Claude Code 2.1.76: MCP Elicitation, PostCompact Hook, Security Bypass Fixed, Opus 4.6 Output Raised to 128K

Two releases landed today and they're both significant. The headline security fix: PreToolUse hooks returning allow were bypassing deny permission rules — including enterprise managed settings. If you're running Claude Code in a managed enterprise environment with security hooks, this was a real bypass. Patch immediately.

Claude Code 2.1.76 ships MCP Elicitation — the first native implementation of a protocol that lets MCP servers pause mid-task and request structured user input via JSON Schema-validated forms or browser URLs. Two new hooks (Elicitation and ElicitationResult) intercept both the request and the user's response before they reach the MCP server. This eliminates the most common agent failure mode: silently guessing when given ambiguous inputs. If you're building MCP servers that need OAuth flows, config resolution, or progressive data gathering, this is the primitive you've been waiting for.

The PostCompact hook fires after context compaction with the full compact_summary, giving you the first deterministic entry point to observe what was lost and act on it. Combined with the new autoMemoryDirectory config for custom memory storage paths, you now have a complete memory lifecycle control surface. The pattern: PostCompact writes critical context back to CLAUDE.md or pushes the summary to an external system, ensuring compaction never silently drops load-bearing context.

Opus 4.6 default max output raised to 64K with a ceiling of 128K. The /copy N command copies the Nth-latest assistant response. A --resume race condition that silently truncated recent conversation history was also fixed — if you've had sessions that seemed to "forget" recent work after resume, that's why.

Also shipping: sparse worktree paths for monorepos via git sparse-checkout, an effort command, and a configurable session quality survey. Source