Claude Code 2.1.215 changed a small behavior with a large operational consequence: /verify and /code-review no longer run automatically. You must invoke them explicitly.
In short, Claude Code now leaves verification invocation to the user, so teams should name, run, and record their checks explicitly.
That is not a reason to trust agent output less. It is a reason to make your team's validation steps explicit.
An implicit check is easy to forget, hard to audit, and difficult to reproduce across tools. An explicit check gives the team a named step, observable evidence, and a clean failure boundary.
Treat verification as a workflow stage
A reliable agent loop has four stages:
- Define the exit criteria before implementation.
- Let the agent change the smallest useful surface.
- Run the relevant checks explicitly.
- Review the evidence before accepting the result.
The important word is relevant. A formatting-only change does not need an integration suite. An authentication change should not stop at lint. Match the check to the risk.
Ask hard questions before code exists
A pre-implementation challenge can expose missing constraints before they become patches. Matt Pocock's grill-me skill on Skillselion (GitHub) is one way to turn vague requirements into batches of concrete questions.
Use it to resolve questions such as:
- What behavior must remain unchanged?
- Which failure states need tests?
- What evidence will prove the task is complete?
This is cheaper than discovering the same ambiguity after the implementation has spread across several files.
Keep feedback close to implementation
The test-driven-development skill on Skillselion (GitHub) keeps the feedback loop close to the change: write a failing test, make it pass, then refactor with the suite still green.
TDD is not a substitute for the final verification step. It covers the behavior you chose to specify. The explicit /verify or /code-review pass asks a different question: did the completed change satisfy the repository's broader contract?
Make the evidence part of the handoff
A useful completion note should say more than "tests pass." Record:
- the exact checks that ran;
- their exit status;
- any checks intentionally skipped and why;
- residual risks that still need human judgment.
For example:
Verified: unit tests, typecheck, and the focused auth integration suite.
Skipped: full end-to-end suite because this environment has no payment sandbox.
Residual risk: session expiry behavior still needs staging confirmation.
This gives the next reviewer something falsifiable. It also prevents a green statement from hiding a missing check.
CI is still the final shared contract
Local agent verification is fast feedback. CI is the team-visible, repeatable contract. Keep required checks in CI even when an agent ran the same commands locally.
The practical response to Claude Code 2.1.215 is simple: do not recreate hidden automation. Put verification into the task definition, invoke it deliberately, and attach the evidence to the handoff.
That is a stronger workflow than hoping the agent remembered.
Building an explicit validation loop? Browse Skillselion for reusable planning, testing, and code-review skills, then add the checks your repository actually requires.
Top comments (0)