Add diff size limits and built-in file exclusions

[matej]

Summary

Adds safeguards to prevent context overflow and improve signal-to-noise ratio by:

1. Limiting diff size to avoid exceeding Claude's context window
2. Auto-excluding generated, binary, and vendor files

Changes

New input: max-diff-lines

• Default: 5000 (keeps context under ~50% of 200k token limit)
• Set to 0 to disable the limit
• PRs exceeding the limit are skipped with a clear warning

Built-in file exclusions

Lock files:

• package-lock.json, yarn.lock, pnpm-lock.yaml
• Gemfile.lock, Pipfile.lock, poetry.lock, composer.lock
• Cargo.lock, go.sum, pubspec.lock, Podfile.lock

Generated files:

• *.min.js, *.min.css, *.bundle.js, *.chunk.js, *.map
• *.pb.go, *.pb.swift, *.generated.*, *.g.dart, *.freezed.dart

Binary files:

• Images: *.png, *.jpg, *.jpeg, *.gif, *.ico, *.webp, *.svg
• Fonts: *.woff, *.woff2, *.ttf, *.eot
• Archives: *.zip, *.tar.gz, *.jar
• Compiled: *.pyc, *.so, *.dylib, *.dll, *.exe

Vendor directories:

• node_modules, vendor, dist, build, .next
• __pycache__, .gradle, Pods, DerivedData

User-specified exclude-directories are combined with built-in exclusions.

Test plan

• Unit tests for built-in exclusions (14 new tests)
• Unit tests for diff size limit parsing
• Full test suite passes (187 tests)

🤖 Generated with Claude Code

[github-actions]

No issues found. Changes look good.

[neonspectra]

So, just a heads up as background:

• The way you are doing code reviews here is different from how I did them in my workflow.
• You cloned the whole repo and are using the default action, while in my first-pass implementation, I called the action contained in the upstream repo as a Marketplace Action step.
• This means that the issues I was experiencing with large diffs are going to be pretty fundamentally different from what's going on here.

That said, I do think there's still room for improvements. I like the use of exclusions, but I don't like ever truncating context. IMO, truncation is bad because if there is a situation that one needs to truncate, then that situation itself is probably blowing up claude's context window too much in the first place with the prompt design.

The better route here imo is to make sure Claude gets items passed as files and has the ability to read them using agentic mode to look for whatever Claude thinks is important. This way we should (almost) never encounter a scenario where we have diffs that are way too big because we'll be passing just files.

I stacked a PR to this PR with my suggestion for this: #19

[matej]

That said, I do think there's still room for improvements. I like the use of exclusions, but I don't like ever truncating context. IMO, truncation is bad because if there is a situation that one needs to truncate, then that situation itself is probably blowing up claude's context window too much in the first place with the prompt design.

We're not truncating. We're skipping code review completely if the diff gets to about ~50% of the context size (where model performance would start to degrade in a lot of cases). I think this is a good pragmatic start before we explore other enhancements.