Issue #33
5 min read

Claude breaks into OpenAI

Claude Code now reads AGENTS.md when no CLAUDE.md exists, auto mode classifier moved server-side to stop billing overhead, and your claude.ai skills now sync to terminal sessions. Anthropic is moving toward an IPO while Claude drives a quarter of its internal engineering work, and attackers recently used Claude in an intrusion at OpenAI.

Projects redesigned: from folder to conversation — Claude Blog

Highlight

Claude Code now reads AGENTS.md when no CLAUDE.md exists

If your repository already carries an AGENTS.md for other coding agents, Claude Code will now read it in projects that have no CLAUDE.md. You can switch which file is used under "Project instructions" in /config. The behavior is not yet available on Bedrock, Vertex or Foundry.


Worth Knowing

Auto mode classifier moves server-side and stops billing overhead

If you use auto mode on the Claude API, Enterprise, Bedrock, Vertex, Foundry or a gateway, the safety classifier now runs on the server by default and you are not charged for its token overhead, with a warning when a billed fallback happens. Set CLAUDE_CODE_AUTO_MODE_SERVER=0 to opt out on the cloud platforms, and check the new Auto mode server row in /status to see which path your session is on. See the classifier billing docs.

Your claude.ai skills and plugins sync to terminal sessions

Skills and plugins you enable on your claude.ai account now appear in terminal sessions signed in with that account, so you no longer maintain two separate sets. Turn it off with syncClaudeAiSkills: false or syncClaudeAiPlugins: false.

New key interrupts the turn and sends queued messages

Press Ctrl+Enter (or Ctrl+X Ctrl+S) to interrupt what Claude is doing and push all your queued messages through at once. Sent and queued messages stay gray until the model receives them, so you can tell what has landed.

TaskOutput tool removed in favor of reading output files

Claude now reads a background task's output file with the Read tool instead of the deprecated TaskOutput tool. If you set taskOutputMaxChars or TASK_MAX_OUTPUT_LENGTH, those values no longer do anything.

npm plugin installs no longer run package scripts

Plugins from an npm source are now fetched with npm pack --ignore-scripts and integrity-verified, so a package's install scripts no longer execute on your machine. You can also pass --accept-command <sha256> to claude plugin install and claude plugin update to approve exactly the command a previous --json run displayed rather than blanket -y.

VS Code adds per-change diff review and task controls

You can now accept or reject each change individually in the proposed-change diff tab instead of taking a whole edit at once. Background shells and other running tasks appear in the agent map with a Stop control and a typed /tasks command, and responses get a Copy button plus /copy.


Under the Hood

Resume and startup crashes from malformed state fixed

Sessions no longer crash at launch or on --resume when ~/.claude.json holds a malformed theme, mcpNeedsAuthNoticed or customApiKeyResponses value, or when a saved transcript contains a malformed message block, and the endless "unexpected tool_use_id" retry loop now self-heals or ends with a clear error.

Rendering, caching, and headless startup speedups

Large diffs and long transcripts render faster with fewer slow frames, per-turn slowdowns from language-server diagnostics and hook progress updates are gone, and SDK and -p sessions no longer wait on the per-directory CLAUDE.md lookup before the first turn.


From Anthropic


In the News

A week weighted toward configuration and cleanup: project instructions can now come from AGENTS.md, auto mode's classifier overhead stops showing up on your bill, and a long list of resume-path crashes are gone.