The writer describes using ordinary ChatGPT conversations as a practical software-development agent pool, despite not using Codex. Custom MCP servers on a Windows computer and a VPS exposed file access, code editing, terminal commands, tests, and computer-use tools. OpenAI’s Secure MCP Tunnel connected those local capabilities to ChatGPT without requiring separate inference spending or custom cloud infrastructure. The writer also modified open-source tools so screenshots could appear directly in chats, tool results were more reliable, and a dashboard could monitor the tunnel, system resources, and tool activity.
For larger projects, a stronger model planned the work and divided it into scoped GitHub issues with requirements and acceptance criteria. Multiple separate chats then tackled different issues in parallel using shared access to the same tools, while premium-model usage was reserved for planning and difficult reviews. The approach worked well for clearly defined tasks, though agents sometimes needed coordination or help from a stronger model. OpenAI’s announced changes will end the effectively unlimited chats the writer relied on and combine chat usage with Codex limits. The writer says the setup used features and tools OpenAI provided, not rule-breaking, and does not seriously believe it caused the policy change. Still, unlimited access was the main reason for maintaining an OpenAI subscription; the writer plans to switch to Claude, preferring its Opus model and finding its subscription more useful.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
Bill Joy created the first BSD distribution in 1977, bundling Berkeley’s Pascal system and the ex editor for use with existing Unix systems. He later expanded it into 2BSD, adding vi, the C shell, and other tools as BSD spread among universities. (fosscomics.com)
AI can handle more implementation and assist with design and testing, but engineers remain responsible for choosing what to build and deciding when it is correct. Durable, reviewable decisions and small, verifiable increments help keep AI-assisted engineering reliable. (picon.dev)
Proton’s AI manifesto sets out how the company uses AI in software engineering while keeping human developers responsible for decisions and code quality. It frames AI as an aid to engineering, not a replacement for expertise. (proton.me)
Linux 7.4’s merge window set a kernel record with nearly 20,000 commits, far above the typical 13,000-14,000 in recent years. The increase reflects growing development activity, including AI-generated bug reports and patches that have overwhelmed maintainers. (phoronix.com)
NixOS can idle at about 96 MB of RAM and complete updates and rebuilds within 512 MB after tuning kernel options and disabling unneeded services. The guide details the changes and trade-offs for running lightweight VPS deployments. (lowendspirit.com)
OmniDiff is an open-source Git difftool written in Rust, designed for fast, accurate, content-aware comparisons. It uses syntax-aware structural diffing to make code changes easier to review. (github.com)
Today's best Hacker News stories, summarized and screenshotted, one email a day.