Version 1.0.77

What: Added support for persistent WebSocket connections as an alternative to stdin/stdout communication How to use:

# Connect to Claude's SDK server with WebSocket
export CLAUDE_CODE_SESSION_ACCESS_TOKEN="your-token-here"
claude --sdk-url=wss://api.claude.ai/sdk/v1/connect \
       --input-format=stream-json \
       --output-format=stream-json

Details:

• Automatic reconnection with exponential backoff (up to 5 attempts)
• Message buffering during disconnections (1000 message capacity)
• Request replay on reconnection to prevent message loss
• Support for both ws:// and wss:// protocols with Bearer token authentication
• Enables real-time bidirectional communication for streaming responses

What: New UI components that explain exactly why Claude needs permission for specific actions How to use: Permission explanations now appear automatically when confirmation is required:

Permission rule Edit(src/**) requires confirmation for this edit.
/permissions to update rules

Hook security-check requires confirmation for this command:
Accessing sensitive configuration files
/hooks to update

Details:

• Shows the specific permission rule or hook that triggered the request
• Provides custom reasoning from hooks when available
• Includes actionable guidance on how to modify permission settings
• Consistent display across all tool types (Edit, Command, WebFetch, etc.)

What: Sophisticated detection and escaping of dangerous shell constructs like heredocs and multiline strings How to use: This protection is automatic when Claude executes shell commands. It prevents injection attacks in complex scenarios:

# Heredocs are now safely handled
cat << EOF
$(potentially_dangerous_command)
EOF

# Multiline strings are properly escaped
echo 'line1
line2'

Details:

• Detects heredoc patterns that could bypass normal escaping
• Identifies multiline strings that might hide malicious commands
• Applies context-aware escaping strategies
• Automatically adds stdin redirection (< /dev/null) to prevent interactive command exploitation
• Protects against arithmetic expansion and nested command substitution attacks

What: Intelligent conversation summarization that respects model token limits How to use: Conversation summaries are generated automatically when saving conversations. The system will indicate if summarizing partial conversations:

Summarizing last 5 of 10 messages (~2500 tokens)

Details:

• Dynamically calculates available tokens based on model context window
• Prioritizes recent messages when full conversation exceeds token budget
• Transparently indicates when summaries are based on partial context
• Uses efficient Haiku model with prompt caching for fast summarization
• Generates concise 5-10 word titles for easy conversation browsing

What changed: Spinner tips are now managed server-side instead of being fetched by the client Impact: Eliminates network latency for tip loading, provides more contextual and personalized tips with sophisticated rotation logic to prevent repetition

What changed: Removed separate trust dialogs for hooks and bash commands in favor of unified permission system Impact: Cleaner configuration with single trust mechanism, reducing complexity while maintaining security

What changed: WebSocket transport includes automatic reconnection and message replay capabilities Impact: More reliable real-time communication that can survive temporary network disruptions without losing messages or state

• Issue: Users didn't understand why specific actions required permission
• Cause: Permission system only showed generic confirmation prompts
• Resolution: Added detailed reasoning display showing exact rules/hooks and custom explanations

• Issue: Complex shell constructs like heredocs could bypass escaping
• Cause: Standard shell-quote libraries don't handle all edge cases
• Resolution: Implemented custom pattern detection and multi-layer escaping strategy

• Issue: Long conversations could exceed model token limits during summarization
• Cause: No token budget management when generating summaries
• Resolution: Added token-aware truncation that prioritizes recent messages within available budget