PR Lens

Why it's ranked here

The engineering choices are specific and written down. Model output is checked against a strict schema and gets exactly one correction round. Commit hashes, line counts and the pull request number come from git and overwrite whatever the model claimed. Every provider is called at temperature zero so the same diff tends to draw the same picture, and the pictures carry no script. It is MIT licensed, runs on your own model key with no PR Lens server in the analysis path, and reaches GitHub, GitLab and Bitbucket. The open question is how accurate the diagrams are, which depends on the model you bring.

What's good

Key handling is careful. The model key is read only from the environment, never a flag, so it stays out of shell history and CI logs, and the GitHub Action passes it the same way. Sign-in files are written owner-only. The local file holding hosted-page write tokens is refused outright if git would commit it. The Action only ever edits a comment posted by one named account, so nobody can plant its marker on their own comment and have it rewritten. Provider support is three request shapes rather than a vendor list, so Ollama, DeepSeek or OpenRouter work by pointing at a URL, with no new release needed.

Tradeoffs

A diagram is only as true as the model that drew it: the schema check proves the document is well formed, not that the architecture it shows is right, and there is a single repair attempt before the run fails. The GitHub Action commits rendered pictures to a separate data branch in your repository and needs a full-history checkout. The agent skill runs whatever CLI is newest at call time rather than a pinned version. The hosted GitHub App and shared canvases are governed by a policy on the project site, not by code in the files examined, and the README says free for open source without stating private-repository terms.

How to use it well

Fits teams that review sizeable changes across several services and want reviewers oriented before reading code, or anyone onboarding into an unfamiliar codebase with a coding agent. Start with the CLI on a dry run to see what would be sent, then wire the Action with a pinned version and a model you trust on code. Treat the picture as a map to check against the diff, not a substitute for reading it. Keep shared canvas edit links private, since they carry write access, and rotate the token if one leaks. It does not lint, test or judge the change; it only draws it.

Observed

License

MIT (LICENSE, Coldtea AI)

Language

TypeScript pnpm monorepo: CLI, schema and renderer packages

Runtime

Node.js 20.11 or later, per package.json engines

Interfaces

CLI on npm, composite GitHub Action, coding-agent skill, hosted GitHub App

Model providers

Gemini by default, OpenAI, or any OpenAI-compatible endpoint, with your own key from the environment

Forge support

GitHub, GitLab CI/CD component and Bitbucket pipe, per README

Output

Self-contained light and dark SVGs with no script or external references

Test runner

Vitest declared in package.json