Claude docs changes for April 11th, 2026 [diff]

Executive Summary

• New documentation for Claude in Amazon Bedrock (research preview): a new Messages API endpoint at /anthropic/v1/messages runs on AWS-managed infrastructure with zero operator access, currently available in us-east-1
• Claude Code 2.1.101 now trusts the OS CA certificate store by default, eliminating extra configuration for enterprise TLS-inspection proxies (CrowdStrike, Zscaler, etc.)
• /loop command significantly expanded: self-paced interval mode where Claude chooses the delay, a built-in autonomous maintenance prompt, and loop.md customization for project-specific defaults
• Security fix in 2.1.101: command injection vulnerability in the POSIX which fallback used by LSP binary detection
• Structured outputs extended to Ruby and PHP SDKs; TypeScript gains jsonSchemaOutputFormat() alongside zodOutputFormat()

New Claude Code versions

2.1.101

New features

• Added /team-onboarding command to generate a teammate ramp-up guide from your local Claude Code usage
• Added OS CA certificate store trust by default, so enterprise TLS proxies work without extra setup; set CLAUDE_CODE_CERT_STORE=bundled to use only the bundled Mozilla CA set
• /ultraplan and other remote-session features now auto-create a default cloud environment instead of requiring web setup first

Existing feature improvements

• Improved brief mode to retry once when Claude responds with plain text instead of a structured message
• Improved rate-limit retry messages to show which limit was hit and when it resets instead of an opaque seconds countdown
• Improved claude -p --resume <name> to accept session titles set via /rename or --name
• Improved settings resilience: an unrecognized hook event name in settings.json no longer causes the entire file to be ignored
• Improved plugin hooks from plugins force-enabled by managed settings to run when allowManagedHooksOnly is set
• Improved beta tracing to honor OTEL_LOG_USER_PROMPTS, OTEL_LOG_TOOL_DETAILS, and OTEL_LOG_TOOL_CONTENT; sensitive span attributes are no longer emitted unless opted in
• Improved SDK query() to clean up subprocess and temp files when consumers break from for await or use await using

Major bug fixes

• Fixed a command injection vulnerability in the POSIX which fallback used by LSP binary detection
• Fixed a memory leak where long sessions retained dozens of historical copies of the message list in the virtual scroller
• Fixed --resume/--continue losing conversation context on large sessions when the loader anchored on a dead-end branch instead of the live conversation
• Fixed a hardcoded 5-minute request timeout that aborted slow backends (local LLMs, extended thinking, slow gateways) regardless of API_TIMEOUT_MS
• Fixed permissions.deny rules not overriding a PreToolUse hook's permissionDecision: "ask" — previously the hook could downgrade a deny into a prompt
• Fixed Bedrock SigV4 authentication failing with 403 when ANTHROPIC_AUTH_TOKEN, apiKeyHelper, or ANTHROPIC_CUSTOM_HEADERS set an Authorization header
• Fixed subagents not inheriting MCP tools from dynamically-injected servers
• Fixed sub-agents running in isolated worktrees being denied Read/Edit access to files inside their own worktree
• Fixed Grep tool ENOENT when the embedded ripgrep binary path becomes stale (VS Code extension auto-update, macOS App Translocation); now falls back to system rg and self-heals mid-session
• Fixed custom keybindings (~/.claude/keybindings.json) not loading on Bedrock, Vertex, and other third-party providers

Claude Code changes

Changed documents

Amazon Bedrock [Source]

• ANTHROPIC_SMALL_FAST_MODEL_AWS_REGION now also applies to the Bedrock Mantle endpoint, not just standard Bedrock. [line 130] [Source]
• Added ANTHROPIC_SMALL_FAST_MODEL_AWS_REGION to the Mantle-specific environment variables table. [line 333] [Source]

Commands [Source]

• /loop updated: both the interval and the prompt are now optional. Omitting the interval lets Claude self-pace between iterations; omitting the prompt runs the built-in maintenance check or .claude/loop.md if present. [line 46] [Source]

