Claude Code 2.1.215: /verify and /code-review Are No Longer Automatic — What Every Operator Pipeline Must Change

Claude Code 2.1.215 ships one change. That is enough to break a category of operator workflows that were treating /verify and /code-review as passive safety nets running quietly in the background.

Starting with this release, neither skill runs automatically. You invoke /verify or /code-review yourself, or they do not run at all.

What happened

The 2.1.215 changelog entry is concise: "Claude no longer runs the /verify and /code-review skills on its own; invoke them with /verify or /code-review when you want them."

Until this release, Claude Code could elect to run these skills autonomously — most commonly at the end of a task, when it judged that a verification or review step made sense given the work it had just done. The judgment call was the model's. The execution was implicit.

In 2.1.215, that changes. Both skills become explicit-only. The model no longer decides unilaterally to run them. A human or an orchestration layer must issue the slash command.

Why it matters for AI engineering teams

The practical impact depends entirely on how your team integrated these skills.

Teams that always invoked them manually — no change. The slash commands work the same way.

Teams that relied on Claude to run them during autonomous sessions — this is the break. If your workflow looked like this:

1. Send a task prompt to Claude Code in background mode or CI.
2. Claude completes the task.
3. Claude runs /verify on its own before finishing, surfacing lint errors, test regressions, or policy violations.
4. The session report reflects the verification result.

— step 3 no longer happens unless your prompt or orchestration layer explicitly triggers it.

The claude code 2.1.215 verify code-review operator change is most consequential for:

• CI pipelines that pipe prompts into claude -p or an equivalent background invocation and assume a verification pass is part of the completion.
• Agentic task orchestrators — tools like Claude Agents, custom harnesses, or OpenAI-compatible proxy flows that send multi-step coding tasks and read the result — where the verification step was implicitly delegated to the model.
• Quality-gate workflows where /code-review was used as a passive catch before committing or merging, and teams never added an explicit invocation because Claude was doing it automatically.

For teams in that last category, the silent removal is the most dangerous outcome: no error, no warning, just a missing check.

The router/operator angle

This change is a shift from implicit quality assurance to explicit governance. From a gateway and orchestration perspective, that is actually the right direction — but it requires a deliberate transition, not an assumption that nothing changed.

There are two concrete things operators need to do:

1. Audit your prompt templates and task definitions.

If your CLAUDE.md or task prompts contain language like "verify your work before finishing" or "run a review before you commit," those natural-language instructions may have been sufficient to trigger automatic skill execution. They still set intent, but the skill no longer self-activates. Replace implicit language with explicit slash command calls in your prompts:

After completing the changes, run /verify and report any findings before finishing.

2. Update your orchestration scripts.

If you're using claude -p with a programmatic pipeline, add an explicit verification step either inside the prompt or as a follow-up tool call. Harnesses that read the session output for verification results must also confirm the verification actually ran — not assume it.

On the routing layer, if you're managing Claude Code sessions through a gateway or proxy that inspects session events, the absence of a skill_run event for /verify after 2.1.215 is expected behavior, not a telemetry gap.

What to watch

Anthropic has been progressively separating skill execution from autonomous model judgment across multiple Claude Code releases. The EndConversation tool (2.1.214), the subagent-cap changes (2.1.212), and now this /verify//code-review explicit-only gate all move in the same direction: the model does less on its own; operators specify more explicitly.

This is a good architecture for teams that need predictable, auditable behavior. It requires updating workflow assumptions that were written when Claude Code had more autonomy.

Check your session logs for the last two weeks. If /verify or /code-review was appearing in session event traces without an explicit invocation in your prompts, audit those workflows now. The gate is gone.