Claude Code Daily Briefing - 2026-04-15

Release Summary

Full Release Notes

New Features & Practical Usage

Prompt Caching TTL Configuration — Direct Token Cost Control (v2.1.108)

v2.1.108 lets you control prompt caching TTL (Time-To-Live) via environment variables. After the recent reduction from 1 hour to 5 minutes that worsened quota depletion for Pro Max 5x users, this puts control back in your hands.

# Enable 1-hour cache TTL (recommended for subscribers)
export ENABLE_PROMPT_CACHING_1H=1

# Force 5-minute cache TTL (fine-grained control for API key users)
export FORCE_PROMPT_CACHING_5M=1

Works across all providers: API keys, Bedrock, Vertex, and Foundry. A related bug where DISABLE_TELEMETRY caused fallback to 5-minute cache instead of 1-hour was also fixed. If you’re burning through quota, apply ENABLE_PROMPT_CACHING_1H=1 immediately.

GitHub v2.1.108

/recap — Session Context Recovery on Return (v2.1.108)

When returning to a session after time away, /recap provides a summary of your current session state. Claude analyzes the conversation history to tell you what you were working on, how far you got, and what to do next.

# Get session context summary on return
/recap

# Auto-activate away summary detection
# Configure via /config or:
export CLAUDE_CODE_ENABLE_AWAY_SUMMARY=1

If /resume reopens a session, /recap is about understanding where you are in it. Particularly useful for next-day work resumption on long-running projects.

GitHub v2.1.108

Routines — Cloud-Based Scheduled Automation (4/14)

Claude Code now supports Routines — scheduled, repeatable automations that run on Anthropic’s web infrastructure. Your local machine doesn’t need to stay online.

While /loop runs locally in your active session, Routines enable workflows like “run code review every day at 9 AM” or “check dependency updates every Friday.” Built on Managed Agents’ Session architecture, this further blurs the line between local and cloud-based agent work.

9to5Mac

Developer Workflow Tips

/recap + /resume Combo for Multi-Project Session Management

Combine v2.1.108’s /recap with the improved /resume picker for smooth multi-project context switching.

# 1. Switch to another project session
/resume
# → Default: shows only current directory sessions
# → Ctrl+A: shows all project sessions

# 2. Get context after switching
/recap
# → Instant "where did I leave off?" summary

The /resume picker now defaults to current directory sessions, making per-project session management much cleaner.

Slash Command Discovery — Claude Invokes Tools Autonomously (v2.1.108)

Starting with v2.1.108, Claude can invoke built-in slash commands like /init, /review, and /security-review via the Skill tool. Instead of requiring explicit slash command input, Claude now judges the appropriate command from context and executes automatically.

Say “initialize this project” and Claude calls /init. Say “do a security review” and it calls /security-review. Custom skills are discoverable the same way.

Plan-First Workflow — Quantitative Evidence

Community testing shows that asking AI coding agents to implement directly yields ~33% success rate. Following a plan-first workflow — draft plan → review → revise → implement — significantly improves outcomes. This is quantitative justification for using /plan mode or Plan Mode consistently.

Security & Limitations

US Treasury Seeks Mythos Access for Financial System Vulnerability Detection (4/14)

Per Bloomberg, the US Treasury has requested access to Anthropic’s Claude Mythos for defensive use — detecting security vulnerabilities in financial systems.

This follows the 4/12 emergency meeting between the Federal Reserve and Treasury, reflecting the assessment that Mythos’s zero-day detection capabilities (181 successful exploits in Firefox JS engine) could protect national financial infrastructure. With federal agencies now seeking access alongside Project Glasswing’s ~50 partner organizations, Mythos’s influence is expanding from the private sector into government.

