Claude Code v2.1.49 Released: Background Agent Kill Fix, Startup Speedups, and Yoga WASM Memory Leaks Squashed
Anthropic ships Claude Code v2.1.49 with roughly 20 changes, including a fix for Ctrl+C and ESC being silently ignored during background agent execution, multiple startup performance improvements via MCP batching and auth-failure caching, an end to Yoga WASM memory growth in long sessions, file editing in simple mode, and new SDK capability fields.
Anthropic published Claude Code v2.1.49 to npm on February 20, 2026 (UTC), roughly one day after v2.1.47. The release contains approximately 20 changes focused on background agent control, startup performance, persistent memory leaks in long-running sessions, and SDK improvements. v2.1.48 does not appear to have reached the latest tag — v2.1.49 is the direct successor.
Background Agent Control Is Now Reliable
One of the more frustrating regressions in recent builds has been fixed: pressing Ctrl+C or ESC while background agents are running and the main thread is idle was silently ignored. Starting with v2.1.49, pressing either key twice within three seconds will terminate all background agents. For users running parallelized agentic workflows, this restores a basic and critical escape hatch.
Startup Performance: Three Layers of Improvement
Anthropic tackled startup latency from multiple angles in this release.
First, MCP tool token counting is now batched into a single API call instead of one call per tool. On setups with many MCP servers installed, this alone reduces startup HTTP traffic substantially.
Second, authentication failures for HTTP and SSE MCP servers are now cached, eliminating repeated connection attempts to servers that require authentication the user hasn’t configured. Previously, every startup would hammer unavailable endpoints.
Third, the non-interactive mode (-p flag) now skips unnecessary API calls during initialization, making scripted and CI/CD usage faster and cheaper.
Yoga WASM Memory Growth Fixed — Twice
Two separate bugs caused Yoga (the layout engine used for Claude Code’s terminal UI) to grow its WASM linear memory without bound during long sessions. Both are fixed in v2.1.49. Separately, the tree-sitter parser — used for syntax-aware code handling — is now periodically reset to prevent its own WASM memory from accumulating. Stale yoga layout references that could trigger rendering glitches have also been cleaned up.
The practical result: sessions running for hours should no longer see continuously growing RAM usage from these components.
Simple Mode Gains File Editing
The CLAUDE_CODE_SIMPLE environment variable enables a stripped-down interaction mode. Previously this mode only exposed the Bash tool. v2.1.49 adds the file edit tool to simple mode, allowing direct file creation and modification without going through shell commands. This makes simple mode usable for a broader class of coding tasks.
Plugin Scope Auto-Detection
plugin enable and plugin disable now automatically detect the correct scope (user or project) when --scope is not passed explicitly. Previously, the commands always fell back to user scope, which could silently operate on the wrong configuration layer when working inside a project with project-scoped plugins.
SDK Improvements
Model capability discovery. The SDK model info object now exposes supportsEffort, supportedEffortLevels, and supportsAdaptiveThinking fields. SDK consumers no longer need to hardcode per-model capability checks; they can query these fields at runtime.
ConfigChange hook. A new ConfigChange hook event fires whenever configuration files change during an active session. Enterprise deployments can use this to implement security auditing or to block unauthorized settings changes through policy.
Permission suggestions on safety blocks. When a safety check triggers an ask response, permission options are now populated and surfaced to SDK consumers, enabling richer UX in tools built on top of the SDK.
Other Notable Fixes
- Prompt suggestion cache regression fixed. Cache hit rates had dropped in a recent build; this release restores expected behavior.
- Verbose mode toggle bug fixed. Using
/configto toggle verbose mode was not updating the thinking block display. The memo comparators now correctly detect verbose setting changes. --resumepicker cleaned up. Sessions that began with commands like/clearwere showing raw XML tags in the session picker. The display now falls through to the session ID fallback correctly.disableAllHookshierarchy fix. ThedisableAllHookssetting was allowing non-managed configurations to disable hooks set by enterprise policy. The fix enforces that managed hooks cannot be disabled by user-level settings (#26637).- Permission prompts now show context. Prompts triggered by path safety and working directory blocks now display the reason for the restriction rather than presenting a bare prompt with no explanation.
Sonnet 4.6 Gets 1M Context on Max Plan
Anthropic is removing Sonnet 4.5 with 1M context from the Max plan. Sonnet 4.6 — the current frontier Sonnet model — now supports 1M context and replaces it. Users on the Max plan who have Sonnet 4.5 selected in /model will need to switch manually.
