Research: Port PR-Agent review taxonomy and prompt engineering

[haasonsaas]

Context

PR-Agent (10.8k stars, Python, Apache 2.0) is the leading OSS AI-powered code review tool with structured diff analysis, confidence scoring, and multi-model support.

What to yoink

• Review taxonomy — categorized findings (security, performance, style, bugs, best practices) with severity levels
• Structured diff analysis prompts — battle-tested prompt engineering for code review across languages
• Incremental review — only review changed hunks, not entire files, reducing token usage and noise
• Confidence scoring — calibrated confidence per finding to reduce false positive fatigue
• /improve command pattern — suggest improvements with auto-applicable patches (suggest + apply)
• Auto-describe — generate PR descriptions from diff analysis

Approach

Diffscope's Rust engine is faster than PR-Agent's Python — the value is in the prompt engineering and UX patterns, not the runtime:

1. Study their prompt templates for each review category
2. Port the review taxonomy and severity classification
3. Borrow confidence scoring calibration approach
4. Adapt the /improve suggest-and-apply pattern for diffscope's output format

Also study reviewdog (9.2k stars, Go, MIT) for composable linter integration patterns — it integrates any analysis tool into PR review comments.

References

• https://github.com/qodo-ai/pr-agent
• https://github.com/reviewdog/reviewdog

Priority

Tier 2 — Prompt engineering and review UX patterns