Source Intelligence
Follow Discord

DisclaimerUnofficial, and not affiliated with Anthropic. Nearly all of this is read straight out of what ships: npm bundles, captured prompts, published docs. Anthropic's own notes go in verbatim, marked as theirs. The rest is my reading, and every entry carries the strings behind it. If one looks wrong, vote it down and say why.

← All releases Home ←olderv2.1.37 v2.1.39newer→

Claude Code v2.1.38

24 entries read diff v2.1.37 → v2.1.38 Markdown

This release introduces session effort levels (low/medium/high) for controlling Claude's response intensity, auto-memory as a research preview for persistent learning across sessions, and browser switching for multi-Chrome automation setups. The update also adds file snapshot persistence for plan recovery, conditional skills that activate based on file patterns, improved elicitation with natural language date/time parsing, and new security checks for command validation.

Find
Pick an entry · j / k steps through
15 entries

Changesopen

Effort Level Control#

What

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

Usage
claude --effort medium
claude --effort high
Details
  • 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
Evidence

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

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Auto-Memory (Research Preview)#

What

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.

Details
  • 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"
Evidence

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):")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Related

Other releases about the same thing. Found by shared names or similar wording; neither means one caused the other.

Browser Switching for Chrome Automation#

What

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.

Details
  • 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
Evidence

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

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Conditional Skills#

What

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

Details
  • 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
Evidence

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

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Related

Other releases about the same thing. Found by shared names or similar wording; neither means one caused the other.

File Snapshot Persistence for Plan Recovery#

What

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

Details
  • 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
Evidence

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

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Enhanced Elicitation with Natural Language Date/Time Parsing#

What

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

Details
  • 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
Evidence

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

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Auto-Compact Notification#

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."

Evidence

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

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Related

Other releases about the same thing. Found by shared names or similar wording; neither means one caused the other.

Memory Search Improvements#

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"
Evidence

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

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Thinking Mode Deprecation Warning#

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.

Evidence

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

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

WebSocket Reconnection Improvements#

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

Evidence

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

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Output Format Deprecation#

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

Evidence

Deprecation check (search for "output_format is deprecated")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Related

Other releases about the same thing. Found by shared names or similar wording; neither means one caused the other.

Opus 4.6 Extended Context Messaging#

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

Evidence

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

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

MCP Bash Tool Enhancement#

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.

Evidence

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

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Credential Refresh Improvements#

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

Evidence

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

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Related

Other releases about the same thing. Found by shared names or similar wording; neither means one caused the other.

1 entry

Bug Fixesopen

#

  • 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
1 entry

Internalopen

#

  • 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
Verbatim
Official · Anthropic

Anthropic’s official release notes

Published verbatim by Anthropic for v2.1.38. Text is unmodified from the upstream changelog. Everything else on this page came out of the bundle instead, which is why the two lists don't match.

  • Fixed VS Code terminal scroll-to-top regression introduced in 2.1.37
  • Fixed Tab key queueing slash commands instead of autocompleting
  • Fixed bash permission matching for commands using environment variable wrappers
  • Fixed text between tool uses disappearing when not using streaming
  • Fixed duplicate sessions when resuming in VS Code extension
  • Improved heredoc delimiter parsing to prevent command smuggling
  • Blocked writes to .claude/skills directory in sandbox mode
Read it on GitHub →
System prompt

The system prompt was not captured for this release, so this page cannot say whether it moved.

Which releases were captured →