Version Details
- Version: 2.1.49
- Published: February 20, 2026 (UTC)
- npm: @anthropic-ai/claude-code
- Install / upgrade:
npm install -g @anthropic-ai/claude-code@2.1.49
関連記事
Claude Code v2.1.47 リリース - Windows対応強化・メモリ改善・40件超のバグ修正
Anthropicは2026年2月19日、Claude Code v2.1.47をリリースした。Windows環境の描画バグ修正、長時間セッションのメモリ改善、起動速度の向上(約500ms短縮)、コンテキスト圧縮後のプランモード保持など40件以上の修正が含まれる。
Claude Code v2.1.49 リリース - バックグラウンドエージェント制御改善・起動高速化・Sonnet 4.6の1Mコンテキスト対応
Anthropicは2026年2月20日、Claude Code v2.1.49をリリースした。バックグラウンドエージェント実行中のCtrl+C/ESC無視バグの修正、起動パフォーマンスの大幅改善、シンプルモードへのファイル編集ツール追加、Yoga WASMメモリリークの修正などが含まれる。また、MaxプランにおいてSonnet 4.6が1Mコンテキストに対応した。
Claude Code v2.1.50 リリース - Worktreeフック追加・大規模メモリリーク修正・Opus 4.6が1Mコンテキスト対応
Anthropicは2026年2月21日(JST)、Claude Code v2.1.50をリリースした。WorktreeCreate/WorktreeRemoveフックの追加、エージェント定義でのworktreeアイソレーション宣言対応、SSHセッション切断時のデータ損失修正、glibc 2.30以前のLinuxでのネイティブモジュール読み込み不具合修正、長時間セッションを対象とした複数のメモリリーク修正、Opus 4.6(高速モード)への1Mコンテキスト対応拡大などが含まれる。
人気記事
ChatGPT(OpenAI)とClaude(Anthropic)の機能比較 2026年版。コーディング・長文解析・コスト・API料金の違いを検証
ChatGPT(GPT-4o/o3)とClaude(Sonnet 4.6/Opus 4.5)を2026年時点の最新情報で比較する。コーディング能力、長文処理、日本語品質、API料金、無料プランの違いをSWE-benchなどのベンチマーク結果とともに解説する。
【2026年2月20日 所感】「AIがコードを書く」は仮説から現実になった——しかし私たちはその意味をまだ消化できていない
2026年2月20日に観測したコーディングエージェント関連ニュースの総括と所感。Anthropicの自律性研究、cmux、MJ Rathbunのエージェント事故、HN「外骨格 vs チーム」論争、Stripe Minions週1000件PR、Taalas 17k tokens/sec——朝から夜までの流れを通じて見えてきた「AIがコードを書く時代」の実相を考察する。
868のスキルをnpx 1コマンドで——「Antigravity Awesome Skills」が主要AIコーディングエージェントの共通スキル基盤になりつつある
Claude Code・Gemini CLI・Codex CLI・Cursor・GitHub Copilotなど主要AIコーディングアシスタントを横断する868以上のスキルライブラリ「Antigravity Awesome Skills」(v5.4.0)を詳細分析。Anthropic・Vercel・OpenAI・Supabase・Microsoftの公式スキルを統合した設計思想、ロール別バンドル・ワークフロー機能、SKILL.mdによる相互運用性のアーキテクチャを解説する。
最新記事
AIエージェント間通信の標準化競争が始まる——AquaとAgent Semantic Protocolが同日登場
2026年2月23日、Hacker Newsに2つのAIエージェント通信プロジェクトが同日掲載された。Go製CLI「Aqua」とセマンティックルーティングを実装する「Agent Semantic Protocol」は、MCPが解決できないP2P・非同期通信の課題に取り組む。
Claude Sonnet 4.6、無料・Proプランのデフォルトモデルに——社内テストでOpus 4.5を59%の確率で上回る
Anthropicは2026年2月17日にリリースしたClaude Sonnet 4.6を、claude.aiの無料・Proプランのデフォルトモデルに設定した。価格はSonnet 4.5と同額の$3/$15 per 1Mトークン。社内評価ではコーディングエージェント用途でOpus 4.5を上回る結果が出ている。
GoogleがOpenClaw経由のGemini利用ユーザーのアカウントを永久停止——月額$250請求継続のまま
2026年2月23日、Hacker Newsで140pt/107コメントを集めたレポートによると、GoogleはOpenClaw(サードパーティクライアント)経由でGeminiを使用していたGoogle AI Pro/Ultraユーザーを予告なしに永久停止した。技術的・経済的背景を整理する。