Claude Code takes 27 minutes rebasing PR

This is the kind of anecdote that matters more than a flashy benchmark: Claude is now being trusted with the boring, fragile parts of software work, not just code generation.

• –Rebase work is a good proxy for real repo competence because it forces the model to deal with history, conflicts, and project conventions
• –27 minutes is still slower than a human doing a simple rebase, but the value is in offloading attention, not just raw speed
• –If this is Claude Code in practice, it shows the product is moving from autocomplete-style assistance toward autonomous maintenance
• –The downside is obvious: long-running agent tasks need trust, logging, and rollback discipline before teams will hand them more critical branches
• –For AI coding tooling, this is a sign the moat is shifting from “can it write code?” to “can it safely finish repo work?”