v2.1.152

Source: Anthropic Claude Code Date: 2026-05-27 URL: https://github.com/anthropics/claude-code/releases/tag/v2.1.152

Summary

Claude Code v2.1.152 ships three new extension points that deepen the platform’s composability. Skills and slash commands can now set disallowed-tools in frontmatter to remove tools from the model while active — programmable behavioral constraint from the skill layer. A MessageDisplay hook event lets hooks transform or hide assistant text as it renders — a programmable presentation layer. /reload-skills and SessionStart hook reloadSkills: true allow dynamic skill installation mid-session. Auto mode no longer requires opt-in consent. /code-review --fix auto-applies findings. Fallback model support provides session resilience. 20+ bug fixes target background agent lifecycle, plugin deduplication, and remote MCP connectivity.

Implications

• Harness extensibility thread: disallowed-tools is the first mechanism for skills to narrow the model’s tool surface — governance built into the composition layer, not just admin settings. Combined with the Workflow tool (v2.1.147) and admin hard_deny rules (v2.1.136), three separate control surfaces now constrain what the model can do.
• Auto mode consent removal is a policy signal: Anthropic trusts the auto mode mechanism enough to stop gating it behind explicit consent. The operational boundary widens as the behavioral boundary tightens.
• MessageDisplay hook creates a programmable layer between model output and human display — the first hook that mediates what the user sees rather than what the model does. Enables output filtering, formatting, or redaction by third-party code.
• Background agent reliability remains the dominant bug-fix theme: stale thinking-block signatures, cancelled-subagent permission crashes, plugin registry rebuilds dropping branch-tracking updates. The failure modes being fixed are from long-running, composed agent sessions.