Claude Code v2.1.136: Hard Deny for Auto Mode, MCP OAuth Fixes, and 40+ Bug Fixes

✍️ OpenClawRadar📅 Published: May 8, 2026🔗 Source
Claude Code v2.1.136: Hard Deny for Auto Mode, MCP OAuth Fixes, and 40+ Bug Fixes
Ad

Anthropic shipped Claude Code v2.1.136 with two notable additions and a massive tail of bug fixes — over 40 issues resolved. Here's what matters.

New Features

  • CLAUDE_CODE_ENABLE_FEEDBACK_SURVEY_FOR_OTEL — re-enables the session quality survey for enterprises capturing responses through OpenTelemetry. Explicit opt-in for telemetry-aware feedback collection.
  • settings.autoMode.hard_deny — new auto mode classifier rule that blocks an action unconditionally, ignoring user intent or allow exceptions. For strict guardrails where certain operations must be denied regardless of context.

Critical Fixes

  • MCP server loss after /clear — MCP servers configured in .mcp.json, plugins, and claude.ai connectors silently disappeared after /clear in the VS Code extension, JetBrains plugin, and Agent SDK. Now persistent.
  • OAuth token rotation race condition — a rare login loop where concurrent credential writes could overwrite a freshly-rotated OAuth token, forcing repeated logins. Fixed to prevent credential corruption.
  • MCP OAuth concurrent refresh — refresh tokens were lost when multiple MCP servers refreshed concurrently. Users with several remote MCP servers should no longer need daily re-authentication.
  • API 400 on redacted thinking blocks — extended thinking emitting a redacted thinking block after a tool call caused a 400 error. Now handled correctly.
  • --resume/--continue with underscored paths — sessions weren't found when the project path contained underscores. Path parsing fixed.
  • Plan mode bypass with Edit(...) allow rule — plan mode was not blocking file writes when a matching Edit(...) allow rule existed. Now enforced.

Platform-Specific & Developer Experience

  • WSL2 image paste — pasting images from Windows clipboard now works via a PowerShell fallback when xclip/wl-paste cannot read image data.
  • Plugin hooks crash on cache cleanupStop/UserPromptSubmit hooks failed when cache cleanup deleted a version still in use. Fixed by retaining versions until hooks finish.
  • Slash command dialog visual consistency — standardised footer hints, dialog spacing, and arrow-key styling. The dialog frame now appears immediately during loading instead of popping in later.
Ad

Other Noteworthy Fixes

  • Bash output colors at wrong positions in terminal and markdown code blocks
  • ReasonML diff corrupted undefined text artifacts at word-diff boundaries
  • Worktree exit dialog pointed to wrong directory after removal
  • @ file picker missed files in small non-git directories and directories with >100 entries
  • Failed tool calls not click-to-expand in fullscreen when output was truncated
  • Backspace/Ctrl+Backspace swapped after Ctrl+G external editor on terminals with persistent extended-key modes
  • /usage weekly reset showed time-of-day instead of calendar date
  • Welcome banner ellipsis caused column overflow on CJK terminals
  • /insights crash on tool calls with malformed input fields
  • Renderer crash when tool collapsibility classification changes mid-session
  • skills entry in plugin.json hiding default skills/ directory; listing a file path now shows an error instead of failing silently
  • IDE shell-integration lock files ignoring CLAUDE_CONFIG_DIR
  • Trailing whitespace in copied terminal output during streaming
  • Plugin uninstall/enable/disable not matching slugs case-insensitively
  • Tool error truncation marker showed negative count for surrogate-pair strings
  • Env vars from CLAUDE_ENV_FILE SessionStart hooks went stale after /resume or /clear
  • /branch saved multi-line session title when pasted with line breaks
  • Stray leading space on second line of wrapped text at column boundary
  • Esc not dismissing dialogs in /install-github-app, /desktop, /resume, /web-setup
  • /doctor MCP schema errors missing field name and source file path
  • Bash permission prompts showed parser diagnostic instead of user-readable explanation
  • Plugin slash commands with spaces not resolving to namespaced form
  • AskUserQuestion discarding multi-select answers when supplied as array
  • /clear <name> not labeling cleared session for /resume
  • CronList output missing qualifiers and scheduled prompt
  • 'Jump to bottom' overlay left color artifacts on CJK characters in fullscreen
  • Wide markdown tables left stale bordered render in terminal scrollback while streaming
  • Pasted text silently dropped when long prompt with pasted-text placeholder was auto-truncated
  • /release-notes stuck on old version

For details and download, see the GitHub release.

📖 Read the full source: GitHub Claude-Code

Ad

👀 See Also

Developer's experience with Claude AI: From thinking partner to cognitive outsourcing
News

Developer's experience with Claude AI: From thinking partner to cognitive outsourcing

A developer shares an 8-month experience using Claude AI daily, noting a shift from using it to refine existing thinking to outsourcing initial thinking entirely. The post describes two distinct cognitive approaches: AI as a thinking partner versus AI as a first-pass generator.

OpenClawRadar
Study Shows Claude Opus Agent Failures Were Architectural, Not Alignment Issues
News

Study Shows Claude Opus Agent Failures Were Architectural, Not Alignment Issues

A study placed Claude Opus and Kimi K2.5 in a live environment with email, shell access, and persistent storage. The models demonstrated correct values but experienced serious failures due to missing architectural safeguards like stakeholder models and execution boundaries.

OpenClawRadar
OpenClaw v2026.3.12 dashboard redesign consolidates interface elements
News

OpenClaw v2026.3.12 dashboard redesign consolidates interface elements

OpenClaw v2026.3.12 features a complete dashboard redesign that consolidates modular views for chat, config, agents, and sessions, along with command palette, mobile bottom tabs, slash commands, search, export, and pinned messages into a single interface.

OpenClawRadar
AI Agents That Don't Slash Maintenance Costs Will Sink Your Team
News

AI Agents That Don't Slash Maintenance Costs Will Sink Your Team

James Shore argues that doubling AI coding speed without halving maintenance costs leads to net productivity loss within months. Model shows 2x code output with 2x maintenance cost per line yields productivity worse than starting point after ~5 months.

OpenClawRadar