Researched and written with 315,095 tokens (in 10 · out 32,522 · cache-read 66,585 · cache-create 215,978) across 5 Claude calls

Coding agents: releases and model news

• v2.1.311: Claude Code adds --hold-idle-file and --max-hold-min to self-hosted-runner and fixes long startup/edit freezes tied to CLAUDE.md.
• 0.167.1: Codex Rust patch release, changelog only, no notable features.
• Gemini 4 Carbon in GitHub Copilot: Google's latest model is GA in Copilot, pitched at agentic coding and secure code generation.
• Which Claude model should you build with?: Anthropic guide on picking Fable 5.2, Opus 5.5, Sonnet or Haiku, when to tune effort instead of switching models, and when to put subagents on smaller models for cost.
• Claude usage reset: Pro, Max, Ultra and Team users get one free full limit reset across all models, usable before December 10, 2:00PM PT, after bugs burned extra quota.
• AI Engineer Code Summit: OpenAI on New Ways of Working With Codex: Workshop recap from OpenAI DevEx engineers on how Codex workflows have changed over the past year.

Context, tokens and caching

• ranxianglei/billion-context: Compression plugin claiming 5x fewer tokens, month-long single sessions and a 100K window being enough, while holding a 95-97% prefix-cache hit rate (compression costs ≤2%).

Skills, CLAUDE.md and repo setup

• addyosmani/agent-skills: Production-grade skill pack encoding senior-engineer workflow gates across define, plan, build, verify, review and ship.
• gstack: Garry Tan's free skill pack that turns Claude Code into a review team.
• Tested the top 20 Claude Code skills against random text: Test found grill-me and handoff did not beat random same-length text on one model, a useful check on skill hype.
• My optimized global CLAUDE.md: Community global CLAUDE.md tuning, no detail in the signal.
• Matt Pocock's Home Repo: One home repo holds skills, steering files and wiki, and AGENTS.md points to other repos to clone, so one agent works across many repos.
• Modern Web Guidance: Chrome, Edge, Firefox and MDN-backed skills steering agents off legacy JS toward modern web primitives, though critics warn it hands one browser vendor de facto control over LLM web output.

Agentic tooling and products

• OpenVibe: Canvas of real terminals for running AI coding agents side by side.
• Designing for Agents, Zeno Rocha, Resend: Podcast on why agents recommend Resend over incumbents, covering answer-engine optimization (AEO) as a distribution channel.