Claude Code now refuses to clone games

Hi there, this is your daily ☕️ Devshot.

Claude Code now refuses to clone games LINK

• Anthropic has added guardrails to Claude Code that make it refuse to recreate someone else's specific game, citing its IP-infringement policy, after the move drew heavy criticism from developers over the past week.
• Addy Osmani confirmed the firm appended two sentences to Claude Code's prompt on Opus 5.5 steering it toward original games and genres but away from cloning an existing title, even if renamed; original-game work is meant to keep functioning.
• The blocking appears tied to Opus 5.5 via Claude Code, so users may sidestep it by calling Opus through Ollama or a wrapper; similar refusals now hit reverse-engineering tasks, with one Redditor briefly suspended for the attempt.

DeepSeek open-sources its coding agent LINK

• DeepSeek has open-sourced Super-Agent, an experimental coding agent published on GitHub as dsh-superagent under an MIT licence, which the lab calls a research prototype rather than an official product and aims at Anthropic's Claude Code.
• The agent works on a single GitHub repository and can monitor group chats on Feishu, opening a task in its own session when an approved person requests work and replying in the original thread with a pull request, while a web panel flags what needs a human.
• To avoid wasting a team's attention, it acts only when asked or when a team rule permits, though the public code differs from DeepSeek's internal version, some documented ideas aren't built yet, and users are told to review and adapt it before running.

Linux 7.4-rc1 merge window arrives LINK

• Linux 7.4-rc1 shipped today on schedule despite Linus Torvalds being on holiday, closing a merge window packed with AMD Ryzen Zen 6 support, AMDGPU HDMI 2.1 enabled by default, Intel IPU8 for Nova Lake, and faster kernel builds.
• This is the largest merge window ever by commit count, nearly 20,000 against the usual ~15,000, which Torvalds attributes partly to AI-driven work shifting into normal development and partly to a full third of the diff being AMD GPU register dumps.
• Expected to release around December 20 if it stays on schedule, Linux 7.4 is the likely pick for this year's LTS kernel, though Torvalds notes the -rc cycle will reveal whether the AI patch influx holds up.

LocalStack adds local Azure testing LINK

• LocalStack has released a free, self-serve public preview of its Azure emulator, running services locally and accepting existing ARM, Bicep, Terraform, Pulumi, and Azure CLI definitions without rewrites, with general availability targeted for early 2027.
• Coverage is partial and documented per provider, with Storage Accounts and App Service each implementing only a fraction of their operations, and many accepted settings do nothing: HTTPS-only serves HTTP, VNet integration is stored without network effect, and bearer tokens pass unless RBAC enforcement is on.
• Catches to watch: the emulator doesn't route traffic or enforce security rules, Windows-plan apps run on Linux, Oryx builds fail on Apple Silicon, App Service state lives in memory and is lost on stop, and App Inspector, snapshots, and chaos APIs aren't in preview.

Why we ditched DuckDB for our warehouse LINK

• PostHog has abandoned DuckDB for its managed warehouse after three months of re-architecting, concluding that the single-node engine couldn't safely handle customers' increasingly large and complex cloud workloads at a predictable price.
• DuckLake's portable catalog schema blocked engine-specific tuning, making single-table writes take over three minutes because each commit re-read catalog stats; over 99% of those stat rows referenced deleted tables, and purging them sped writes 30-50x before they regrew.
• PostHog dogfooded a 2 PB DuckLake holding roughly 1 trillion rows, but fork-only patches and the pain of scaling direct pgwire connections on port 5432 pushed it to rebuild on its own Trino flavor with a new table format, HogLake.

Tutorials & Advice

> Thinking in Traits: Rust reuses code by describing what types can do through capability-based traits rather than inheritance hierarchies, so independent pieces connect only through shared contracts.

> Lakebase branching for CI/CD: lessons from 100 PR databases: Give each pull request its own instant, copy-on-write branch of production data so tests run against real data in seconds instead of 20-minute database restores.

> Two Circles, One Connection, and border-shape: Use the new CSS border-shape property to draw curved connecting lines that keep their borders, which clip-path erases, between two draggable circles.

Other news you might like

• Grafana 13.3 release: manage variables across dashboards, share findings with Notebooks, a UI refresh, and more LINK

• Cloudflare services for private origins is now generally available LINK

• dbreunig/nesformer LINK

• HackerTube LINK

• CUDA Oxide 0.3 Vastly Improves Rust Support For NVIDIA CUDA Programming LINK

• Simplifying runtime caching on Vercel LINK

• Beating the compiler at signed 32-bit division LINK

• Backporting security fixes at scale with Codex LINK

Trending tools

HackerBooks: aggregates ~16,500 books mentioned by at least two HN commenters since 2007, linking each book back to the original comments that referenced it. LINK

Die With Me: a Mac app that tracks Claude and Codex usage, compares it with friends, and opens a shared chatroom during AI cooldown periods. LINK

Switch: open-source workspace connecting people and AI agents across Slack, Teams, Discord, and Mattermost, preserving context during handoffs without vendor lock-in. LINK

Kilo Code: open-source mobile AI coding agent that notifies you when sessions need input, letting you dictate prompts and attach files remotely. LINK

Tadata: an AI Slack assistant that connects your tools and sends daily briefings on meetings, contacts, and relevant news before work starts. LINK

Kombai Gallery: browse hundreds of categories of web and mobile UI designs for inspiration, then remix them or hand them to your coding agent. LINK

Trending papers & reports

Prompt-tuning optimizers get most of their gains from building on the best surviving version rather than from showing the model richer error summaries, meaning memory of past fixes beats fancier feedback across 1,750 tested runs. LINK

Code reasoning tests show that even a model answering whether two programs behave the same with ~95% accuracy pinpoints which edit changed behavior just ~54% of the time, revealing shallow understanding standard tests miss. LINK

Ethereum cost-cutting tactics are mapped from 400 real developer code changes into 8 broad categories and 36 specific methods, showing how teams actually lower the per-operation fees that make blockchain apps economically viable. LINK

AIFaultBench compiles 770 real-world bugs from 105 AI software projects, reproducing ~85% under documented conditions, giving teams a reliable testbed to benchmark debugging tools and study why AI systems actually break. LINK

Bug-reproduction automation rebuilds the exact conditions behind deep-learning software failures from incomplete reports, reproducing 85 of 106 real-world bugs and helping developers fix problems ~57% faster. LINK

See you tomorrow for a new dose of ☕️ Devshot!