CLAUDE CODE V2.1.76: MCP ELICITATION, MONOREPO SPARSE CHECKOUTS, AND SOLID HARDENING
Anthropic shipped Claude Code v2.1.76 with MCP elicitation, sparse monorepo worktrees, new hooks, a model effort knob, and a long list of reliability fixes. Th...
Anthropic shipped Claude Code v2.1.76 with MCP elicitation, sparse monorepo worktrees, new hooks, a model effort knob, and a long list of reliability fixes.
The release adds MCP elicitation so servers can request structured input mid-task, plus new Elicitation/ElicitationResult hooks for governance and interception. You also get a PostCompact hook, a /effort command to tune model effort, and a CLI --name flag. Monorepo users get worktree.sparsePaths for claude --worktree, which uses git sparse-checkout to pull only the directories you need release notes.
A bunch of papercuts were addressed: spurious context-limit errors on 1M sessions, LSP plugin registration order, tmux clipboard over SSH, Bash permission matching with #, and a circuit breaker for auto-compaction retries. The 1M-context fix pairs well with broader 1M GA and standard pricing for Opus/Sonnet noted here context window roundup.
Structured elicitation and hooks make it easier to enforce guardrails and collect the exact inputs your tools need.
Sparse worktrees cut monorepo checkout time and disk use, making agent loops faster and cheaper.
-
terminal
Benchmark claude --worktree with and without worktree.sparsePaths on a large monorepo (cold checkout time, disk usage, first-command latency).
-
terminal
Prototype an MCP server that requests form fields, then use Elicitation/ElicitationResult hooks to validate, transform, or block payloads before execution.
Legacy codebase integration strategies...
- 01.
Enable worktree.sparsePaths for known hot paths in legacy monorepos to speed CI/dev containers without restructuring the repo.
- 02.
Add policy in hooks to gate risky tool calls; rely on the new compaction circuit breaker and Bash permission fixes to reduce flaky runs.
Fresh architecture paradigms...
- 01.
Design tools around MCP elicitation for typed inputs and smoother multi-step flows from day one.
- 02.
Standardize on 1M-context models and validate that skill invocations no longer trigger false context-limit errors.