Version 2.1.38

Control how much effort Claude puts into responses during a session.

claude --effort medium
claude --effort high

• Valid levels: low, medium, high
• The max level exists but is restricted:
• Not available in interactive mode
• Not available for Claude.ai subscribers
• Use high for complex analysis, low for quick tasks

CLI flag definition (search for "--effort <level>")

Claude can now automatically learn and remember patterns across sessions, writing insights to a persistent memory directory.

Activation: Auto-memory has a three-layer activation cascade (checked in order):

1. Environment variable (highest priority): CLAUDE_CODE_DISABLE_AUTO_MEMORY=true forces off, =false forces on
2. User setting: autoMemoryEnabled in ~/.claude/settings.json (per-project, overrides the feature flag)
3. GrowthBook feature flag tengu_oboe (default fallback): server-side rollout control, defaults to false

If the flag is off for your account and you haven't set the user setting, auto-memory is disabled by default. Remote sessions (CLAUDE_CODE_REMOTE=true) also disable it unless CLAUDE_CODE_REMOTE_MEMORY_DIR is set.

How to enable: The toggle is in the /memory panel (not /settings). Type /memory to see "Auto-memory (research preview): off" at the top, then confirm to toggle it on. This writes autoMemoryEnabled: true to settings, overriding the feature flag. After enabling, "Open auto-memory folder" and per-agent memory entries also appear in the /memory panel.

Note: The feature flag tengu_oboe existed in v2.1.37 but with no user-facing toggle — auto-memory was purely server-gated. v2.1.38 added the autoMemoryEnabled setting, the /memory panel toggle, and the bidirectional env var override, making it opt-in for everyone.

• Per-project setting: autoMemoryEnabled controls whether Claude reads/writes to auto-memory
• Memory guidance updated from "write down key learnings as you complete tasks" to "save patterns worth preserving across sessions"
• New guidelines for what to save:
• Stable patterns and conventions confirmed across multiple interactions
• User preferences for workflow, tools, and communication style
• Key architectural decisions and important file paths
• Solutions to recurring problems and debugging insights
• What NOT to save:
• Session-specific context (current task details, in-progress work)
• Speculative conclusions from reading a single file
• Information that duplicates CLAUDE.md instructions
• Supports explicit user requests: "always use bun", "never auto-commit", or "forget X"

Activation logic (search for tengu_oboe), settings schema (search for "Enable auto-memory for this project"), /memory panel toggle (search for "Auto-memory (research preview):")

Switch between multiple Chrome browsers when using Claude's browser automation features.

Usage: Invoke the switch_browser tool when you want to connect to a different Chrome instance.

• Broadcasts a connection request to all Chrome browsers with the Claude extension installed
• User clicks 'Connect' in the desired browser to switch
• Only available with bridge connections (shows error otherwise)
• Useful for multi-browser development/testing workflows

Tool definition (search for "Switch which Chrome browser is used for browser automation")

Skills can now activate automatically when you work with files matching specific patterns.

• Skills are stored with file pattern matchers
• When a matching file is touched, the skill activates automatically
• Log message shows: [skills] Activated conditional skill '<name>' (matched path: <path>)
• Tracked via telemetry for optimization

Skill activation system (search for "Activated conditional skill")

Plan mode now persists file snapshots to the conversation transcript, enabling recovery if files are deleted.

• File snapshots are stored as system messages with subtype file_snapshot
• Recovery sequence when resuming a conversation:

1. Try to read plan file from disk
2. If missing, recover from file snapshot in message history
3. If no snapshot, attempt recovery from serialized plan content

• Log messages indicate recovery source: "Plan recovered from file snapshot"
• Protects against accidental file deletion during long planning sessions

Recovery system (search for "Plan recovered from file snapshot")

Configure directories to symlink when using git worktrees, preventing disk bloat from duplicated dependencies.

Usage: Add to your settings:

{
  "worktree": {
    "symlinkDirectories": ["node_modules", ".cache", ".bin"]
  }
}

• No directories are symlinked by default — must be explicitly configured
• Common candidates: node_modules, .cache, .bin
• Prevents duplicating large directories across worktrees
• Part of the --worktree flag configuration

Settings schema (search for "Directories to symlink from main repository")

MCP elicitation now supports natural language date/time input with intelligent parsing.

• Accepts natural language: "tomorrow", "next Monday", "jan 1st 2025", "in 2 hours"
• Converts to ISO 8601 format automatically
• Prefers future dates over past dates when ambiguous
• For dates without times, omits the time component
• Supports multiple formats: uri, email, date, date-time
• Shows current context (date, time, timezone, day of week) to help parsing
• Includes minItems/maxItems validation for selection elicitations
• Gracefully handles invalid input with clear error messages

Date parser prompt (search for "You are a date/time parser")

Claude now displays a helpful message when auto-compact is enabled: "Auto-compact is enabled. When the context window is nearly full, older messages will be automatically summarized so you can continue working seamlessly. There is no need to stop or rush — you have unlimited context through automatic compaction."

Notification message (search for "Auto-compact is enabled")

Added structured guidance for searching past context:

1. Search topic files in memory directory first (fast)
2. Session transcript logs as last resort (large files, slow)

• New tip: "Use narrow search terms (error messages, file paths, function names) rather than broad keywords"

Memory search section (search for "Searching past context")

When using thinking.type=enabled, Claude now warns that this is deprecated and recommends thinking.type=adaptive for better performance, with a link to documentation.

Deprecation warning (search for "thinking.type=enabled' is deprecated")

Changed reconnection handling from a fixed retry count to a time budget approach, providing more predictable behavior during network instability.

Transport logging (search for "Reconnection time budget exhausted")

The output_format parameter is now deprecated in favor of output_config.format. Using both together produces a warning.

Deprecation check (search for "output_format is deprecated")

Clear error message when trying to use Opus 4.6 with 1M extended context on unsupported plans, with link to documentation.

Error message (search for "Opus 4.6 with extended context (1M) is not available")

New noOutputExpected boolean field in command schemas indicates when a command is expected to produce no output on success, improving semantic interpretation of exit codes.

Schema definition (search for "Whether the command is expected to produce no output")

Improved handling of token refresh failures with invalid_grant errors, including automatic credential invalidation and token clearing for smoother re-authentication.

Token handling (search for "Token refresh failed with invalid_grant")

• Fixed handling of orphaned permissions when updatedInput is undefined, now falls back to original tool input (search for "Orphaned permission for")
• Added security check for commands containing single-quoted backslash patterns that could bypass security checks (search for "single-quoted backslash pattern")
• Improved sync agent error recovery with automatic retry mechanism (search for "Sync agent recovering from error")
• Fixed compaction error handling when response is not text as expected (search for "Expected text response for compaction")
• Added validation for elicitation responses with proper "This field is required" messaging

• Updated Azure Identity SDK integration with additional credential types (AzurePipelinesCredential, VisualStudioCodeCredential, etc.)
• Added bridge staging URL for development/testing environments
• Improved Symbol.dispose/Symbol.asyncDispose handling for resource cleanup
• Enhanced session title update logging with success/failure tracking