Computer use [Source]

• New section "Screenshots are downscaled automatically": Claude Code downscales every screenshot before sending it to the model, so Retina and high-DPI displays do not need resolution changes. For example, a 16-inch MacBook Pro at native Retina resolution captures at 3456×2234 and downscales to roughly 1372×887. [line 99] [Source]

Costs [Source]

• Updated cost benchmarks based on enterprise deployment data: average is now ~$13/developer/active day and $150–250/developer/month (previously $6/day and $100–200/month). 90th-percentile cap updated to $30/active day. [line 3] [Source]

Environment variables [Source]

• ANTHROPIC_SMALL_FAST_MODEL_AWS_REGION description updated to mention it also applies to Bedrock Mantle. [line 34] [Source]
• CLAUDE_BASH_MAINTAIN_PROJECT_WORKING_DIR description updated to include PowerShell commands (not just Bash). [line 47] [Source]
• New CLAUDE_CODE_CERT_STORE variable added: comma-separated list of CA certificate sources (bundled, system). Default is bundled,system. [line 51] [Source]

Hooks guide [Source]

• Added a full JSON example showing how to correctly merge multiple hook event types (e.g., PostToolUse and Notification) into the same hooks object, clarifying the correct nesting structure. [line 35] [Source]

Hooks [Source]

• Clarified that allowManagedHooksOnly exempts hooks from plugins that are force-enabled via enabledPlugins in managed settings, allowing organizations to distribute vetted hooks through a marketplace while blocking everything else. [line 161] [Source]

MCP [Source]

• MCP scope hierarchy and precedence now documented as an explicit ordered list: Local (1) > Project (2) > User (3) > Plugin-provided servers (4) > claude.ai connectors (5). Also clarified that plugins and connectors deduplicate by endpoint/URL, not by name. [line 1163] [Source]

Monitoring usage [Source]

• PowerShell subprocesses now also inherit the TRACEPARENT environment variable for distributed tracing (previously only Bash subprocesses were mentioned). [line 107] [Source]

Network configuration [Source]

• New "CA certificate store" section: documents that Claude Code now trusts both the bundled Mozilla CA set and the OS certificate store by default, making enterprise TLS-inspection proxies work out of the box. Covers the CLAUDE_CODE_CERT_STORE env var (bundled, system, or both), notes that system CA integration requires the native binary (not the Node.js runtime), and shows how to restrict to one source. [line 42] [Source]

Permissions [Source]

• Clarified that the :* suffix in Bash permission rules (e.g., Bash(ls:*)) is not deprecated — it is an equivalent way to write a trailing wildcard, behaving identically to Bash(ls *). [line 86] [Source]
• Updated allowManagedHooksOnly description to clarify that hooks from plugins force-enabled in managed settings enabledPlugins are exempt from the block. [line 237] [Source]

Run prompts on a schedule [Source]

• Major /loop documentation overhaul. Added a summary table of all three invocation modes: interval+prompt (fixed schedule), prompt-only (Claude self-paces), and bare /loop (built-in maintenance prompt). [line 29] [Source]
• New section "Let Claude choose the interval": when the interval is omitted, Claude dynamically picks a delay (1 min–1 hour) based on observed activity. May use the Monitor tool to stream events instead of polling. [line 57] [Source]
• New section "Run the built-in maintenance prompt": bare /loop runs an autonomous maintenance loop — continues unfinished work, tends to the PR (review comments, CI failures, conflicts), and runs cleanup passes when nothing else is pending. Irreversible actions only proceed if authorized in the transcript. [line 72] [Source]
• New section "Customize the default prompt with loop.md": .claude/loop.md (project-level) or ~/.claude/loop.md (user-level) replaces the built-in maintenance prompt. Edits take effect on the next iteration; content over 25,000 bytes is truncated. [line 90] [Source]

