Maya: Hey builders! Here's what's on the radar today on Claude Code Review. Mods come to Claude Code, a prompt-audit command flags stale instructions, a repair release tightens permissions, Projects gives each request its own agent, and mods can now read selected text. I'm Maya.

James: And I'm James. Let's get into it.

Maya: On October first, Anthropic added mods to Claude Code. Mods are short pieces of TypeScript, shipped inside plugins, that run before, after, or in place of actions like sending a prompt, calling a tool, or asking permission. Claude can even write one for you.

James: That's the heart of it, Maya. You install them with slash plugin in the CLI or desktop app, and slash plugins lists what you've got. Examples include token weather, blast radius showing what a delete would remove, and replay theater stepping through recent diffs.

Maya: And the constraint matters, James. Mods are not sandboxed; they have the same machine access as Claude Code, so install only from sources you trust. On Team and Enterprise, a built-in mod called sec-default loads first and limits what risky user mods can do.

James: Exactly right. Now, story two. Running slash claude-api prompt-audit makes Claude Code review your standing instructions, your skills, your CLAUDE.md, your AGENTS.md, and flags rules that no longer help Claude Opus five point five.

Maya: And it targets prompts written for older models, James. It flags verification rituals like double-check your work, ALL-CAPS NEVER rules, pressure language, and hard word caps. Each finding carries the file, the line, quoted evidence, and a proposed diff you can apply.

James: That's the useful part, Maya. The audit itself changes nothing on disk; you ask Claude to apply edits in whatever batches you prefer. It lives inside the bundled claude-api skill, so it refuses to run if you disabled that skill through skillOverrides.

Maya: Right, and a walkthrough suggests running it right after a model upgrade, since moving from Opus five to five point five shifted context handling, effort levels, and subagent caching. Just expect some false positives on dynamic paths.

James: Fair warning. Story three: Claude Code two point one point two eight nine, published October third, is mostly a repair release, and the most consequential fixes are in permissions. Deny and ask rules are the settings that block a command or force a confirmation.

Maya: And several paths where those rules did not hold got closed, James. Under sandbox auto-allow, a rule missed commands behind an environment variable prefix, like TZ equals HOME rm dash rf build, and missed a bare variable assignment before the command.

James: Those are sneaky gaps, Maya. Read deny rules also skipped files at-mentioned or selected through a symlink. On stability, the terminal froze on short code blocks with many unclosed script tags or deeply nested substitutions, and those freezes are fixed.

Maya: And one reversal worth noting: Anthropic reverted a two point one point two eight eight change to claude auth status that may have made sign-outs more frequent. For agent builders, it adds agent dot spawn for teammates and new idle and waiting states.

James: Good catch. Story four: Claude Code Projects is a mode where a single chat, run by an orchestrator that coordinates the project, delegates work to several agents at once. You give it context, a GitHub repo, a file, a folder, or a Google Drive folder.

Maya: And for each request, James, the orchestrator creates a thread and assigns it a dedicated agent. Threads appear in a right-hand panel with status, in progress or waiting for input. You can open any one to see its steps while the others keep working.

James: That's the model, Maya. The main chat itself runs on Low effort by default, since it mostly coordinates threads. You can override that in Project settings, and the post is asking users who had to raise it to explain where Low falls short.

Maya: Last one, James. Claude Code two point one point two eight eight, released October second, adds dollar dot ui dot selection for mods. The call returns the text you last selected in fullscreen mode, plus the transcript row if the selection sits inside one.

James: And two typing changes land, Maya. A prompt cleared with Ctrl-C can be recovered; press Up on the empty prompt and the draft returns, pasted text and images included. Also, a dangerous rm inside a bash dash c script no longer runs without a prompt.

Maya: Exactly, and session reliability got a long list. Non-interactive sessions now continue from the partial response after a mid-response API timeout, and thinking-only responses are retried. Also, claude project purge is now just claude purge, with the old name still working.

Maya: That's your Claude Code Review daily brief. Check claudecodereview.ai for the full stories. Keep building.