In response to a user who prefers switching model families for adversarial code review, bcherny points out that '/code-review low' outperforms other models on internal evals at less than $0.01 per run.
In response to DHH's skepticism about agent permission models, a Claude Code team member highlighted that auto mode actively enforces hard instructions in the system prompt. For example, telling the model not to push to git will be reliably respected, and the mode has never degraded model behavior.
The watermark detection API that accompanies Anthropic's new content watermarking feature will be available at no cost.
Chrome extension sessions now carry over to desktop, web, and mobile. Conversations are saved, and skills and connectors work in the browser. Available on Max and Team plans today, rolling out to Pro in coming weeks.
LLM-generated bugs have shifted from off-by-ones to system design flaws and missing context. Adversarial code review, via a simple prompt or the built-in /code-review slash command (with optional severity levels like low or medium), is a practical way to surface them.
A bug has been confirmed in the /code-review slash command that causes it to run inside the current bloated session context instead of spinning up a subagent with fresh context as documented. A fix is on the way.
The /code-review command uses no subagents at the `low` effort level and scales up to 10 parallel subagents at higher effort levels.
The /code-review command has used a forked agent (fresh context) since version 2.1.218, released July 21. Users experiencing stale session context issues should run `claude update` to get the fix.
For users hitting token limits with /code-review, the `low` effort level is a token-efficient option. One user reported consuming over 700k tokens with `/code-review medium` on just 200 lines of code.
