Claude Code v2.1.206: /doctor Trims CLAUDE.md, /commit-push-pr Auto-Allow Git Push, Background Agent Upgrade

Anthropic shipped Claude Code v2.1.206 with a stack of quality-of-life improvements and critical fixes for MCP servers, background agents, and git integrations.
New Features
/cdpath suggestions: The/cdcommand now shows directory path completions, matching/add-dirbehavior./doctortrim proposal: A new check suggests cutting content from checked-inCLAUDE.mdfiles that Claude can derive directly from the codebase./commit-push-prauto-allow git push: Now automatically allowsgit pushto the repo's configured push remote (remote.pushDefault, or the sole remote when only one is configured) in addition toorigin.- Gateway
/login: Now supports Anthropic-operated public gateway endpoints. - EnterWorktree confirmation: Prompts before entering a git worktree outside
.claude/worktrees/. - Background agent upgrade: Background agents now upgrade in the background immediately after a Claude Code update, avoiding the slow stale-session upgrade on attach.
Fixes
- Expired login error message — now properly prompts for
/logininstead of showingThere's an issue with the selected model
. claude --resume/--continue— keyboard input now works on startup.- MCP
request_timeout_ms— per-server timeout settings from--mcp-configor.mcp.jsonare now respected; long-running MCP tool calls no longer time out at the 60s default in fresh sessions. CLAUDE_CODE_EXTRA_BODY— now correctly passed toclaude agents/--bgbackground workers.- OAuth MCP servers — no longer require manual re-authentication after a single failed token refresh.
--permission-prompt-tool— no longer crashes withMCP tool not found
on cold start before the server connects./modelpicker — fixed rows showing wrong price, stopped quoting first-party list prices on providers that don't bill them, and fixed misplaced server-provided rows when entitlement/allowlist restrictions drop a reference row.- Desktop sessions — no longer stuck showing
running
after mid-turn slash commands. - Windows agents view — keyboard input no longer ignored when a setup prompt appears before bare
claude --resume. claude rm— now properly removes the job from the daemon roster; deleted jobs stay deleted./remote-control— now explains how to sign in instead of showingUnknown command
when logged out.- Left arrow — now steps back out of a phase or agent in the workflow detail view.
/status— no longer lists the same broken-install warning twice.- LSP plugins — false
disused plugin
tips and skewed telemetry fixed. /doctorupdate check — for Homebrew installs, now compares against the cask's channel instead of the settings channel.- Fullscreen jump-to-bottom pill — fixed Ctrl+End suggestion on macOS, rebound chords now show, and the pill no longer wraps over the transcript.
- Bedrock — fixed multi-minute startup hang when using
awsCredentialExporthelper on networks with restricted egress.
Improvements
/code-review— findings quality improved onclaude-opus-4-8across all effort levels.- Agents view status column — now uses full terminal width instead of truncating at 64 characters.
- Ctrl+X in agents view permanently removes a completed session; deleted background jobs stay deleted. Sessions no longer render twice.
📖 Read the full source: GitHub Claude-Code
👀 See Also

Claude 4.6 Opus Reasoning Distilled to 14GB for Apple Silicon via MLX Quantization
A developer has quantized a Qwen 3.5 27B model distilled from Claude 4.6 Opus reasoning trajectories from 55.6GB to 14GB using MLX for Apple Silicon, achieving ~16 tokens/sec on an M4 Pro while maintaining the model's analytical reasoning capabilities.

OpenClaw Kubernetes Operator with Embedded Ollama Support
A community member has created an OpenClaw Kubernetes operator that includes embedded Ollama support, allowing AI agents to run with local models in the same namespace. The setup includes installation commands, configuration details for both local and cloud Ollama models, and dashboard access instructions.

How Clawdbot Coordinates 6 AI Agents with a Production-Stable Work Queue
Clawdbot's team built a work queue system to coordinate 6 AI agents (design, code, marketing, ops) for their AI-operated store. The system features atomic task claiming, a state machine, retry logic with backoff, task chains, heartbeat tracking, and a daemon orchestrator.

Relational Memory for LLMs: Three-Layer System Models User Relationships
An open-source Python tool that adds relational memory to LLMs by modeling user-AI relationships across seven psychological dimensions, using a three-layer narrative structure instead of flat fact storage.