Claude Code v2.0.52

What: Large tool outputs (>400KB) are now saved to disk instead of being truncated, allowing Claude to access the full data later.

How it works: When a tool produces output exceeding 400KB with ENABLE_TOOL_RESULT_SIZE_LIMIT enabled, the output is:

1. Saved to ~/.claude/projects/<project>/tool-results/<tool>-<timestamp>/result.{txt|json}
2. A 2KB preview is shown inline
3. Claude receives instructions on how to explore the full file

Example output:

<persisted-output>
Output too large (1.5 MB). Full output saved to: ~/.claude/projects/.../tool-results/Bash-1701234567890/result.txt

Preview (first 2.0 KB):
[preview content...]
...

You can explore this file using:
- Read tool to view portions of the file
- Grep tool to search for patterns
- Bash with jq to query JSON data
- Bash with head/tail for beginning/end
</persisted-output>

Details:

• Automatically detects JSON output and creates .json files with jq usage examples
• Tool result files are automatically granted read permission (no prompts needed)
• Falls back to truncation if file writing fails
• Evidence: $_2() at line 403698, if5() at line 403760, nf5() at line 403783

What: New CLAUDE_ENV_FILE environment variable allows loading custom session environment scripts.

How to use:

export CLAUDE_ENV_FILE=/path/to/my-env-setup.sh
claude

Details:

• Scripts from CLAUDE_ENV_FILE are loaded before hook files from session-env directory
• Multiple sources can be combined (env file + hooks)
• Independent error handling - failures in one source don't prevent loading others
• Useful for personal environment configuration that doesn't belong in hooks
• Evidence: kqB() at line 232402, environment read at line 232407

What: Pro and Max plan users can now see their extra usage balance directly in the /usage command output.

How to use:

/usage  # Now shows extra usage section for eligible users

Details:

• Shows three states:
• Not enabled: "Extra usage not enabled • /extra-usage to enable"
• Unlimited: Shows "Unlimited" with no progress bar
• Metered: Shows utilization bar with spending (e.g., "$5.23 / $30.00 spent")
• Displays monthly reset date
• Controlled by feature flag claude_code_usage_show_extra_usage
• Only visible to Pro and Max subscribers
• Evidence: PM3() at line 481986, RJ0 title at line 482045

What: New detection capability for when Claude Code is running inside a Bubblewrap sandbox on Linux.

How it works:

• Checks process.platform === "linux" AND process.env.CLAUDE_CODE_BUBBLEWRAP === "1"
• Infrastructure for future sandbox-aware behavior

Details:

• Currently unused in application code (preparation for future features)
• Part of environment detection module
• Evidence: OV6() at line 228083

What: Debug logging now uses intelligent buffering instead of direct filesystem writes.

Details:

• Accumulates log entries in memory (up to 100 entries)
• Flushes every 1000ms or when buffer is full
• Reduces filesystem I/O by 10-100x when debug logging is enabled
• Automatic immediate mode when using --debug flags (for real-time debugging)
• Graceful disposal ensures logs are flushed on exit
• Evidence: QC0() at line 2348, nw9() at line 2404

What: When startup profiling is enabled, reports are now saved to persistent files for later analysis.

How to use:

export CLAUDE_CODE_PROFILE_STARTUP=1
claude <command>
# Report saved to: ~/.claude/startup-perf/<timestamp>.txt

Details:

• Previously only logged to console (ephemeral)
• Now saves to ~/.claude/startup-perf/<timestamp>.txt
• Directory created automatically if missing
• Still logs to console in addition to file
• Evidence: WV9() at line 507095, cP3() at line 507106

What: Permission checks for session files (bash outputs, session memory, plan files, tool results) are now centralized.

Details:

• New FP3() helper function consolidates all special file permission logic
• Tool result files automatically granted read permission (new)
• Cleaner codebase with less duplication
• Evidence: FP3() at line 506192

What: Added support for parsing wildcard tool permission rules like ToolName:*.

Details:

• New Wt8() function parses wildcard patterns in permission rules
• New Mq() function parses tool names with optional rule content in format toolName(ruleContent)
• Enhanced permission system flexibility
• Evidence: Wt8() at line 190167, Mq() at line 502273

What: The ConEmu/Ghostty progress reporting system was refactored from an independent side-channel to an integrated render pipeline.

Details:

• Removed: O81 class (ConEmu progress reporter), T1A class (progress manager), LOA global instance
• Added: ink-progress DOM elements, sGB() progress extraction from DOM
• Progress now updates synchronously with UI renders
• Same ANSI escape sequences generated, but through render pipeline
• Evidence: Removed classes at lines 403283-403407 (v2.0.51), new functions at lines 218873, 220775

What: Added telemetry for A/B test experiment exposure tracking.

Details:

• Tracks which users are exposed to which experiment variations
• Events include: experiment ID, variation ID, device ID, user attributes
• Batched transmission via OpenTelemetry log pipeline
• Internal analytics only - no user-visible impact
• Evidence: rNB() at line 231011, tNB() at line 231925, aNB() at line 230979

What: When running in GitHub Actions, Claude Code now collects additional metadata.

Details:

• New fields collected: actor, actor ID, repository, repository ID, owner, owner ID
• Sent via github_actions_metadata telemetry field
• Only when GITHUB_ACTIONS=true environment variable is set
• Evidence: Metadata collection at lines 228565-228573, telemetry at line 231700

What: New CLAUDE_CODE_REMOTE_SESSION_ID environment variable for tracking remote sessions.

Details:

• Read from environment and included in telemetry
• Sent as HTTP header x-claude-remote-session-id in API requests
• Enables correlation of events across remote Claude Code sessions
• Evidence: Environment read at line 231577, telemetry at line 231676, header at line 301327

What: Refactored how hook progress and resolution counts are tracked during tool execution.

Details:

• New J29() function pre-computes hook tracking maps
• W29() and F29() provide efficient lookup of sibling tool IDs and pending hooks
• Improves performance during parallel tool execution
• Evidence: J29() at line 503835, W29() at line 503879, F29() at line 503889