Claude Code v2.1.236: New Default Model Env Var, Sandbox Fixes, and More
Claude Code v2.1.236 is out, focusing on developer workflow improvements and critical bug fixes. Key additions include a new environment variable for default model selection, cross-session idle notifications, and tighter macOS sandbox behavior.
New ANTHROPIC_DEFAULT_MODEL Environment Variable
You can now set ANTHROPIC_DEFAULT_MODEL to define the starting model for new sessions. Unlike ANTHROPIC_MODEL, a /model selection overrides this default and persists across restarts. This gives teams more flexible control over model defaults per environment without locking users in.
Cross-Session Idle Notifications
The SendMessage API now supports notify_when_idle on macOS and Linux. It lets a Claude Code session request another local session to send a one-shot notice when it next goes idle. This is opt-in and requires no polling, useful for orchestrating background tasks that depend on another session's state.
Sandbox Improvements on macOS
On macOS, wildcard read-deny rules (e.g., **/.env) now correctly take precedence inside allowed read regions. The rules also cover directory contents and can't be bypassed by renaming the file. This strengthens security for sandboxed file access.
Notable Fixes
This release fixes several regressions from 2.1.229 onward, including:
- Clipboard copy, background housekeeping, background sessions, and local MCP logs breaking when the working directory was deleted.
- Fullscreen renderer failing permanently after a single bad start—now falls back to the classic renderer.
- /model picker rendering taller than the terminal—now scrolls within the window height.
- SendMessage calls being rejected due to a malformed closing tag.
- Unhandled promise rejections when subprocesses like
powershell.exefail on WSL with Windows interop disabled (2.1.234 regression).
Improved Auto Mode and Other Enhancements
Auto mode now treats Monitor allow rules like Bash commands when active, and the classifier uses consistent defaults across Bedrock, Vertex AI, Foundry, and telemetry-disabled setups. Git status checks can no longer be fooled by status.showUntrackedFiles=no. Other changes:
/goalsessions now check in after 30 minutes (then 1h, 2h) when idle behind background work./usageshows usage-credits spend for Team/Enterprise members.- SIGTERM in print/SDK mode no longer records interrupted turns or synthetic tool denials.
- Slash-command typos report errors instead of fuzzy matching.
- Session recaps capped at 400 characters to prevent runaway output.
📖 Read the full source: GitHub Claude-Code
👀 See Also
Context Quality Degradation in AI Agents: Hallucination Rates Increase with Token Count
Testing shows hallucination rates increase from ~3% at 10K tokens to ~28% at 200K tokens, with recall accuracy dropping below 90% for early-session information once context exceeds 50K tokens.
OpenClaw 3.31 Update Resets Agent Permissions and Settings
OpenClaw update 3.31 automatically disabled all agent tools, computer access permissions, and sub-agents, requiring manual re-enabling in Settings. The update also changed how permission requests work, no longer prompting for approval during use.
Claude Code v2.1.208: Screen Reader Mode, Vim Remaps, Memory Leak Fixes, and More
Claude Code v2.1.208 adds opt-in plain-text screen reader mode, vim insert-mode sequence remaps, a process wrapper for corporate launchers, and fixes dozens of bugs including memory leaks in long sessions.
Kimi K2.6 vs Claude Opus 4.7: A Practical Coding Showdown on a Minetest Mod + Google Sheets Integration
A developer tested Kimi K2.6 and Claude Opus 4.7 on building a Minetest bounty board mod with a TypeScript backend and Google Sheets logging. Opus succeeded in both tasks; Kimi passed the local task but failed the integration. Costs: Opus ~$3.59 local, $16.03 integrated; Kimi $0.39 local, $5.03 failed.