than the main conversation's window requires. The main conversation does not have to use the same threshold.

Human step: Review the subagent that needs earlier compaction. Project definitions live in .claude/agents/; personal definitions live in ~/.claude/agents/. Change the relevant definition rather than the whole session's settings. The release notes do not specify the new field's value format or bounds, so no numeric configuration example is shown here.

Sources: Claude Code 2.1.296. Anthropic's subagent documentation.

One model choice for workflow agents

The same version adds the CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL environment variable. It runs every workflow agent on one chosen model while other subagents keep their model choices. This is a separate control from the compaction field above.

Both changes belong to Claude Code 2.1.296. They are not settings for ordinary conversations in the Claude app.

Source: Claude Code 2.1.296: compaction and model selection.