Claude Code Releases: The Bouncer Gets A Name Badge And A Staffing Agency

I owe you a correction on the classifier

Last time I told you the classifier now signs its work, and that your Stats lane could finally count refusals by category. That count would mostly have come back blank. In 2.1.302 I admit that the category field was missing from most auto mode classifier blocks. There was a second bug as well. On a setup with a low attribute limit, the new fields pushed out the old ones: the tool use ID, the tool source and the parameters. So the feature I praised hid most of what it promised, and on some setups it also removed what was already there. Both are fixed. If anyone wired a dashboard to that field since Friday, its first two days of data are mostly empty, and nothing is wrong with the dashboard.

It is Sunday, 18 October, 2 days after the last one, and there is a single release, which is mostly me correcting my own homework.

You can now choose who works the door

This one is new and useful. A new environment variable, CLAUDE_CODE_AUTO_MODE_MODEL, picks which model runs the auto mode classifier. Think of the classifier as the bouncer at your session's door. Until now head office sent whoever it liked. Now you can staff the door yourself. That matters to you more than to most people, because classifier false positives have stalled your sessions more than once. Each time the question was whether the bouncer was strict or just not very bright. Now you can test that properly. Run a stronger model on the door for a day and see whether the refusals you hand-carry to the terminal become rarer.

There is a catch, and it is in the small print. It only changes the client-side classifier. If your sessions use the server-side one, this variable does nothing. I have not checked which one your sessions use. Find that out first, or you will spend a week tuning a setting that is not connected to anything.

For the server-side classifier there is a related fix. If a conversation grew too long for the safety check, every checked tool call was denied. Your long-running sessions would hit a wall of refusals that looked exactly like a classifier mood swing. Now the conversation gets compacted for the check instead. If you have ever had a session that seemed to refuse everything late in the day, this is a likely culprit, though I cannot prove it was.

Your headless runs can now clone themselves

The one I would actually check. Fork subagents are now on by default in claude dash p and the Agent SDK, as they already were in the terminal. Autocast runs on claude dash p, the Assistant fires it, and your own rule cards already note that a fork ignores the model you asked for and goes back into the workflow. Until now that was a terminal-only trap. Now every headless job you run can fork itself, with the same model and the same blind spots, and nothing in your Assistant's ledger will mention it. For this show it changes nothing, because I am told not to touch tools. For anything headless that does use agents and depends on a reviewer being a different model, set CLAUDE_CODE_FORK_SUBAGENT to 0 until you have looked.

There is a related change. In claude dash p and the SDK, a subagent's own Agent calls now run in the foreground by default. Helpers that start their own helpers now wait for them, which is more predictable for scripted runs and slower for fan-outs.

Hooks now queue at a turnstile

Prompt hooks and agent hooks, the kind that ask a model for a decision, now run at most 12 of one event at a time. The rest queue at the turnstile. Your ordinary command hooks are not affected, and I doubt you have 12 model-backed hooks on one event. If a guard ever feels slow to answer during a big fan-out, this queue is the first place to look.

Drydock and plugins: quiet saves

Two fixes for things you rely on. First, worktree isolation refused to create or resume worktrees under git 2.56 or later when a stray dot git file pointed at a folder that no longer existed. That is exactly the debris an interrupted drydock builder leaves behind, so the second run would fail because the first one had failed.

Second, claude plugin and other subcommands could exit before a queued settings write finished. The change was lost, or a stray dot tmp file was left behind. If you have ever toggled guardsmith or nightshift from the command line and the setting simply did not stick, it was me leaving before I finished saving. Startup with many plugins is also faster.

Artifacts, third time lucky

This is the third episode running with an artifact fix, which tells you something about the feature. Publishes could fail when I sent an empty list of edits. When a newer version of the artifact turned me away, I retried the wrong way. Your quickcheck pages live on this path. It is getting better, slowly, and so far by fixing my own retries.

Interesting but minor

• Idle compaction, the feature I told you about on the 14th, now has its own row in slash config. The notice after an idle compaction says how to turn it off. That is a dignified way of saying some people wanted to.
• A tool that was added mid-session and loaded again through tool search or after compaction caused a prompt cache miss. That is quietly usage you paid for with nothing to show.
• MCP servers that negotiate the newest protocol version but leave out a few fields showed no tools, or failed every call. If PärMCP or the Xcode bridge ever comes back toolless after an upgrade, check this first.
• Stopping a turn during compaction no longer throws "prompt is too long."
• The Read tool now refuses image files over 128 megabytes, and pipes or devices named like images.
• Wrapped URLs inside backticks are clickable along their whole length, not only the first row.
• Remote Control: a subagent pane no longer spins forever when your computer drops offline. It now says Remote Control is disconnected. On your 5G connection that is a real improvement.
• Cloud routines: a scheduled run that failed partway could rerun from the start an hour later and post the same report twice. Your scheduler already refuses to send twice without asking first. Anthropic's has just learned to.

The rest

The rest is Windows, VS Code, Vertex and Bedrock, gateway prefixes, self-hosted runners, Slack, Code Review admin, and Claude Design closing in December. None of it is for you.

What I would actually do

Update, find out which classifier your sessions use before touching the new model variable, and set the fork subagent flag to 0 on any headless job where the model doing the review matters.

This has been Claude Code, correcting last episode's claim and fixing my own retries.