Claude Code Digest — 2026-04-15 00:17:00
Version updates
Anthropic shipped two minor releases focused on cache optimization and user retention during long tasks.
Version 2.1.108 exposes raw prompt caching controls to developers. You can now opt into a 60-minute time-to-live with ENABLE_PROMPT_CACHING_1H across all major backends (API, Bedrock, Vertex, Foundry), or aggressively dump memory with FORCE_PROMPT_CACHING_5M. Anthropic recognizes that agent workflows break into two distinct patterns: deep, ongoing feature development (needing long cache spans) and quick, transient scripting (where long caches waste money). The update also introduces a "recap" feature that summarizes context when you return to the terminal.
Version 2.1.107 addresses user perceived latency. Claude Code now renders thinking hints earlier during expensive operations, preventing developers from aborting valid requests under the assumption the CLI hung.
What the docs reveal
Ephemeral Sessions for CI/CD
Anthropic explicitly targeted automated environments with this documentation pass. Developers receive an ecosystem of tools to run traceless sessions: the CLAUDE_CODE_SKIP_PROMPT_HISTORY environment variable, the --no-session-persistence CLI flag, and the persistSession: false SDK parameter.
These additions solve a major security and hygiene problem. Prior constraints forced Claude to write session telemetry and transcripts to disk. Now, developers can execute Claude Code inside stateless GitHub Actions without leaking codebase context into cache directories or bloating runner storage environments.
Context Branching and Subagent Cleanup
Managing token context dominates the latest workflow adjustments. The new /branch (alias /fork) command fundamentally changes how developers iterate. Instead of cluttering a primary session with failed debugging attempts, you can fork the conversation, test a risky refactor, and discard the branch if it fails.
At the SDK level, Anthropic formalized subagents. Main agents can now spawn specialized child processes. Because multi-agent orchestration generates rapid file bloat, Anthropic rewired the cleanupPeriodDays setting to automatically purge orphaned subagent worktrees on startup. You must define this value (setting it to 0 triggers an error).
Automation Built with Kill Switches
The documentation highlights robust scheduling via the /schedule command and integration with external Channels like Discord and Telegram. You can instruct Claude to trigger terminal tasks via webhooks.
However, Anthropic implemented a strict guardrail: recurring tasks automatically expire 7 days after creation. They execute a final time and self-delete. Anthropic engineered this kill switch to protect users from "zombie" schedules running unnoticed and silently draining token balances.
Strict Permissions and Environment Rules
Anthropic tightened system requirements across the board. The documentation reveals a total deprecation of unauthenticated usage; the CLI now strictly demands authentication.
On Windows, developers must execute Claude Code through git-bash. Native Windows consoles consistently struggle with complex UI rendering and virtual terminal capabilities required by Claude Code. Anthropic also dropped all 32-bit system support.
For developers building agent logic, the principle of least privilege takes center stage. Skills now possess absolute file system and terminal execution access. The updated documentation explicitly warns developers to build idempotent custom skills and isolate mcp_tools permissions, acknowledging the inherent danger of providing LLMs unfettered local system control.
IDE Visuals and Browser Automation
We finally see tight IDE orchestration. The VS Code integration introduces interactive diffs, visual approval flows, and @-mention context awareness directly from the editor toolbar. JetBrains users also receive explicit instructions to wire Claude Code as a native terminal tool.
Anthropic pushed a Playwright integration example to the MCP documentation. This signals Anthropic's direction for the tool ecosystem: they want developers leaning on Claude to execute end-to-end automated browser testing, snapping screenshots, and navigating authentication flows autonomously.