Show HN: I made a heatmap diff viewer for code reviews (0github.com)

A developer released an open-source "heatmap diff viewer" that color-codes every changed line and token in a GitHub pull request by how much human attention it likely needs. To try it you replace github.com with 0github.com on any PR URL; the system clones the repo into a VM, spins up a gpt-5-codex model for each diff, asks the model to produce a JSON score map, and renders that as a per-line/token heatmap. The goal isn’t binary “is this a bug?” classification but nuanced prioritization — surfacing hard-coded secrets, odd crypto modes, tangled logic, and other patterns that merit a second look. This matters because it reframes automated review from issue detection to reviewer triage, potentially speeding code review and focusing scarce human attention where it most increases safety and correctness. Key technical implications: model-per-diff inference (latency and cost per PR), token-level scoring and JSON output, and a VM-based pipeline that processes repository code — which raises privacy and security tradeoffs for private repos and sensitive code. For the AI/ML community this is an interesting practical use of LLMs as attention heuristics, but it also highlights evaluation needs (false positives/negatives), potential bias in what the model flags, and the appeal of on-prem or fine-tuned models to avoid data leakage. The project is open source, enabling experimentation and safer deployments.