Claude docs changes for May 27th, 2026 [diff]

Executive Summary

• New security-guidance plugin documentation: automatically reviews Claude's code changes for vulnerabilities (injection, unsafe deserialization, DOM injection, etc.) and fixes them in the same session, at three layers: per-edit pattern match, end-of-turn diff review, and commit/push review
• New comprehensive guide for monorepos and large codebases covering per-directory CLAUDE.md files, sparse worktrees, code intelligence plugins, and centralized convention management
• Version 2.1.152 adds a MessageDisplay hook for transforming assistant message text, SessionStart enhancements (session title setting, skill reloading), and disallowed-tools frontmatter for skills/commands
• Version 2.1.152: auto mode no longer requires opt-in consent, and /code-review --fix now applies findings directly to the working tree
• Corrected documentation across Bedrock, Vertex AI, and Microsoft Foundry: only /logout is unavailable when using cloud provider credentials (not /login)

New Claude Code versions

2.1.152

New features

• /code-review --fix now applies review findings to the working tree after the review; /simplify now invokes /code-review --fix
• Skills and slash commands can now set disallowed-tools in frontmatter to remove tools from the model while the skill or command is active
• Added /reload-skills command to re-scan skill directories without restarting the session
• SessionStart hooks can now return reloadSkills: true to re-scan skill directories, making skills installed by the hook available in the same session
• SessionStart hooks can now set the session title via hookSpecificOutput.sessionTitle on startup and resume
• Added a MessageDisplay hook event that lets hooks transform or hide assistant message text as it is displayed
• Added pluginSuggestionMarketplaces managed setting: admins can allowlist org marketplaces whose plugins may be suggested via context-aware tips
• claude plugin marketplace remove now accepts --scope user|project|local for symmetry with other marketplace commands
• Auto mode no longer requires opt-in consent
• Vim mode: / in NORMAL mode now opens reverse history search (like Ctrl+R), matching bash/zsh vi-mode

Existing feature improvements

• Claude Code now switches to the configured --fallback-model for the rest of the session when the primary model is not found, instead of failing every request
• The /usage breakdown now includes large session files; files are scanned with a streaming read so memory usage stays flat
• Thinking summaries in the collapsed group now stay readable for at least 3 seconds, render as markdown, and cap at 10 lines (Ctrl+O shows full thinking)
• In fullscreen mode, the "Thinking for Ns" indicator now counts up live while the model is thinking and keeps its value if you interrupt mid-thought
• Simplified the Workflow tool's inline progress display — live agent counts now show only in the persistent workflow status row
• The post-response timer now shows "Waiting for N background agents/workflows to finish" when backgrounded work is still running, and reports cumulative time once results are processed
• Added the session entrypoint as an OpenTelemetry metric attribute (app.entrypoint, opt-in via OTEL_METRICS_INCLUDE_ENTRYPOINT=true)

Major bug fixes

• Fixed terminal styling degrading in very long sessions by recycling the renderer's style pool
• Fixed sessions getting stuck after a model or login switch left stale thinking-block signatures in history
• Fixed cache_creation_input_tokens reporting as 0 in transcript and result usage when the API reports cache writes only via the nested cache_creation breakdown
• Fixed remote MCP servers failing to connect in Claude Code Remote sessions when the egress proxy is enabled
• Fixed plugin MCP servers with the same command but different environment variables being incorrectly deduplicated
• Fixed plugins that track a git branch silently no longer receiving updates after the plugin registry was rebuilt

Claude Code changes

New Documents

Large codebases [Source]

New guide: "Set up Claude Code in a monorepo or large codebase." Covers layering per-directory CLAUDE.md files, excluding irrelevant CLAUDE.md files with claudeMdExcludes, blocking reads of generated/vendored code with Read deny rules, using code intelligence plugins to replace file scans, configuring worktree.sparsePaths for lightweight worktrees, granting cross-package access with additionalDirectories, adding per-directory skills, and centralizing conventions via plugins when directory-level layering stops scaling.

Security guidance [Source]

New guide for the security-guidance plugin, which reviews Claude's own code changes for common vulnerabilities and fixes them in the same session. Describes three review layers: per-edit pattern match (no model call), end-of-turn diff review by a separate Claude instance, and an agentic commit/push review with surrounding code context. Covers custom rules via .claude/claude-security-guidance.md and .claude/security-patterns.yaml, usage costs, how to disable individual layers, and how the plugin fits alongside /security-review and PR-time code review.

Changed documents

Admin setup [Source]

• Added link to the new monorepo/large codebase guide for organizations deploying into a monorepo. [line 123] [Source]

Agent SDK - Overview [Source]

• Commands (formerly "Slash commands") are now described as the legacy format; the description explicitly says to use skills for new custom commands. [line 321] [Source]
• Plugins description updated to "Extend with skills, agents, hooks, and MCP servers" (removed "custom commands"). [line 323] [Source]

Agent SDK - Plugins [Source]

• The SDK init message now includes a skills field listing available plugin skills (separate from slash_commands); the code example is updated to show both. [line 143] [Source]
• Troubleshooting now says to verify that a skill appears in the skills list (not slash_commands) when debugging plugin skill invocation. [line 225] [Source]

Amazon Bedrock [Source]

• Corrected: when using Bedrock, only the /logout command is unavailable — /login is still available. [line 158] [Source]

Best practices [Source]

• Added guidance to have Claude show evidence of success (test output, command result, screenshot) rather than asserting it — reviewing evidence is faster than re-running verification. [line 37] [Source]
• Added tip that the most useful specs are self-contained: they name files and interfaces involved, state what is out of scope, and end with an end-to-end verification step. [line 338] [Source]
• Corrected non-interactive streaming example: --output-format stream-json now requires --verbose. [line 415] [Source]
• New "Add an adversarial review step" section: guidance on using a subagent in fresh context to review diffs against the plan before treating a task as done, including how to use /code-review and how to avoid over-engineering from reviewer findings. [line 499] [Source]

CLI reference [Source]

• Updated usage examples for --include-hook-events, --include-partial-messages, and --replay-user-messages to include --verbose, which is now required alongside --output-format stream-json. [lines 78-79, line 99]

Common workflows [Source]

• Added reference to the new monorepo/large codebase guide in the "Understand new codebases" section. [line 25] [Source]

Discover plugins [Source]

• Added new "Automatic security review" section listing the security-guidance plugin, which reviews each change Claude makes for common vulnerabilities and instructs Claude to fix them in the same session. [line 90] [Source]

Google Vertex AI [Source]

• Corrected: when using Vertex AI, only the /logout command is unavailable — /login is still available. [line 130] [Source]

Hooks guide [Source]

• Added reference to the security-guidance plugin as a production example of hooks that run a separate model review and feed findings back into the session. [line 93] [Source]

MCP [Source]

• Clarified scope precedence behavior: when the same server is defined in multiple scopes, the entire server entry from the highest-precedence source is used — fields are not merged across scopes. [line 304] [Source]

Memory [Source]

• Added note that to block an action regardless of what Claude decides, use a PreToolUse hook rather than CLAUDE.md instructions. [line 23] [Source]
• Corrected the maximum import recursion depth for @path imports in CLAUDE.md from five to four hops. [line 87] [Source]
• Added reference to the new monorepo/large codebase guide for CLAUDE.md layout and per-directory rules. [line 137] [Source]

Microsoft Foundry [Source]

• Corrected: when using Microsoft Foundry, only the /logout command is unavailable — /login is still available. [line 59] [Source]

Security [Source]

• Added link to the security guidance plugin in the related resources section. [line 132] [Source]

API changes

No significant API documentation changes today.