DEV Community

Ahab
Ahab

Posted on • Originally published at indieseek.co

Claude Code v2.1.215 Verify and Code Review Workflow

#claudecode #ai #programming #productivity

Claude Code verification after v2.1.215: run /verify and /code-review explicitly

Quick answer

Claude Code v2.1.215, released July 19, 2026, no longer invokes the /verify and /code-review skills on its own. The commands still exist, but a task that merely looks finished should not be expected to trigger either skill automatically.

If those skills were part of your quality gate, replace the hidden assumption with an explicit finish sequence:

implement
-> run deterministic checks
-> /verify
-> /code-review 
-> fix findings and rerun checks
-> human merge decision
Enter fullscreen mode Exit fullscreen mode

Do not treat this as evidence that Claude Code stopped running tests in every situation. The official release note is narrower: it changes autonomous invocation of two named skills. Your prompts, hooks, scripts, and CI may still run other checks.

Who this is for

This guide is for developers who use Claude Code interactively and assumed that a completion request, CLAUDE.md rule, or the agent's own judgment would automatically launch its bundled verification or review workflow.

It is also useful for teams building unattended coding-agent jobs. In that setting, a slash command is helpful evidence, but deterministic commands and CI should remain the enforceable gate. For broader instruction-file placement, see the AGENTS.md, CLAUDE.md, and Copilot instructions guide. For untrusted repositories, establish the coding-agent sandbox boundary before running any project command.

What changed on July 19

Anthropic's v2.1.215 release contains one focused behavior change: Claude no longer runs /verify and /code-review by itself; users should invoke the command they want.

That fits Claude Code's documented skill model. Skills are normally available to both the user and Claude, while disable-model-invocation: true makes a skill user-only. The release effectively gives these two bundled workflows a manual boundary even though other skills may still be selected automatically.

The practical risk is not that the commands disappeared. It is that an older workflow may reach “done” without the review stage its author thought was implicit.

Separate the three gates

Gate Best evidence What it catches Should it block completion?
Deterministic checks Exit codes and test reports Build, lint, type, unit, integration, or product assertions Yes
Claude Code skills /verify and /code-review output Missing validation, suspicious diff behavior, reuse and correctness issues Yes when your workflow requires them
Human review Reviewed diff, risk decision, approval Product intent, acceptable tradeoffs, deployment judgment Yes for material changes

Keep the gates separate. A clean /code-review result does not prove the app builds. A green test suite does not prove the change solves the intended problem. A human approval without recorded checks makes regressions harder to diagnose.

Build an explicit finish workflow

1. Name deterministic commands before implementation

Put the smallest relevant commands in the task or repository instructions:

Required checks
- npm test -- --runInBand
- npm run lint
- npm run build
- one smoke test for the changed route
Enter fullscreen mode Exit fullscreen mode

Do not write only “verify the change.” That phrase leaves the evidence shape undefined. Include the expected command, environment, and one behavior-level assertion.

2. Invoke /verify as its own step

After implementation and deterministic checks, run:

/verify
Enter fullscreen mode Exit fullscreen mode

The public v2.1.215 note names the command but does not document a stable argument contract. Use the bare command unless your installed version shows supported arguments, and record what it actually ran. If /verify only summarizes earlier evidence, that summary is not a substitute for missing test output.

3. Give /code-review an intentional scope

Claude's Code Review documentation says the local command reviews commits ahead of the upstream branch plus uncommitted changes by default. It can also accept a file, pull request number, branch, or ref range.

For a branch intended to merge into main, make the comparison visible:

/code-review main...my-feature
Enter fullscreen mode Exit fullscreen mode

Use /code-review --comment only when posting findings to the pull request is intended. Use --fix only after reviewing the proposed findings and preserving unrelated worktree changes.

4. Rerun the deterministic gate after fixes

Review-driven edits can invalidate earlier test results. After addressing findings, rerun the relevant build and tests, then record the final commit and outcome. Do not reuse a green result from before the last code change.

Local review is not managed PR review

The local /code-review command runs inside a Claude Code session. Anthropic's managed GitHub Code Review is a separate service with repository-level triggers such as PR creation, every push, or manual @claude review comments. The v2.1.215 release note does not announce a change to those managed triggers.

Choose by workflow:

  • Interactive local task: run deterministic checks, /verify, and local /code-review before handing off.
  • Unattended agent task: enforce scripts and CI outside the model; treat skill output as an additional artifact.
  • GitHub pull request: keep required CI and human approval; optionally add managed Code Review for a second pass.
  • Tiny documentation change: use format, link, and build checks; run code review only when the diff warrants it.

Use a compact evidence block

Add this to the task handoff or pull request:

Verification evidence
- Claude Code version:
- Final commit:
- Deterministic commands and exit codes:
- Product smoke test:
- /verify result:
- /code-review scope and result:
- Findings fixed:
- Remaining risk / human decision:
Enter fullscreen mode Exit fullscreen mode

This makes a skipped skill visible instead of silently relying on model behavior. It also keeps an unattended run auditable when logs are compacted or the session is resumed later.

Common mistakes

  • Assuming “finish and verify” will still make Claude invoke both skills automatically.
  • Adding a CLAUDE.md sentence that asks the model to choose /verify, then treating it as an enforceable hook.
  • Confusing local /code-review with managed GitHub Code Review.
  • Running review before tests, then never rerunning tests after review-driven edits.
  • Letting an AI review replace required CI or human approval.
  • Posting raw logs that contain tokens, environment variables, or private repository URLs.

FAQ

Were /verify and /code-review removed?

No. The release says Claude no longer runs them on its own and tells users to invoke them explicitly.

Will Claude Code still run tests?

It may run commands requested by your prompt or instructions. The confirmed change only concerns automatic invocation of the two named skills, so verify the actual command log instead of generalizing the release note.

Can CLAUDE.md force these skills to run?

Treat repository instructions as guidance, not an enforcement mechanism. For a hard gate, call the command explicitly in an interactive session or enforce deterministic scripts in CI.

Does /code-review replace human review?

No. Use it to find and prioritize evidence. Product intent, risk acceptance, secrets exposure, and merge authorization still require deterministic policy and human judgment.

Sources

Top comments (0)