Claude Code v2.1.184 to v2.1.186 Major Updates

This is Ishikawa from the Cloud Business Division. I'll summarize the Claude Code updates from v2.1.184 to v2.1.186 (2026-06-20 to 2026-06-22). Note that v2.1.184 was not published to npm and is a missing number, so this article covers v2.1.185 and v2.1.186 that were actually published.

Update Summary

This covers 2 versions (v2.1.185 to v2.1.186, 2026-06-20 to 2026-06-22). In addition to new features such as claude mcp login / claude mcp logout for authenticating to MCP servers from the CLI, important fixes are included for streaming failures after sleep recovery and permission rules not being applied when generating sub-agents. There are also default behavior changes, such as Claude automatically responding to bash command output executed with !, so please also check the "Breaking Changes" section described later.

Notable Updates

• New Feature: Added claude mcp login / claude mcp logout to authenticate to MCP servers from the CLI without opening the interactive menu (v2.1.186)
• Bug Fix: Fixed an issue where streaming requests would fail after the machine wakes from sleep (v2.1.186)
• Security: Fixed an issue where Agent(type) deny rules and allowed type restrictions were not applied when generating named sub-agents (v2.1.186)

Target Versions and Period

New Features

• Added claude mcp login / claude mcp logout: You can now authenticate and log out of MCP servers from the CLI without opening the interactive /mcp menu. Supports stdin redirection via --no-browser, allowing authentication to be completed over SSH as well (v2.1.186).
• Added status filtering to /workflows: Pressing the f key in the agent detail view allows filtering by status (v2.1.186).
• Added AWS credential refresh option to /login: When awsAuthRefresh is configured, "Claude Platform on AWS - refresh credentials" is displayed in /login, allowing you to refresh AWS credentials (v2.1.186).
• Added teammateMode: "iterm2" setting: You can now specify iTerm2 as the teammate display destination. A warning is displayed if the it2 CLI is not found in auto mode (v2.1.186).
• Added "Skills" section to the Installed tab of /plugin: Makes it easier to check the skills provided by installed plugins (v2.1.186).

Security

Several fixes have been made around permissions and isolation.

• Fixed permission rules not being applied when generating named sub-agents: Agent(type) deny rules and Agent(x,y) allowed type restrictions were not being applied when generating named sub-agents. This ensures that configured permission controls now work as intended (v2.1.186).
• Fixed feature-gated tools passing through via --tools: Fixed an issue where --tools would pass feature-gated tools before flags were loaded during a cold initial startup (v2.1.186).
• Fixed Chrome tab group isolation not being applied: Fixed an issue where Chrome tab group isolation was not applied when the in-product permission gate was off in concurrently running CLI sessions (v2.1.186).

Improvements

• Changed hint display when stream stalls: The display when a response stalls has changed from "No response from API · Retrying in …" to "Waiting for API response · will retry in …", and the trigger timing has also been relaxed from 10 seconds to 20 seconds of no response (v2.1.185).
• Memory feature improvements: When the MEMORY.md index approaches its size limit, the agent is now prompted to compact it (v2.1.186).
• Flexible key notation for skill frontmatter: The display-name, default-enabled, fallback, and metadata.* keys are now accepted in kebab-case, snake_case, or camelCase notation (v2.1.186).
• Unified review engine for /review <pr>: /review <pr> now uses the same review engine as /code-review medium (v2.1.186).
• Improvements to claude mcp get / claude mcp remove: Now suggests the closest configured server name on typos and abbreviates long server lists (v2.1.186).
• Improved handling of invalid SKILL.md frontmatter: When YAML frontmatter is invalid, the skill body is now loaded with empty metadata instead of silently failing (v2.1.186).

Fixes (Major)

Excerpting fixes mainly related to stability and operability.

• Fixed streaming failures after sleep recovery: Fixed an issue where streaming requests would fail with "Content block not found" or JSON parse errors after the machine wakes from sleep (v2.1.186).
• Fixed Esc / Ctrl+C being unresponsive while background agents are running: Fixed an issue where Esc and Ctrl+C were unresponsive while background agents were still running after the main turn ended (v2.1.186).
• Fixed session costs not displaying for Enterprise and Team: Fixed an issue where session costs were not displayed for usage-based Enterprise and Team subscribers (v2.1.186).
• Fixed --effort inheritance for agent teams: Teammates spawned via the tmux/pane backend now inherit the leader's --effort level (v2.1.186).
• Fixed Workflow schema validation loop: Fixed an issue where agent({schema}) sub-agents in Workflows would loop infinitely instead of stopping after 5 attempts when schema validation repeatedly failed (v2.1.186).
• In addition, numerous minor bugs have been fixed, including sub-agent/agent panel display, background session screen rendering, /plugin scroll display, Markdown strikethrough rendering, permission prompt option number misalignment, and confusing "MCP server disconnected" notifications for deprecated tools.

Breaking Changes

Changes that alter default behavior are included. Depending on how you have been using these features, you may be affected, so please review the contents before updating.

Claude now automatically responds to bash command output from ! (v2.1.186)

Claude will now automatically respond to the output of bash commands executed with !.

The key point is not "whether the output is included in the context" but "whether Claude automatically responds." Previously, commands executed with ! and their output were already incorporated into the conversation context. However, Claude would not automatically respond to that output and would wait until you explicitly gave the next instruction (this is what the changelog refers to as context-only behavior). From v2.1.186 onwards, Claude will automatically respond to that output (with explanations, suggestions for next actions, etc.).

To revert to the previous context-only behavior, set the following in settings.json.

{
  "respondToBashCommands": false
}

Please be aware of the behavioral differences in the following cases.

• When running commands for confirmation purposes: In cases where you ran !git status or !ls just to "include the output in the context" and intended to give the next instruction yourself, Claude will now react to that output each time and return explanations or suggestions for next actions. There is a possibility of unintended response turns and increased token consumption.
• When streaming large amounts of output or logs: If you output long build logs or test results with !, Claude may automatically react and begin lengthy explanations.
• When output contains sensitive information: Even if you ran a command output containing environment variables or credentials for confirmation purposes, Claude may reference or process the content.
• Workflows that run multiple ! commands in sequence: In workflows where commands are stacked up, a response turn will be inserted after each command, changing the previous tempo and progression.

If you have been using ! exclusively for context ingestion, you can maintain the previous behavior with respondToBashCommands: false.

Changed permission prompt behavior for background sub-agents (v2.1.186)

When a background sub-agent requests permissions, instead of being automatically denied as before, a prompt is now displayed in the main session. The dialog indicates which agent is making the request, and pressing Esc denies only that tool. Please note that flows that previously ran unattended with "automatic denial" as a premise may now block waiting for a prompt.

CLAUDE_CODE_MAX_RETRIES capped at 15 (v2.1.186)

CLAUDE_CODE_MAX_RETRIES is now capped at 15. In environments where a value greater than 15 was set, only up to the cap will be applied. For unattended execution sessions, using CLAUDE_CODE_RETRY_WATCHDOG is recommended instead.

In Closing

In v2.1.185 to v2.1.186, in addition to new features such as MCP authentication from the CLI and workflow filtering, fixes related to reliability and security have been consolidated, including streaming failures after sleep recovery and permission rules not being applied to sub-agents. On the other hand, there are also changes to default behavior, such as automatic responses to ! bash commands and permission prompts for background sub-agents. If there are any changes that interest you, please update and check them out.

References

https://github.com/anthropics/claude-code/blob/main/CHANGELOG.md

https://code.claude.com/docs/en/changelog