Introduction
Since I’m using Claude Code as my main development tool, I’ve developed a habit of checking the release notes as soon as they come out. v2.1.32 was released on February 5th, and this update has quite a few substantial changes, so I’m going to summarize them.
It would be boring to simply end with “This feature has been added,” so I’m also including practical usage tips on how to actually use them. I hope it helps someone.
1. New Features
A. Claude Opus 4.6 Model Support
The latest Opus model has been added. You can select it with the --model option in Claude Code, and you can also set it as the default.
Usage Tip: Model Usage Strategy
You don’t need to use Opus for every task. If you divide and use them according to the situation, you can save costs and improve speed.
| Situation | Model | Reason |
|---|---|---|
| Architecture design, complex refactoring | Opus | Requires broad context understanding + precise judgment |
| Simple bug fixes, code formatting | Sonnet | Fast and cheap, good enough for this |
| File exploration, search-oriented tasks | Haiku | Fast when run as a sub-agent |
When designing with the Plan agent, use Opus, and when actually modifying the code, use Sonnet for good cost-effectiveness.
B. Agent Teams (Research Preview)
This feature allows multiple agents to collaborate by sending messages to each other. It’s still in the experimental stage, so environment variables need to be set.
# Enable Agent Teams
export CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1
Usage Tip: When is it useful?
Honestly, it still consumes a lot of tokens, so it’s burdensome to use on a daily basis. But it’s worth trying in these situations.
- Large-scale multi-module refactoring: When modifying module A requires modules B and C to change in a chain reaction.
- API spec changes: When the backend DTO changes, all layers from Remote → Data → Domain → Presentation need to be modified simultaneously.
- Migration tasks: For example, tasks spanning the entire project, such as Gson → Kotlin Serialization.
The difference from the existing Task agent (sub-agent) is that Task delegates work in one direction, while Agent Teams allows agents to communicate with each other. Since it’s still a research preview, I recommend using it for experimental purposes rather than production work.
C. Auto Memory
I personally think this is the most practical feature in this update.
Claude Code automatically records what it learns while working in a MEMORY.md file and refers to it in the next conversation. It is stored in the .claude/projects/.../memory/ path inside the project directory.
Usage Tip: Memory Management
If you leave the memory unattended, it can be filled with unnecessary content. This management can help.
# MEMORY.md (keep under 200 lines)
## Project Structure
- feature module is Compose, comics module is XML + DataBinding
- Remote API must return DataResponse<T>
## Common Mistakes
- Do not use try-catch directly without using safeApiCall
- Do not use viewModelScope.launch directly in ViewModel, use onMain/onIO
## Detailed Note Links
- Build Issues: debugging.md
- Code Patterns: patterns.md
The key is to keep MEMORY.md concise because it will be truncated if it exceeds 200 lines, and separate the details into separate files. It’s similar to writing project rules in CLAUDE.md, but MEMORY.md is different in that Claude automatically learns and accumulates content.
D. “Summarize from here” Feature
You can summarize the conversation from a specific point in the message selector.
Usage Tip: Context Management
When working with Claude Code for a long time, there comes a point when the context window becomes insufficient. In the past, I would just start a new session or leave it to automatic compression, but now I can summarize based on the desired point.
This pattern was effective.
- Complete the exploration/investigation phase (reading files, understanding structure)
- Run “Summarize from here” here
- Start the implementation phase (exploration results are compressed into a summary, focus on code modification)
The implementation quality is improved because unnecessary exploration logs do not take up context, while key information is maintained.
E. –add-dir Skill Auto-Loading
.claude/skills/ files in directories added with --add-dir are also automatically recognized.
Usage Tip
You can collect common skills in a separate directory and load them with --add-dir for each project. With this update, skills are automatically loaded, making command reuse much easier.
2. Bug Fixes
A. @ File Autocomplete Path Correction
When running Claude Code in a subdirectory, there was a problem with the relative path being twisted when referencing @ files. For example, if you ran it in feature/offerwall/, the @src/main/... path would not be properly captured, but this has been fixed.
B. Bash Heredoc Template Literal Error
This is a welcome fix because I experienced this problem myself. In the Bash tool, if a JavaScript template literal (like ${index + 1}) was included in a heredoc, a “Bad substitution” error would occur. The cause was that the shell was interpreting ${} as variable substitution, but it is now properly escaped.
# Previous: This code would cause an error if it was in a heredoc
cat <<EOF
const item = items[${index + 1}] // Bad substitution!
EOF
# v2.1.32: Works normally
C. Thai/Lao Vowel Rendering
The problem of Thai vowels being broken in the input field has been fixed. (This does not directly affect Korean users, but it is good to know when working with multilingual tasks)
D. [VSCode] Fixed Slash Command Malfunction
The problem of slash commands being executed incorrectly when pressing Enter with text entered in front of them in VSCode has been fixed. This will be noticeable for those who use Claude Code as a VSCode extension.
3. Improvements
A. –resume Agent Auto-Reuse
When continuing a previous conversation with --resume, the --agent value used in that conversation is automatically reused.
Usage Tip
This is useful when creating and using custom agents. For example, if you are working with a code review agent and then disconnect the session, the review agent settings will be automatically restored when you continue with claude --resume later. In the past, you had to reattach the --agent flag.
B. Skill Character Budget Expansion
The character budget allocated to skill descriptions has been expanded to 2% of the context window. This is good news for those who register and use many skills. If you use a large context window, more skill descriptions will fit without being truncated.
C. [VSCode] Conversation List Loading Spinner
Minor, but a loading spinner has been added when loading the past conversation list. For UX improvement.
4. Update Application Routine
For reference, I’ll share the order in which I check the release notes and apply them.
# 1. Update
npm update -g @anthropic-ai/claude-code
# 2. Check Version
claude --version
# 3. Initial MEMORY.md Setup (When using Auto Memory for the first time)
# Write a summary of the project's core rules in the .claude/projects/.../memory/MEMORY.md file
# Claude will add it automatically, but setting the initial value improves the directionality
# 4. Test New Features
# Check Opus 4.6 performance with a simple task
claude --model claude-opus-4-6 "Analyze this project structure"
Conclusion
v2.1.32 feels like an update that improves real-world usability rather than flashy, large-scale features. In particular, Auto Memory and Summarize from here are features that directly solve the chronic problems of long sessions (lack of context, loss of context), so I think I will use them often.
Agent Teams is still in the experimental stage, but the direction itself is interesting. I believe that the structure of agents collaborating with each other could change the landscape of large-scale codebase work once it is stabilized.
And the heredoc bug fix… honestly, I’m personally most happy about this fix because I remember struggling with it once.
I hope this article helps those who use Claude Code, even just a little.