Three things worth knowing
- Claude Code v2.1.215 ships one behavioral change:
/verifyand/code-reviewno longer trigger automatically. Both run only when you invoke them. - The skills themselves are unchanged. Only who decides when to run them has moved from the agent's judgment to yours.
- Signed binaries ship for eight platform targets: macOS, Linux, Linux musl, and Windows, each in
arm64andx64.
Autonomous skill invocation is one of those things that looks helpful until it isn't. Claude Code would decide mid-session that your work deserved a verification or review pass, add extra turns, and bill you for checks you hadn't asked for and sometimes didn't need. v2.1.215 ends that.
The release is a single-line changelog. But single-line behavioral changes in agentic tools tend to matter more than multi-feature releases, because they shift who's in control.
What Happened
Version 2.1.215 landed on July 19. The full changelog: "Claude no longer runs the /verify and /code-review skills on its own; invoke them with /verify or /code-review when you want them."
Both skills previously could fire autonomously during a session. Now they require explicit invocation. Everything else about the skills is unchanged: the commands work the same way, run the same checks, and produce the same output. Only the trigger moved.
The release packages prebuilt binaries for darwin-arm64, darwin-x64, linux-arm64, linux-x64, and musl variants of both Linux targets, plus win32-arm64 and win32-x64. Each asset ships with a SHA-256 hash; the release also includes SHASUMS256.txt.sig for signature verification before you install.
Key Features
/verifyis now opt-in: Claude no longer decides when your work needs a verification pass. You invoke it when you want it./code-reviewis now opt-in: Same change. The review skill runs only when you type the command.- No new syntax to learn: The commands, their behavior, and their output are unchanged. Only the trigger moved.
- Eight platform builds:
darwin-arm64,darwin-x64,linux-arm64,linux-x64,linux-musl-arm64,linux-musl-x64,win32-arm64,win32-x64. - Signed checksums:
SHASUMS256.txtandSHASUMS256.txt.sigship with every release for verification before install.
Why It Matters
Autonomous skill invocation costs you in two ways. The obvious one is token spend: extra model turns you didn't request. The less obvious one is timing. When Claude decides to run a review mid-session, it interrupts the flow at a point you didn't choose, sometimes during a mechanical edit that didn't need one, and not after the logic change where a review would have been useful.
This change makes agent behavior predictable, which matters especially if you're running Claude Code in CI pipelines or standardized team workflows. An agent that improvises extra steps is a liability in scripted environments. Explicit invocation means the pipeline does what it's configured to do, nothing more.
The direction here is worth tracking as a signal for how Anthropic is thinking about agent autonomy more broadly. Claude keeps its judgment for the coding work itself. The meta-work, verifying and reviewing, now needs a human trigger. That's a deliberate line, and it suggests more changes in this direction are likely.
Example Use Case
You're mid-refactor on a TypeScript service: renaming a module, updating thirty import paths, no logic changes. Before v2.1.215, Claude might decide that the pass warranted a verification cycle, adding turns to a change you already knew was mechanical. On v2.1.215, the session stays scoped to the edits you asked for.
Then, when you finish the part that touches business logic, you run /code-review yourself before opening the PR, and /verify if you want a correctness pass on top of that. The skills become deliberate checkpoints in your workflow rather than surprises in the middle of it. One review at the point of maximum value beats three scattered wherever the agent guessed they belonged.
Competitive Context
Agentic coding tools split on how much the agent should do without being asked. v2.1.215 is a clear vote for explicit control over agent initiative, at least on these two skills. Claude keeps its autonomy for coding work; the layer above it now requires a human trigger.
For teams choosing between Claude Code and more autonomous alternatives, that trade-off is the real differentiator. If you want an agent that self-polices every change without prompting, this release moves Claude Code away from that. If you want an agent that does what you asked and stops there, it moves closer. Given how widely deployed Claude Code is, even a one-line behavioral change like this reaches a lot of daily workflows.
My Take
The change is small. The signal it sends is not. Autonomous skill invocation has been one of the legitimate complaints about agentic tools: the agent doing extra work you didn't ask for, at the worst possible moment, and billing you for it. Moving /verify and /code-review to explicit invocation is the right call.
The question is whether this extends further. Autonomy in agentic tools has been a race to show off capability rather than a careful design of what the agent should decide versus what the developer should decide. If Anthropic keeps drawing that line deliberately, release by release, it will matter more than any individual feature.
Run your software agents at scale
Cosmos gives your agents the context, tools, and feedback loops they need to get better with every workflow.
Written by
Paula Hingel
Technical Writer
Paula writes about the patterns that make AI coding agents actually work — spec-driven development, multi-agent orchestration, and the context engineering layer most teams skip. Her guides draw on real build examples and focus on what changes when you move from a single AI assistant to a full agentic codebase.