Give the diff another pair of model eyes.

Define the review target

Name the change and the kind of failure you care about. “Review my uncommitted changes” is workable; “Review my uncommitted changes for double settlement under retries and missing regression tests” gives the reviewer a sharper job.

Supply enough evidence

A diff alone may omit contracts, tests or callers needed to judge correctness. Khimba’s local preparation can include selected supporting files while respecting ignore rules. Inspect the displayed path list and remove irrelevant or sensitive material before confirmation.

Evaluate the findings

Prioritise specific claims tied to code paths. Reproduce suspected bugs and ask whether the finding is introduced by the change or already exists. Treat style preferences separately from correctness defects. False positives are normal in AI review, so do not make mechanical changes merely because two models agree.

Example request

“Ask DeepSeek to review my uncommitted changes. Focus on correctness, security boundaries, idempotency and missing tests. Cite files and functions for every substantive finding, and distinguish blockers from optional improvements.”

The reviewing model cannot browse URLs, invoke tools or fetch unapproved files. The opinion is a bounded reading of the supplied text, not a complete security audit.