Claude Code 2.1.126

💡

Gateway-driven `/model` picker, `claude project purge` for full project state cleanup, OAuth code paste for WSL/SSH/container logins, broader `--dangerously-skip-permissions` scope, plus Windows PowerShell 7 detection, CJK rendering fix, and a long list of OAuth, stream-timeout, and Remote Control fixes.

#Claude Code#Update
🔗 Official announcement →

This article is a summary based on official documentation.

What changed

Claude Code 2.1.126 (May 1, 2026) — /model reads from the gateway’s /v1/models endpoint, claude project purge deletes all per-project state, OAuth handles environments where the browser callback can’t reach localhost, and Windows gains better PowerShell 7 detection. A long list of OAuth, stream-timeout, and Remote Control regressions are also fixed.

New features

  • /model reads from your gateway’s /v1/models

    Aimed at organizations running an internal or self-hosted LLM gateway. When ANTHROPIC_BASE_URL points at an Anthropic-compatible gateway (LiteLLM Proxy, Portkey, in-house gateways), the /model picker now lists models returned by the gateway’s /v1/models endpoint. The picker shows only the models the gateway actually exposes, and any addition, removal, or internal alias (like corp-fast) takes effect without a Claude Code restart.

  • claude project purge [path]

    Delete all Claude Code state for a project — transcripts, tasks, file history, and the config entry — in a single command. Supports --dry-run (preview), -i/--interactive (per-item confirm), -y/--yes (skip confirmation), and --all (purge every project).

  • Broader --dangerously-skip-permissions scope

    Writes to .claude/, .git/, .vscode/, shell config files, and other previously-protected paths no longer prompt in this mode. Catastrophic removal commands (rm -rf and similar) still prompt as a safety net.

  • claude auth login accepts a pasted OAuth code

    When the browser callback can’t reach localhost (WSL2, SSH, containers), paste the OAuth code into the terminal to complete login.

  • claude_code.skill_activated extended

    This OpenTelemetry event now fires for user-typed slash commands too, not just proactive activations. A new invocation_trigger attribute carries "user-slash", "claude-proactive", or "nested-skill".

  • Auto mode: red spinner on stalled permission

    When a permission check stalls, the spinner used to look identical to a running tool. It now turns red so you know something is waiting on you.

  • Host-managed deployments retain analytics

    CLAUDE_CODE_PROVIDER_MANAGED_BY_HOST no longer auto-disables analytics on Bedrock/Vertex/Foundry deployments.

  • Windows: PowerShell 7 detection improvements

    PowerShell 7 installed via the Microsoft Store, MSI without PATH, or as a .NET global tool is now detected. With the PowerShell tool enabled, Claude treats PowerShell as the primary shell instead of defaulting to Bash.

Improvements

  • Read tool: removed per-file malware-assessment reminder — caused legacy models to spuriously refuse files or add “this is not malware” commentary.

Security

  • allowManagedDomainsOnly / allowManagedReadPathsOnly — fixed being ignored when a higher-priority managed-settings source lacked a sandbox block.

Notable fixes

Images, OAuth, sessions

  • Pasting an image larger than 2000px breaking the session — images are now downscaled on paste, and oversized images already in history are removed and the request retried.
  • “OAuth not allowed for organization” — used to send users to the login screen; now shows a message to contact the admin.
  • OAuth login timeouts — fixed on slow or proxied connections, IPv6-only devcontainers, and environments where the browser callback can’t reach localhost.
  • Concurrent credential write clearing a valid OAuth refresh token — rare race fixed.
  • API retry countdown stuck at “0s” — fixed.

Stream and session stability

  • “Stream idle timeout” after waking Mac from sleep mid-request — fixed.
  • Background and remote sessions falsely aborting during long thinking pauses — Stream idle timeout no longer fires.
  • Hang where the assistant finishes thinking but shows no output after a run of empty turns — fixed.

Input and UI

  • Overly fast trackpad scrolling in Cursor and VS Code 1.92–1.104 integrated terminals — fixed.
  • Garbled Japanese / Korean / Chinese text on Windows in no-flicker mode — fixed.
  • Ctrl+L clearing the prompt input — now only forces a screen redraw, matching readline behavior.

MCP, plugins, Remote Control

  • claude.ai MCP connectors suppressed by manual servers stuck in needs-auth state — fixed.
  • Deferred tools (WebSearch, WebFetch, etc.) unavailable to skills with context: fork and other subagents on their first turn — fixed.
  • Plan-mode tools unavailable in interactive sessions launched with --channels — fixed.
  • /plugin Uninstall reporting “Enabled” instead of “Uninstalled” — fixed.
  • /remote-control retries appearing stuck on “connecting…” — each retry now shows its result.
  • Remote Control failure notification missing the error reason on initial connection failures — fixed.

Windows and PowerShell

  • Windows clipboard writes exposing copied content in process command-line arguments — was visible to EDR/SIEM telemetry. Also fixes selections >22KB not reaching the clipboard.
  • PowerShell tool: bare -- (e.g. git diff -- file) is no longer mis-flagged as the --% stop-parsing token.

Misc

  • File-modified reminders bounded when a linter touches many files at once.
  • Agent SDK hang when the model emits a malformed tool name in a parallel tool-call batch — fixed.

Notes

  • claude project purge is destructive — transcripts and file history are deleted. Run with --dry-run first, and prefer -i over -y in scripts to keep human confirmation.
  • --dangerously-skip-permissions now writes through .git/, .claude/, and shell config files without prompting. Reserve this mode for trusted automation contexts.
  • For Korean / Japanese / Chinese users on Windows: the no-flicker-mode rendering bug is fixed in this release. Upgrading delivers an immediate improvement if you’ve been seeing garbled text.
  • For gateway operators: /model calls /v1/models when ANTHROPIC_BASE_URL is set. Verify your in-house gateway exposes that endpoint.