Server-managed settings [Source]

• Removed "(public beta)" from the title and beta-related language throughout the page; server-managed settings are now generally available. [line 1] [Source]

Settings [Source]

• Updated allowManagedHooksOnly description and behavior documentation to clarify that hooks from plugins force-enabled in enabledPlugins are loaded; trust is granted by full plugin@marketplace ID. [lines 143, 398]

Sub-agents [Source]

• Added clarification that a subagent starts in the main conversation's current working directory; cd commands inside a subagent do not persist between Bash/PowerShell calls and do not affect the main conversation's working directory. Use isolation: worktree to give the subagent an isolated copy. [line 207] [Source]

Tools reference [Source]

• Expanded Bash tool working directory behavior: cd in the main session carries over to later Bash commands as long as it stays within the project directory or an added directory; moving outside causes a reset with a Shell cwd was reset to <dir> message appended. CLAUDE_BASH_MAINTAIN_PROJECT_WORKING_DIR=1 disables carry-over entirely. Subagent sessions never carry over working directory changes. [line 48] [Source]
• Added note that the same main-session working-directory reset behavior and CLAUDE_BASH_MAINTAIN_PROJECT_WORKING_DIR apply to PowerShell commands as well. [line 109] [Source]

Troubleshooting [Source]

• Added CRYPT_E_NO_REVOCATION_CHECK (0x80092012) as an additional Windows TLS error code that indicates the certificate revocation lookup is blocked by a corporate firewall, alongside the existing CRYPT_E_REVOCATION_OFFLINE entry. [line 294] [Source]

API changes

New Documents

Claude in Amazon Bedrock (research preview) [Source]

New guide for the Claude in Amazon Bedrock research preview — a new Messages API endpoint at /anthropic/v1/messages running on AWS-managed infrastructure with zero operator access. Covers three authentication methods (Bedrock service role, IAM assumed roles, bearer tokens), SDK installation, supported models (Claude Mythos Preview and Claude Haiku 4.5 with anthropic. prefix), feature availability (extended thinking, prompt caching, tool use, citations, structured outputs supported; Anthropic-defined tools, Managed Agents, and Batches API not supported), quotas (2M input TPM default, up to 4M without additional approval), data retention, CloudWatch/CloudTrail observability, and support contacts. Currently available in us-east-1 by invitation only.

Changed documents

Claude on Amazon Bedrock [Source]

• Intro updated to clarify this page covers the current Bedrock integration (InvokeModel/Converse APIs), and now links to the new research preview page for the Messages API endpoint. [line 3] [Source]

Code execution tool [Source]

• Replaced prose model compatibility description with a detailed table listing every supported model and which tool versions (code_execution_20250825, code_execution_20260120) each one supports. Clarifies that code_execution_20260120 (REPL persistence + programmatic tool calling) is limited to Opus 4.5+/Sonnet 4.5+ and newer. [line 17] [Source]

Programmatic tool calling [Source]

• Replaced the generic compatibility reference with an explicit table of the four models that support code_execution_20260120: Claude Opus 4.6, Claude Sonnet 4.6, Claude Opus 4.5, and Claude Sonnet 4.5. [line 17] [Source]

Structured outputs [Source]

• Clarified that only the Python SDK's client.messages.parse() accepts output_format as a convenience parameter; other SDKs require output_config directly. [line 113] [Source]
• TypeScript now supports jsonSchemaOutputFormat() in addition to zodOutputFormat() for typed JSON Schema literals. [line 119] [Source]
• PHP SDK now supports structured outputs via classes implementing StructuredOutputModel with outputConfig: ['format' => MyClass::class]. [line 121] [Source]
• Added a complete CLI example for structured outputs using a YAML heredoc body and --transform with GJSON. [line 196] [Source]
• Automatic schema transformation (removing unsupported constraints, updating descriptions) now documented for Ruby and PHP SDKs in addition to Python and TypeScript. [line 224] [Source]