Claude Code v2.1.302: Auto Mode Model Env Var, Idle Compaction Toggle, and 30+ Fixes
Claude Code v2.1.302 is out. Mostly a bug-fix release with two new knobs worth knowing about, plus a pile of fixes around plugins, MCP, and Windows file writes.
What's new
CLAUDE_CODE_AUTO_MODE_MODELlets you pick the model the client-side auto mode classifier uses. It does not affect server-side auto mode — those are separate code paths.- A new Idle compaction row in
/configto turn idle compaction off or back on. The notices shown after an idle compaction now include instructions for turning it off. /design-syncand/design-loginnow print a notice that standalone Claude Design (claude.ai/design) shuts down by December 14, 2026.
Notable fixes
- Read tool memory blowup: the Read tool was loading image files of any size — including pipes or devices named like an image — entirely into memory. Sources over 128 MB are now refused.
- Vertex / Bedrock mid-session switch: requests were failing with
diagnostics: Extra inputs are not permittedafter switching providers mid-session. - Gateway-prefixed Opus:
/autocompactand/effortwere saving to the wrong model when a gateway named Opus 4 or 4.1 with a prefix, e.g.anthropic/claude-opus-4-1. - Prompt cache miss: a tool added mid-session and reloaded via ToolSearch or after compaction no longer busts the prompt cache.
- Plugin subcommands:
claude pluginand friends could exit before a queued settings/config write finished, losing the change or leaving a.tmpfile behind. - reftable ref format: plugin marketplaces re-cloned on every refresh and newly installed git plugins showed version
unknownwhen git defaults to reftable. Also fixesbad object refs/headsfetches on self-hosted runners reusing a reftable checkout. - PowerShell tool: now finds commands in a plugin's
bin/folder, and no longer errors when a search that finds nothing (rg,grep,findstr,git grep) is piped intoSelect-Objector another formatting cmdlet. - MCP protocol 2026-07-28: servers negotiating that version but omitting
resultType,ttlMs, orcacheScopeno longer show zero tools or fail every call. - Credit-billed models: prompt suggestions are now skipped after a reply billed to credits, so you stop burning usage credits on models your plan pays for only with credits.
- OTel:
classifier_categorywas missing fromtool_decisionevents for most auto mode classifier blocks. Also fixed newtool_decisionattributes pushing outtool_use_id,tool_source, andtool_parameterswhenOTEL_LOGRECORD_ATTRIBUTE_COUNT_LIMITis low. - Windows exFAT/FAT32: every file write was refused or prompted for projects on those drives, including sessions started with
--plugin-dir. - Worktree isolation: no longer refuses to create or resume worktrees under git 2.56+ when a stray
.gitfile points at a nonexistent directory.
Improvements
- Plugin command and agent files now take far fewer filesystem calls to load — biggest win on Windows, noticeable at startup with many plugins enabled.
- Server-side auto mode: a conversation too long for the safety check is now compacted instead of having every checked tool call denied.
- Background sessions reopened from another folder with no LLM gateway now say why, instead of a truncated message.
If you run Claude Code with plugins on Windows, or switch between Anthropic API and Vertex/Bedrock mid-session, this is worth pulling.
📖 Read the full source: GitHub Claude-Code
👀 See Also

Analysis of 413K AI Agent Runs Reveals What Makes Them Succeed
An analysis of 413,278 AI software engineering agent runs from the CoderForge-Preview dataset shows that human software engineering best practices often harm agent performance. The data reveals specific patterns that separate successful from failing runs on the same problems.

Normalization of Deviance in AI: Why Your Agentic System Will Fail
The AI industry repeats Challenger-like cultural failures: treating unreliable LLM outputs as safe because nothing bad happened yet. Real examples of agents formatting hard drives, wiping databases, and creating GitHub issues.

What's missing in the 'agentic' story: a well-defined user agent role
Mark Nottingham argues that current AI agents lack a clear user agent role, creating a trust gap between what users expect and what agents actually do.

Claude's Five-Seat Minimum Creates Privacy Gap for Solo Practitioners
Anthropic's business-tier privacy protections require a five-seat minimum, forcing solo practitioners to either pay for empty seats or use consumer plans with inadequate privacy terms. This gap contrasts with Google Workspace and OpenAI Business Plans, which offer enterprise-grade privacy at single-seat pricing.