Claude Code 2.0

TL;DR Highlight

Anthropic's terminal-based coding agent Claude Code updated to 2.0 with native VS Code extension, conversation rewind (/rewind), Claude Agent SDK, and more.

Who Should Read

Developers already using or considering Claude Code. Especially backend/full-stack developers building agent-based coding workflows.

Core Mechanics

Native VS Code extension launched. Claude Code, previously terminal-only, now works directly within VS Code UI. Coincidentally timed with Cursor launching its own CLI.
Added /rewind command — a checkpoint feature to rewind conversations and undo code changes when you went the wrong direction. One of the most celebrated features.
New /usage command to check current plan usage limits from within the agent.
/model sonnet[1m] must be explicitly specified for the 1M context version of Sonnet. Not the default, so requires separate configuration for long-context work.
Tab key toggles thinking mode, and this setting persists (sticky) across sessions.
Claude Code SDK renamed to Claude Agent SDK. The --agents flag enables dynamic sub-agent addition for multi-agent collaboration defined in code.
Hooks bugs fixed — PostToolUse event tool_use ID and tool_result block mismatch resolved.

Evidence

Reaction to /rewind (checkpoints) was especially enthusiastic. 'FINALLY checkpoints!' — a long-requested feature.
Claude Code being used as a general-purpose agent beyond coding was shared. One user had it analyze baby video, calculate crop coordinates, and execute ffmpeg commands. A GitHub repo of 'non-coding Claude Code use cases' was linked.
CJK (Korean/Chinese/Japanese) IME users reported issues — IME uses Enter for candidate confirmation, but Claude Code treats it as submission, breaking input. Shift+Enter for submission or IME detection logic needed.
Discussion about whether Sonnet 4.5 surpassing Opus 4.1 on several benchmarks makes the Plan=Opus / Code=Sonnet split no longer optimal. Some argued unifying on Sonnet may now be better.

How to Apply

Use /rewind as a checkpoint: before starting a big refactor or experimental change, remember the current state. If results don't satisfy, /rewind and try a different direction.
For multi-agent workflows, use Claude Agent SDK's --agents flag to dynamically add sub-agents. For example, separate code-writing and test-writing agents running in parallel.
For large codebase work needing 1M token context (e.g., full legacy code analysis), run /model sonnet[1m] at session start to prevent mid-session context truncation.
Claude Code can be used beyond coding (video editing, file organization, data processing). If CLI tools like ffmpeg and imagemagick are installed, Claude Code assembles appropriate commands from natural language instructions.

Terminology

Claude Agent SDKThe renamed Claude Code SDK. A library for building multi-agent systems where multiple agents collaborate, defined in code.

PostToolUse HookAn automatic post-processing script triggered after the agent uses a tool (e.g., file write, command execution). Can be configured to auto-run tests after file modifications.

IMEInput Method Editor. Software for composing characters in languages like Korean, Chinese, and Japanese. Enter key serves as character confirmation, which can conflict with app submission functions.

Context WindowThe maximum text length an LLM can process at once. 1 million tokens is roughly equivalent to dozens of long novels at once.

Sub-agentAn independent agent instance spawned by the main agent for specific subtasks. Useful for parallel processing or role separation.