Claude Code 2.1.302: Classifier Settings and Client Fixes

Claude Code v2.1.302 was released on 16 October 2026, a day after v2.1.301. The headline is a new environment variable, CLAUDE_CODE_AUTO_MODE_MODEL. It chooses the model the client-side auto mode classifier uses, and it does not…

Published:

Updated:

Claude Code 2.1.302: Classifier Settings and Client Fixes

Claude Code v2.1.302 was released on 16 October 2026, a day after v2.1.301. The headline is a new environment variable, CLAUDE_CODE_AUTO_MODE_MODEL. It chooses the model the client-side auto mode classifier uses, and it does not change the model for server-side auto mode.

This is a client release with settings and fixes. It is not a new model.

The short answer

  • Classifier model: CLAUDE_CODE_AUTO_MODE_MODEL picks the client-side classifier model only.
  • Idle compaction: /config has a new “Idle compaction” row to turn it off or back on.
  • Settings writes: a fix for changes that could be lost when a subcommand exits early.
  • Images: the Read tool now refuses image sources over 128 MB.
  • Not “unchanged”: the notes do not show that the server-side classifier is unchanged. The same release lists an improvement to it.

What is in the release?

The release is v2.1.302, published 16 October 2026 (the page shows day and month without the year), with about 60 items. The previous releases are v2.1.301 on 15 October and v2.1.300 on 14 October. The four items covered here were all first listed in 302, and none appear in 301 or 300.

What are the four new items in 302?

Change Release-note wording
Classifier model variable “Added CLAUDE_CODE_AUTO_MODE_MODEL to choose the model the client-side auto mode classifier uses
it does not change the model for server-side auto mode”
Idle compaction “Added an Idle compaction row to /config for turning idle compaction off or back on
the notices shown after an idle compaction now say how to turn it off”
Settings writes “Fixed claude plugin and other subcommands sometimes exiting before a queued settings or config write finished, which could lose the change or leave a .tmp file behind”
Image reads “Fixed the Read tool loading an image file of any size, or a pipe or device named like an image, entirely into memory
image sources over 128 MB are now refused”

Is the server-side classifier “unchanged”?

No, that claim is not supported. The notes say only that the new variable does not change the model for server-side auto mode. They do not say the server-side classifier is unchanged overall.

The same release lists an improvement to it: “a conversation too long for the safety check is now compacted, instead of every checked tool call being denied.”

What other auto-mode and compaction fixes are in 302?

  • Fixed classifier_category missing from OpenTelemetry tool_decision events for most classifier blocks. Also fixed new attributes pushing out tool_use_id, tool_source and tool_parameters when OTEL_LOGRECORD_ATTRIBUTE_COUNT_LIMIT is low.
  • Fixed a “Prompt is too long” error when stopping a turn during compaction.
  • Fixed /autocompact and /effort saving to the wrong model when a gateway names Opus 4 or 4.1 with a prefix such as anthropic/claude-opus-4-1.
  • Fixed a prompt cache miss when a tool added mid-session was loaded again through ToolSearch or after compaction.

302 also lists a notice that standalone Claude Design closes on 14 December 2026. It lists behavior changes too (fork subagents on by default, hooks capped at 12 concurrent), plus Vertex, Bedrock, PowerShell and plugin fixes.

What was already in 301?

  • Added classifier_verdict and permission_mode to tool_decision events to tell classifier decisions apart from permission rules. Also added classifier_category when OTEL_LOG_TOOL_DETAILS=1. That is why the 302 classifier_category fix is a follow-up rather than a new item.
  • Fixed Read treating binary files such as .xlsm, .parquet, .epub and .heic as text (with no 128 MB limit).
  • Improved switching to a model with a smaller context window: the previous model now summarizes a too-long conversation first.

What do the official docs say?

  • Idle compaction: the idleCompaction setting defaults to true (requires v2.1.290 or later), and turning auto-compact off turns idle compaction off too. CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS defaults to 200,000 tokens, and values below 100,000 are raised to 100,000.
  • Classifier model: CLAUDE_CODE_AUTO_MODE_MODEL does not appear in the parts of the env-var and settings docs that were readable. The nearest entry is CLAUDE_CODE_AUTO_MODE_SERVER, which controls whether the server reviews auto mode actions. The “server-side classifier review” section of the permission-modes page was not read.
  • Not in the docs read: the 128 MB image limit and the settings-write behavior. Rely on the release notes for both.

What should you do next?

We suggest these steps. They are ours, not part of the release notes.

  1. If you use auto mode, decide whether you want a different model for the client-side classifier. Server-side auto mode is not affected by this variable.
  2. If compaction notices bother you, set idleCompaction from /config.
  3. Update if you lost settings changes when plugin commands exited.

Questions and answers

Is 302 a new model?

No. It is a client release with settings and fixes.

Does the new variable change the server-side classifier?

No, according to the notes. It affects the client-side classifier only.

What is the Read image limit?

128 MB. Larger image sources are refused.

How do I turn off idle compaction?

Use the new Idle compaction row in /config. The setting idleCompaction defaults to true, and turning auto-compact off also turns idle compaction off.

Terms in plain language

  • Auto mode classifier: the check that decides whether a tool call in auto mode is safe to run.
  • Compaction: shortening a long conversation so it fits in the model’s working memory.
  • Environment variable: a named setting you give your system before starting a program.
  • OpenTelemetry: a standard way for software to report logs and events.

Sources

Facts last checked against the linked pages on 18 October 2026. Release-note wording is quoted from the GitHub release.

All sections All articles

Latest articles