Claude Code's Latest Update Adds /recap, 1-Hour Prompt Caching, and Skill-Invokable Built-In Commands

Anthropic shipped version 2.1.108 of Claude Code on April 14, with a follow-up 2.1.109 on April 15. The update is mostly focused on session management, performance, and a handful of features that have been requested since the tool went mainstream.

/recap: Context When You Come Back

If you’ve ever returned to a long Claude Code session after a break and immediately lost track of where things stood, /recap is the fix. It generates a summary of what happened in the current session, what’s been done, and where things were left off.

The feature is configurable from /config, and you can invoke it manually at any time with /recap. If you have telemetry disabled (DISABLE_TELEMETRY), you can force the away summary with the environment variable CLAUDE_CODE_ENABLE_AWAY_SUMMARY.

Prompt Caching Controls: 1-Hour TTL and Forced 5-Minute TTL

Two new environment variables give you direct control over Claude’s prompt cache TTL on self-managed API setups.

ENABLE_PROMPT_CACHING_1H opts into a 1-hour cache TTL. It works across API key setups, AWS Bedrock, Google Vertex AI, and Microsoft Foundry. If you were previously using ENABLE_PROMPT_CACHING_1H_BEDROCK, that variable still works but is now deprecated in favor of the unified option.

FORCE_PROMPT_CACHING_5M does the opposite: it locks in a 5-minute TTL regardless of other settings. Useful if you’re testing cache behavior or working in an environment where stale cache is a real problem.

There’s also a new startup warning if you have DISABLE_PROMPT_CACHING* variables set, so you can’t accidentally run an expensive session with caching off and only notice after the fact.

Skill Tool Can Now Call Built-In Slash Commands

Previously, the Skill tool could only invoke user-defined skills. Now it can also discover and call built-in slash commands: /init, /review, and /security-review.

In practice, this means you can build workflows where Claude programmatically kicks off a code review or security scan as part of a larger automated task, rather than waiting for a human to type the command. The changelog notes a minor inconsistency in how this is documented versus how it was originally announced (as “merged slash commands and skills”), but the behavior is what matters: built-in commands are now first-class Skill targets.

/undo as an Alias for /rewind

Small quality of life: /undo now works as an alias for /rewind. If you’ve been instinctively typing /undo and hitting an error, that stops now.

/model Warnings Before Mid-Conversation Switches

Switching models mid-session now prompts a warning. The reason: when you switch models, the next response re-reads the full conversation history from scratch without the benefit of cached context. On long sessions, that can mean a significant latency hit and token cost spike. The warning makes that explicit so it’s not a surprise.

/resume Picker Scoped to Current Directory

The /resume session picker now defaults to showing sessions from the current directory rather than all sessions across all projects. If you want the full list, press Ctrl+A. This makes it faster to pick up where you left off on the project you’re in without scrolling through sessions from other repos.

Better Error Messaging

Rate limit errors now distinguish between server rate limits (Claude’s infrastructure is throttling requests) and plan usage limits (you’ve hit your subscription cap). They were previously reported with identical messaging, which made it hard to know whether to wait a few seconds or upgrade your plan.

Any 5xx or 529 error now includes a direct link to status.claude.ai. Unknown slash commands suggest the closest known command, so a typo like /reviw will offer a suggestion rather than just failing.

Memory Footprint Reduction

Language grammars for syntax highlighting now load on demand instead of all at once at startup. For users working across many languages in a session, this cuts the base memory footprint for file reads, edits, and syntax highlighting. The difference is most noticeable on machines with less RAM or in long-running sessions where multiple files across different languages have been opened.

Notable Bug Fixes

A few fixes worth knowing about:

• Paste was broken in the /login code prompt since 2.1.105. Fixed.
• Users with DISABLE_TELEMETRY set were accidentally falling back to the 5-minute prompt cache TTL instead of 1-hour even when they had opted in to the longer TTL. Fixed.
• Diacritical marks (accents, umlauts, cedillas) were being dropped from responses when the language setting was configured. Fixed.
• claude --resume <session-id> was losing the session’s custom name and color set via /rename. Fixed.
• Remote Control session titles set in the web UI were being overwritten by auto-generated titles after the third message. Fixed.

The full list is in the official changelog.