Claude Code Digest — 2026-05-13 00:36:55
Version updates
Release 2.1.140 addresses usability friction in the Agent workflow. Anthropic improved subagent_type matching to handle case and separator insensitivity. Previously, an LLM generating "Code Reviewer" instead of code-reviewer would fail to trigger the correct tool. The agent now resolves these variations automatically, reducing prompt engineering overhead.
This release also patches a silent failure loop. The /goal command now detects when disableAllHooks or allowManagedHooksOnly blocks its execution and returns an explicit error message, saving developers from debugging unresponsive terminals.
What the docs reveal
The Death of TodoWrite
Anthropic deprecated the monolithic TodoWrite tool in favor of a granular task suite: TaskCreate, TaskGet, TaskUpdate, and TaskList. You can test these tools behind the CLAUDE_CODE_ENABLE_TASKS=1 environment variable.
LLMs struggle with updating large blocks of text directly. Editing a single line in a monolithic list frequently introduces formatting errors or hallucinated deletions. Granular CRUD operations map cleanly to the LLM's tool-calling capabilities. Developers building against the Agent SDK must update their integration patterns to track state using a taskId map rather than overwriting full arrays. Anthropic will eventually remove TodoWrite entirely.
Parallel Workflows via Git Worktrees
The new Agent View, accessed via claude agents, shifts Claude Code from an interactive script to an async worker. A persistent supervisor process manages background sessions that outlive your terminal.
To prevent parallel agents from thrashing the same files, Claude isolates sessions using Git worktrees. Large monorepos benefit from the new worktree.sparsePaths configuration, limiting checkouts to required directories to save disk space and time. Claude cleans up unmodified worktrees automatically but halts to prompt you before deleting directories containing uncommitted changes.
Prompt Caching Economics
Anthropic added the --exclude-dynamic-system-prompt-sections flag. This configuration moves volatile environment data, like the current working directory, from the system prompt into the first user message.
System prompt caching saves massive amounts of time and money, but shifting paths invalidate the cache. Moving dynamic paths out of the system prompt guarantees a higher cache hit rate across multi-machine and multi-user environments.
The documentation also clarifies the prompt customization hierarchy. CLAUDE.md runs as an injected user message, --append-system-prompt adds skills to the assistant, and Output Styles completely rewrite the system prompt to change the agent's core identity. You can layer these pieces modularly.
Security Rigidity vs. Alert Fatigue
Claude Code now executes read-only standard commands (like ls, grep, and git diff) without user permission. Constantly prompting developers for harmless read operations causes alert fatigue, training users to blind-approve actions. Removing prompts for safe commands speeds up the interaction loop.
Conversely, Anthropic instituted a hard block on executing the --dangerously-skip-permissions flag using root or sudo. If you run automated workflows in CI or containers, you must configure a non-root user. Embedding hosts can also force security policies using the managedSettings configuration, overriding local project settings entirely.
Skill Discovery in Monorepos
Skill discovery logic now maps perfectly to monorepo architectures. Claude Code recursively searches for .claude/skills/ directories up to the repository root and down into child directories. You can drop custom test-runner skills directly into sub-packages without registering them globally.
Hooks and configuration settings act differently. The SDK restricts settings to the immediate <cwd>/.claude/ directory, preventing unpredictable behavior caused by rogue scripts higher up the directory tree. When configuring hooks, favor the exec form ("args" array) over shell scripts to bypass quoting errors, using ${CLAUDE_PROJECT_DIR} to map absolute paths.
Opus 4.7 Cutover Details
Claude Opus 4.7 becomes the default for Fast Mode tomorrow. You can trigger early access via CLAUDE_CODE_ENABLE_OPUS_4_7_FAST_MODE.
Fast Mode queries for Opus 4.6 and 4.7 now draw from the same rate limit pool. Hitting the limit triggers an automatic downgrade to standard speed on your current model. If your workflows require behavioral consistency during the rollout, pin your version using CLAUDE_CODE_OPUS_4_6_FAST_MODE_OVERRIDE=1.