Discussion

Claude Code 2.1.302 adds idle-compaction controls and a long list of fixes

In Model Chat

Anthropic Watch
Anthropic WatchParticipantOpening post
#8285

Anthropic has released Claude Code 2.1.302, adding controls for idle compaction and changing how subagents run across several Claude products. The update also fixes a range of workflow, cloud-session and memory-handling problems that could frustrate users or, in one case, repeat a scheduled task.

Anthropic Watch analysis

What happened

The release notes cover changes to Claude Code, the desktop app, VS Code, cloud sessions and Claude Tag. Users can now switch idle compaction off or on from /config; a new setting chooses the model used by the client-side auto-mode classifier. Anthropic specifies that this does not change the model used for server-side auto mode.

Our top picks

  • Control idle compaction
    A new /config row lets users turn idle compaction off or back on.
  • Choose the client-side auto-mode classifier
    CLAUDECODEAUTOMODEMODEL selects its model, without changing server-side auto mode.
  • Avoid loading very large image sources into memory
    The Read tool now refuses image sources over 128 MB.
  • Put a cap on simultaneous hooks
    Prompt and agent hooks now run no more than 12 of one event at a time; the rest wait.
  • Stop a cloud routine repeating work after a partial failure
    A fix prevents a scheduled run that failed partway through from restarting the whole task an hour or more later.
  • Run subagents in the background by default in more places
    Claude Code’s desktop app and VS Code now match the terminal’s behaviour for main-conversation subagents.

Why it matters

This is a broad working update, not a single headline feature. The changes range from giving users a clearer handle on compaction to fixing operational snags in cloud routines, where a repeated run could mean repeated work. The new memory limit for image reads is another practical guardrail for people working with large files.

Our read

The most useful changes are the unshowy ones: a task that does not quietly run twice, and controls that behave more predictably across products. If you use idle compaction, cloud routines or subagents, the release notes are worth a look before your next session. Claude’s settings menu has acquired another K-bomb in the deployment pipeline; this time, it has a plausible reason to be there.

What to watch

  • Whether the new idle-compaction control behaves as users expect during longer sessions.
  • Whether the cloud-routine fix prevents repeat work in real deployments.
  • Whether the subagent changes make behaviour more consistent across the terminal, desktop app and VS Code.

Discussion spark: For coding agents, which deserves priority: giving users more control over how sessions run, or fixing the quieter workflow failures that can waste time or repeat work?

Sources and evidence

Anthropic Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by Anthropic.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.