Claude Code v2.1.301 and v2.1.302: Auto Mode Classifier Model, Idle Compaction Toggle and Subagent Defaults

Claude Code shipped two releases in two days, and the second one changes defaults rather than just fixing bugs. v2.1.301 landed on October 15, 2026 and v2.1.302 on October 16, according to the Claude Code releases page. Between them they add a setting for the model behind auto mode's safety classifier, a switch for idle compaction, a cap on concurrent hooks, and new subagent defaults across the SDK, desktop app and VS Code.

This post goes through the items most likely to affect your cost, your safety settings and your scripts. We read the release notes for both versions and did not run the builds ourselves, so behavior descriptions are the notes' wording. For background on the settings involved, see our guides to Claude Code permission modes and managing the context window.

TL;DR: what to check

Auto mode: a model setting for the classifier

Auto mode lets Claude Code approve routine tool calls with a classifier instead of asking you every time. v2.1.302 adds CLAUDE_CODE_AUTO_MODE_MODEL, which, per the release note, chooses the model the client-side auto mode classifier uses. It does not change the model for server-side auto mode.

Why it matters: the classifier is a safety layer. A cheaper or faster model there lowers cost and latency; a stronger one may catch more risky calls. The notes do not say which is better, and we have no measurements, so treat any choice as something to test on your own tasks.

Related fixes in the same release:

• Auto mode with the server-side classifier now compacts a conversation that is too long for the safety check, instead of denying every checked tool call.
• OpenTelemetry tool_decision events now carry classifier_category in most classifier blocks, and v2.1.301 added classifier_verdict and permission_mode, so you can tell classifier decisions from permission rules in your traces.
• Prompt suggestions no longer spend usage credits on models your plan covers only with credits.

If you export telemetry, these attributes make it possible to audit what the classifier blocked. See our permission modes explainer for how auto mode fits with the other modes.

Idle compaction gets a switch

v2.1.302 adds an Idle compaction row to /config for turning idle compaction off or back on, and the notices shown after an idle compaction now say how to turn it off. Compaction summarizes a long conversation to free context; doing it while you are away can save tokens and keep sessions usable, but it can also surprise you when you return to a summarized history.

Related context fixes in the two releases:

• A "Prompt is too long" error when stopping a turn mid-compaction is fixed.
• Switching to a model with a smaller context window now summarizes the conversation with the previous model first (v2.1.301), and a fix stops a case where that switch could spend usage credits without asking.
• Conversations are no longer compacted repeatedly on a model you did not pick behind an LLM gateway that reports a different model name.
• A prompt cache miss when a tool added mid-session was loaded again through ToolSearch or after compaction is fixed.

If you tune cost by hand, the cache and compaction items are the ones to watch. Our context window guide covers the mechanics.

Subagent defaults changed

This is the change most likely to alter behavior without you noticing. According to the v2.1.302 notes:

• Fork subagents (subagent_type: "fork") are now on by default in claude -p, the Agent SDK, the desktop app and VS Code, as they already were in the terminal. CLAUDE_CODE_FORK_SUBAGENT=0 turns them off, and claude mcp serve never offers them.
• Background by default in desktop and VS Code. The desktop app and VS Code now always run the main conversation's subagents in the background, as the terminal does, instead of letting Claude choose. A subagent's own Agent calls in claude -p, the Agent SDK, the desktop app and VS Code run in the foreground by default.
• Hooks cap. Prompt and agent hooks now run at most 12 of one event at a time; the rest wait.
• Permission default on third-party providers. claude -p and Agent SDK sessions with no permission mode now start in default mode on third-party providers, with telemetry off or on a fresh install; --permission-mode auto keeps auto.

For CI jobs and SDK scripts, the last item is the one to test: a job that quietly relied on a permissive starting mode may now stop to ask. For how to use subagents deliberately, see our guide to subagent effort levels.

Security and robustness fixes worth noting

• Image reads. The Read tool no longer loads an image file of any size, or a pipe or device named like an image, entirely into memory; image sources over 128 MB are refused.
• Binary files. Read no longer treats binary files such as .xlsm, .parquet, .epub and .heic as text (v2.1.301).
• Worktrees. Quitting no longer removes a worktree made by a WorktreeCreate hook without asking when it may hold unpushed commits (v2.1.301), and worktree isolation works again under git 2.56 or later when a stray .git file points at a missing directory.
• MCP. Servers that negotiate protocol version 2026-07-28 but omit resultType, ttlMs or cacheScope no longer show no tools or fail every call. Our MCP 2026-07-28 spec post explains that revision.
• Writes and config. claude plugin and other subcommands no longer exit before a queued settings write finishes, and Windows projects on exFAT or FAT32 drives no longer have every file write refused.
• Error reporting. Grep and Glob now return an error saying the search failed instead of reporting no matches when the search program fails unexpectedly.
• Sign-in. Several fixes avoid unexpected sign-outs and misleading messages, such as a full disk blocking login refresh.

The Grep and Glob change deserves a second look if you have scripts that treated "no matches" as success: a failure used to look identical to an empty result.

Mods, VS Code, cloud sessions and Claude Tag

The notes are long, and much of it concerns surfaces beyond the terminal:

• Mods. v2.1.300 through v2.1.302 keep extending the mod API: virtual prompt text, agent transcript views, multi-line text fields, and changes to how hook .catch handlers and $.env.set behave. If you build mods, see our mod tutorial and the list of open-source mods.
• VS Code. v2.1.301 added LaTeX math rendering and a "Get more usage" link for Pro and Max plans; v2.1.302 fixes MCP sign-in links with query strings and a misleading browser-connected banner.
• Cloud sessions. Fixes cover rewind, pull request status flicker, routine runs shown as completed after a session stopped, and a scheduled routine rerunning from the start and possibly posting the same report twice. A new banner explains why a routine did not start, such as hitting the concurrent cloud session limit.
• Claude Tag in Slack. A long list of fixes, including image types that never reached Claude, link formatting that broke copied commands, direct messages dropped when a lookup stalled, and a four-tab reorganization of admin settings (Bundles, Connectors, Skills and plugins, Channels).
• Remote Control, self-hosted runners and Windows. Smaller fixes for git reftable repositories, hung subagent panes and PowerShell command lookup.

Practical upgrade checklist

1. Read the changed-defaults list above, then run your usual CI prompts once in a staging job.
2. Decide whether you want idle compaction on, and set it in /config.
3. If you use auto mode, decide whether the classifier model setting is worth testing for cost or caution, and compare on a handful of real tasks.
4. If you export OpenTelemetry events, add the new classifier attributes to your dashboards.
5. If you run hooks, check whether anything fires more than 12 concurrent instances of one event.
6. If you use claude.ai/design, note the December 14 closing date; our Dashboards and Motion post covers the migration steps.

For an extra safety layer on what agents can do regardless of the harness defaults, AgentBeam, the agent security platform from the explainx.ai team, is built to stop AI agents before they take dangerous actions.

Related reading

• Claude Code subagent effort levels
• Claude Code permission modes explained
• Claude Code context window and limit management
• MCP 2026-07-28 stateless spec
• Claude Dashboards, Motion and the Claude Design shutdown
• How to build a Claude Code mod

Sources: Claude Code releases (v2.1.301, v2.1.302) and the anthropics/claude-code repository.

Release details are accurate as of October 17, 2026; later versions may change these defaults.