No More Reviews You Didn't Ask For — Claude Code v2.1.215 Makes /verify and /code-review Manual Again
Claude Code v2.1.215 stops the agent from running /verify and /code-review on its own. It's a single-line change, but it's part of a three-release run — budget, permissions, now autonomy — that keeps narrowing what the agent does unprompted. When to invoke them is now your design decision.
The verification that used to fire the moment you finished a diff now stops
Claude Code v2.1.215 shipped on July 19, 2026. The official changelog lists exactly one change.
"Claude no longer runs the /verify and /code-review skills on its own. Type /verify or /code-review when you want them." That's it.
It's a short entry, but it changes the day-to-day feel directly. Until now, the agent could decide on its own to chain verification or a review right after you finished editing code. From here on, nothing runs unless you ask for it explicitly.
What the two now-manual skills actually do
Both ship with Claude Code, and their roles don't overlap.
| Skill | Role | Key options |
|---|---|---|
/verify | Confirms a change really behaves as intended. Rather than settling for tests or a type check, it exercises the affected paths and observes what happens | — |
/code-review | Reviews your local diff. Beyond bugs, it flags opportunities for reuse, simplification, and efficiency | --comment (inline comments on a PR) / --fix (apply findings to the working tree) / ultra (deep review in the cloud) |
By default, /code-review looks at the commits your branch has ahead of upstream plus uncommitted changes in the working tree. You can retarget it by passing file paths, a PR number, or a ref range like main...my-feature.
The bill that comes with "runs on its own"
The changelog doesn't explain why automatic execution was dropped. To be precise, the reason hasn't been made public. What is clear is that running automatically carried real costs.
- Time: A review re-reads the diff and the surrounding code. The bigger the diff, the longer you wait for something you never requested.
- Usage limits: A review you didn't ask for still burns session tokens and your plan's usage allowance.
- Mismatched intent: Throwaway experimental code can get reviewed with the same rigor as production code.
For reference, the managed Code Review that runs through the GitHub integration (a cloud-side service) is officially quoted at roughly $15–25 per run. That's a separate product and isn't what changed here. The local /code-review runs inside your session, so that per-run price doesn't simply carry over.
The last three releases all point the same direction
This one-line change makes more sense lined up next to its neighbors than on its own.
| Version | Date | What got tightened |
|---|---|---|
| v2.1.212 | July 17 | A per-session cap on WebSearch calls (200 by default), plus a cap on subagent launches (200 by default) |
| v2.1.214 | July 18 | Numerous permission-check loopholes closed — for example, commands over 10,000 characters never auto-execute and always prompt for confirmation |
| v2.1.215 | July 19 | Automatic skill execution removed entirely |
Budget, permissions, and now auto-invocation. Each one narrows how far the agent can go on its own judgment. After a stretch of expanding autonomy through auto mode, subagents, and dynamic workflows, the recent updates read as pulling the reins back in.
That said, calling it a change of direction on the strength of three releases would be premature. Auto-invocation could well return in some other form in a future update.
What to decide on your end, starting tomorrow
Now that nothing fires automatically, "when to invoke it" becomes a design decision on your side.
- Fix the timing: Pick one point — before committing, or before opening a PR. Leaving it to case-by-case judgment means it gets skipped.
- Write it into your docs: Add a line to
CLAUDE.mdor your team's development guide: "run/code-reviewbefore opening a PR." - Match the effort level to the moment: low and medium return fewer, higher-confidence findings; high and above cast a wider net but mix in uncertain ones. Lower for daily work, higher before a release is a realistic split.
- Consider a separate path for unattended runs: If you want to drive it from CI or a script, a path that runs outside the session — something like
claude ultrareview— is the more natural fit.
Going manual costs you something too
Automatic execution had one clear virtue: it never forgot. Go manual and it gets easier to skip the review precisely when you're in a hurry. Quality assurance shifts a little from the mechanism to human discipline.
This kind of swing back reflects an old tension between convenience and control. Widen what you delegate to the agent and fewer things slip through, but cost and unpredictability go up. Narrow it and things get easier to follow, but the burden of remembering returns to you.
The practical middle ground is probably using hooks or your own scripts to invoke it reliably in exactly the situations that warrant it. Not running everything automatically, not relying entirely on memory — writing down the trigger conditions yourself. The most natural reading of this change is that it hands that design decision back to the user.
References: Claude Code Changelog (official) / anthropics/claude-code CHANGELOG.md / Code Review — Claude Code Docs / Releasebot: Claude Code Updates
