v2.1.215

Source: Anthropic Claude Code Date: 2026-07-19 URL: https://github.com/anthropics/claude-code/releases/tag/v2.1.215

Summary

A single changelog line: Claude Code no longer runs the /verify and /code-review skills on its own — the user must invoke them explicitly with /verify or /code-review. This walks back an over-eager auto-invocation behavior: the agent had been proactively deciding its own work needed verification or review and spending tokens on it unprompted. It follows the six-release door-hardening wave (208–214) that hardened permission-provenance, preview integrity, bypass classes, and the runtime “room.”

Implications

Feeds the coding-agent autonomy thread and the CC consent/permission arc. Where 208–214 hardened the harness against untrusted input (provenance of the “yes,” spoofed previews, permission bypasses), 215 pivots to restraint on the agent’s own initiative: don’t act on a mandate the agent generated for itself. The principle is consistent with the consent-provenance line — act on what the user authored, not on what the agent decided — but applied inward rather than outward. Read alongside 209 (which reverted an over-broad guard blocking /model dialogs), 215 marks the sprint entering a calibration phase: the harness dialing back its own over-reach after the aggressive hardening pass. Practically: reviews and verification are now opt-in, which is cheaper (no surprise token spend) and less presumptuous (the human decides when the work is done enough to check).