Quiet week on the headline front, loud week in the changelogs. Two big open-weight coding models landed within a day of each other, GitHub quietly turned security scanning into an agent job, and Claude Code shipped four releases in three days. Here's what mattered if you ship code with AI.

This Week in AI Coding

• Reflection's Beam goes Apache 2.0 — 501B-parameter sparse MoE with 23B active, built for coding and long-horizon agents, weights on Hugging Face with day-zero vLLM/SGLang support. Real permissive license, not a "free for most" one. The model card has the benchmarks; I'd run it on your own repo before believing anyone's leaderboard.
• Meta's Muse Spark Open — 816B total / 42B active, 1M context, and Meta's demo runs it on two M5 Ultra Mac Studios. Note the license is Meta's own Muse Model License with limited commercial terms, so read it before you build a product on it.
• Gemini 4 Argon lands in GitHub Copilot — gradual rollout to Pro+, Max, Business and Enterprise across VS Code, JetBrains, Xcode, the CLI and more. Billed at provider list rates under usage-based billing, so a model picker switch is now a budget decision too.
• GitHub's agentic autofix for code scanning is GA — the agent explores the codebase, proposes a fix, and reruns CodeQL to confirm the alert is gone before opening a PR. That verify-before-PR loop is the part I actually want from every coding agent.
• AI Scan for pull requests is GA — AI-driven detection for stuff CodeQL doesn't cover yet (PHP, Shell/Bash, Dockerfiles). Findings are advisory and can't block merges, which is the right default for now.
• Cursor ships cloud agent canvases (Oct 16) — cloud agents can build interactive canvases you share by link, plus a beta Environments API and environments that load 5x faster with up to 300 repos. Earlier in the week the Origin App Marketplace opened to public submissions.
• Claude Code 2.1.299 to 2.1.302 — idle compaction (compacts a 200K+ token session left idle ~54 minutes, before the prompt cache expires), UserPromptSubmit hooks now run before a skill expands, and fork subagents are on by default in claude -p and the Agent SDK. Heads up: MCP allow rules now match only the named server, so write mcp__<server>__* if you relied on prefix matching.

What We Shipped

Mostly maintenance this week: three more content posts went out across the portfolio sites, the daily billing tracker ran clean, and the weekly analytics snapshots are in. No big launches, and I'd rather say that than pad it.

Tool of the Week

Claude Code idle compaction. If you leave long sessions open over lunch, the one-hour prompt cache expires and your next message re-pays for the whole context. Idle compaction compacts it just before that happens. Toggle it with idleCompaction in settings.json or the new row in /config. Small change, real money saved on long-running sessions.

Quick Links

• GitHub changelog: Gemini 4 Argon in Copilot
• GitHub changelog: agentic autofix GA
• Reflection: Beam is available
• Claude Code changelog

Building in public: endofcoding.com · vibecodingebook.com · llmhire.com · YouTube @endofcoding

— Guy