Claude Code Digest — 2026-05-15 00:35:27
Version updates
Release 2.1.142 transforms Claude Code from a single-threaded assistant into a background fleet manager. The claude agents command now accepts a robust set of configuration flags—including --settings, --mcp-config, --model, and --effort. Fast mode now defaults to Opus 4.7, confirming Anthropic's confidence in the minor version's latency improvements for rapid-fire developer interactions.
What the docs reveal
Securing and Constraining Hooks
Anthropic aggressively locked down the hook lifecycle. The documentation reveals a new 10,000-character limit on hook outputs (stdout, additionalContext, and systemMessage). Claude now truncates larger outputs to a preview. This prevents poorly written tools—like a database query returning thousands of rows—from instantly blowing out the context window.
Security policies also grew stricter. Hooks can no longer access /dev/tty directly to manipulate the terminal. Instead, developers must return a terminalSequence JSON field for operations like setting window titles or triggering desktop notifications. Coupled with a strict 30-second default timeout for UserPromptSubmit hooks, these changes prioritize session responsiveness over tool flexibility. Anthropic clearly saw uncontrolled hooks freezing the CLI and poisoning the model state.
You control agent retry loops better now. By returning permissionDecision: 'deny' along with a permissionDecisionReason, you explicitly block tools and prevent the model from infinitely retrying a failed action.
Context Economy and Checkpoints
Context windows fill quickly during complex refactors, degrading model performance and increasing API costs. The new checkpointing documentation addresses this workflow friction directly. Claude Code now snapshots your files prior to every prompt.
Pressing Esc twice (or typing /rewind) accesses targeted summarization. You choose between "Summarize from here" and "Summarize up to here." Anthropic added a critical developer UX win: summarizing up to a specific checkpoint now auto-restores the original prompt into your input field. When you go down a debugging rabbit hole, you can compress the dead-end history, keep the meta-context, and instantly adjust your previous approach. The summarized segments remain in the session transcript. You save tokens; you lose no context.
Background Sessions as Compute Instances
Anthropic positions background agents as provisionable compute instances. The new claude agents CLI flags allow you to assign persistent defaults for --model, --effort, and --cwd (working directory). You isolate sessions by project scope.
The security model requires attention. You cannot launch a headless agent using --permission-mode bypassPermissions or auto until you approve those permissions at least once in an interactive claude session. Anthropic balances the efficiency of autonomous agent fleets with the danger of blind execution.
Agent organization also scaled up. Claude Code supports recursive directory structures within .claude/agents/. When you invoke a plugin agent stored in a subfolder, you must use the full scoped path (e.g., plugin-name:subfolder:agent-name).
Enterprise Identity and CI/CD Operations
Claude Code anticipates complex environments. The new ANTHROPIC_WORKSPACE_ID environment variable supports workload identity federation, allowing enterprise teams to attribute cloud-based API costs to specific workspaces.
Continuous Integration environments receive a dedicated fix. CI runners typically lack configured SSH keys, causing agent plugin clones to fail. Setting CLAUDE_CODE_PLUGIN_PREFER_HTTPS forces Git to use HTTPS, keeping automated deployment pipelines running smoothly.
Model Context Protocol (MCP) servers bridge the gap to authenticated corporate data. When a remote MCP server returns a 401 Unauthorized or 403 Forbidden status code, Claude Code automatically triggers a local OAuth flow. Custom MCP servers parse the standard HTTP WWW-Authenticate header to route developers to the correct authorization server. You can write custom MCP servers that broker access to secure internal APIs without hardcoding secrets into your .claude.json configuration.
WSL Audio Engineering
Voice dictation under Windows Subsystem for Linux (WSL) remains complicated. The documentation specifies that WSL users must run WSLg and install libsox-fmt-pulse to enable PulseAudio support for SoX. Voice dictation requires bare-metal microphone access; it fails entirely over remote SSH sessions or web-based instances. Stick to the keyboard if you use remote dev boxes. Use CLAUDE_CODE_SIMPLE_SYSTEM_PROMPT to strip boilerplate from the system instructions and save context space.