Bloomberg

v2.1.108 — 16 Bug Fixes

Key stability fixes:

• /login paste broken: v2.1.105 regression — auth code paste didn’t work in login prompt
• DISABLE_TELEMETRY cache fallback: Telemetry-off users fell back to 5-minute cache instead of 1-hour
• Auto mode permission error: Agent tool unnecessarily requested permission when safety classifier transcript exceeded context window
• CLAUDE_ENV_FILE trailing #: Bash tool produced no output when env file ended with a # comment
• --resume session name loss: Names set via /rename or --name were lost on resume
• --teleport terminal corruption: Escape codes appeared as garbage after teleport
• Silent transcript write failure: Disk-full transcript write failures were silently dropped

GitHub v2.1.108

Ecosystem & Plugins

Anthropic LTBT Appoints Vas Narasimhan to Board (4/14)

Anthropic’s Long-Term Benefit Trust appointed former Novartis CEO Vas Narasimhan as a board member. The LTBT is the independent body overseeing Anthropic’s safety policies, influencing Claude’s behavioral guidelines and model release decisions. A pharma/biotech leader joining AI safety oversight suggests “clinical trial-level systematic verification” frameworks may be applied to AI safety evaluation.

Cursor, Claude Code, Codex — The Unplanned Unified Stack

According to The New Stack’s analysis, Cursor (IDE), Claude Code (terminal agent), and Codex (code generation) are converging into a de facto unified AI coding stack. In early April, Cursor rebuilt its parallel agent orchestration interface, OpenAI published an official Claude Code plugin, and early adopters began running all three simultaneously. Competing tools being used complementarily signals that AI coding is evolving from single-tool to ecosystem-level adoption.

The New Stack

Community News

• Routines launch reaction — excitement meets security concerns: Developers are welcoming the end of local machine dependency, while simultaneously raising concerns about cloud agents having 24/7 repository access. Network access scope will be a key issue for organizations using self-hosted GitLab or private networks. 9to5Mac

• “Harness Engineering” as a new paradigm: GeekNews weekly newsletter (GN#353) highlighted the shift from Prompt Engineering → Context Engineering → Harness Engineering. The key message: copying others’ harnesses is ineffective. True harness engineering means selecting 3-4 skills matching your workflow and customizing them for your project’s conventions. GeekNews

• v2.1.108 startup caching warning: If prompt caching is disabled when starting a session, a warning now appears. Running without caching significantly increases token costs — verify immediately unless you disabled it intentionally.

Minor Changes

• /undo = /rewind alias: More intuitive name for undo operations
• /resume picker improvement: Defaults to current directory sessions; Ctrl+A shows all projects
• /model switch warning: Mid-conversation model switch now warns that the next response will re-read full history uncached
• Reduced memory footprint: Lazy-loading language grammars for file reads, edits, and syntax highlighting
• Better error messages: Server rate limits distinguished from plan usage limits; 5xx/529 errors link to status.claude.com
• Slash command suggestions: Unknown commands now suggest the closest match
• Ctrl+O verbose indicator: “Verbose” indicator added to detailed transcript view

Recommended Reads

• “The Skills Era: Build Your Own Harness” (GeekNews GN#353): Traces the paradigm evolution from Prompt Engineering through Context Engineering to Harness Engineering. The core message: don’t copy someone else’s harness. Select 3-4 skills matching your workflow and customize them for your project conventions. A practical framework for thinking about how to structure AI coding workflows beyond prompt optimization. GeekNews GN#353

• “Cursor, Claude Code, and Codex Are Merging Into One AI Coding Stack Nobody Planned”: The New Stack’s analysis of how three competing AI coding tools are organically converging into complementary roles — IDE + terminal agent + code generation. The insight that AI coding is an ecosystem-level problem, not a single-tool problem, has implications for how teams structure their development toolchains. The New Stack

Interesting Projects & Tools

• Happy — Mobile/Web Remote Client for Claude Code & Codex: Control Claude Code and Codex remotely from mobile (iOS/Android) and web. Features end-to-end encryption, push notifications, and zero telemetry. Captures terminal state and reconstructs it within milliseconds on the target device. Install via npm install -g happy-coder, then run with happy claude. Combined with Routines, enables an always-connected development environment on the go. GitHub | happy.engineering

• damn-my-slow-kt — Automated ISP SLA Enforcement Tool: Automatically measures internet speed daily and files SLA breach complaints when speed drops below 50% of the contracted rate. Built with Playwright for browser automation and supports Discord/Telegram notifications. While Korea-specific (KT Telecom), the pattern of automating repetitive bureaucratic processes through code is universally applicable. GitHub