Claude Code 2.1.296: per-subagent compaction and longer MCP tool descriptions

what changed

Claude Code 2.1.296 shipped on 9 October. Three changes stand out for teams that build on it.

Subagents can now set autoCompactWindow in their frontmatter or in --agents definitions, so a subagent can compact its context earlier than the main conversation. A new environment variable, CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL, runs every workflow agent on one model while other subagents keep theirs. And the default limit on MCP tool descriptions and server instructions sent up front doubled, from 2,048 to 4,096 characters.

The release also fixes managed PreToolUse hooks that deny a call with "continue": false. Before the fix, they refused the call but did not end the turn.

why it matters

Subagent context is where long agent runs go wrong. Being able to compact one busy subagent early, without touching the main session, is a cheap fix for runs that drift. The workflow model setting gives you one place to control cost for fan-out work.

The MCP change cuts both ways. Longer descriptions help Claude pick the right tool, but every server you connect now has more room to fill the context before work starts.

If you enforce policy with managed hooks, the hook fix matters most. A deny that does not stop the turn is not a full control.

what we would do

Upgrade, then check how many tokens your MCP servers use at the start of a session. Trim the tool descriptions you control. If you use managed hooks to block actions, test again that a deny now ends the turn. For long subagent runs, try a lower autoCompactWindow on the noisiest subagent first.