Tags
An admin pins Claude Code to the one model version the team vetted. In 2.1.283, setting the model match to "exact" keeps every newer release blocked until that admin adds it to the allowed list.
Plugin submissions for Claude got their own portal, where builders can track the review and see usage. Over the coming weeks, those plugins land in one discovery experience across Claude and Claude Code.
"3x faster in two weeks" is what the ClaudeDevs account says about claude.ai, and the team credits Claude itself with measuring and debugging the slowdowns. The write-up comes with the exact prompts and methods they used.
Your agent switches thinking off for quick steps to save tokens, and on Opus 5.5 that same request comes back as a 400 error. Anthropic's migration guide flags it as one of four breaking changes for teams moving from Opus 5.
A canary test revealed that Claude Code skips AGENTS.md whenever telemetry is off, because loading it waits on a remote flag. Nothing warns you, and per the linked GitHub issue, Bedrock and Vertex users hit the same wall.
"Certain actions always happen," the Claude Code hooks guide promises, but only if your script exits 2. Exit 1, the usual Unix failure code, blocks nothing, and the agent carries on.
11 open-source bugs slipped past CodeRabbit's production reviewer and got caught by Opus 5.5, though Opus 5.5 missed nine that the old reviewer found. Swapping models changes which bugs get through.
Anthropic's browser sandbox for the Claude API answers to Playground now, and every call comes back with its tokens, cost and latency, though it keeps no prompt history to go back to.
Of the 33 tools AgentsMesh generates for, 17 read a root AGENTS.md and Claude Code makes 18. Though only when no CLAUDE.md sits in your working directory or above it: keep both files and the AGENTS.md is ignored entirely.
You clone a repo with an AGENTS.md in it and no CLAUDE.md, and Claude Code 2.1.277 reads it as the project instructions. Bedrock, Vertex and Foundry don't get that fallback yet.
Only 1 of 216 Claude Code skills passed clean 50 of 87 public Claude Code subagents declare no tools list, so each inherits the caller's whole toolbox, Bash included. Permission prompts still gate it, but a reader-only worker carries shell.
Close the laptop and the handed-off report keeps moving: Cowork and chat are merging into one Claude, rolling out to Pro and Max over the next few weeks, with Claude Docs, Slides and Design in beta on all paid plans.
Five new request headers let an LLM gateway see which Claude Code request is which: class, agent type, compaction state. They stay off until you set CLAUDE_CODE_GATEWAY_HINT_HEADERS=1 in 2.1.273.
Testers poking at Claude Code with CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1 claude are trying out what Anthropic now calls Claude Mods, and the team says it ships on the scale of weeks in lieu of days or months.
CI jobs at Anthropic grew 25x in six months, and three quick fixes to the test selection service bought 70 days, then 29 days, then less than a day. The rewrite took one engineer three weeks.
"Scored, reproducible results": Claude Code 2.1.269 adds claude plugin eval, which runs a plugin's eval suite against Claude Code and hands back JSON plus an HTML report.
Every turn had been failing with HTTP 400 on third-party Anthropic-compatible endpoints since 2.1.265, and Claude Code 2.1.268 traces it to a regex in the Artifact tool's input schema that those endpoints reject.
With "auto", Claude reviews each tool call in Claude Managed Agents against your intent in user.message events and decides whether to run it, deny it, or ask you for input. Anthropic announced the mode on X.
Per a post on X, a pane in the Claude Code desktop app pops out into its own window, so the diff can sit on a second screen while Claude works in the main one. Anthropic's docs mention drag-and-drop pane layout, not detachable windows.
An unreadable managed allowlist in Claude Code admitted everything through, allowedHttpHookUrls and allowedChannelPlugins included. Release 2.1.267 flips that failure mode to admitting nothing instead.
