File triage: classify files before expensive review

[haasonsaas]

Problem

DiffScope reviews every changed file with the same model and depth. CodeRabbit uses a cheap model to classify each file as NEEDS_REVIEW or APPROVED (cosmetic/formatting change) before spending expensive model tokens. This reduces cost and noise significantly.

How CodeRabbit Does It

1. A lightweight model classifies each changed file:
   • "Does this file contain logic/functionality changes, or is it purely cosmetic/formatting?"
   • Classification: NEEDS_REVIEW or APPROVED
2. Files classified as APPROVED skip the detailed review entirely
3. Rate limits: Free=150 files max, Pro=300 files max

How Qodo Does It

• Files sorted by main language first, then by token count descending
• patch_extension_skip_types = [".md", ".txt"] — certain file types auto-skipped
• Deletion-only hunks removed via omit_deletion_hunks() before the expensive call

Proposed Solution

Add a triage step before the main review:

Implementation

#[derive(Debug)]
enum TriageResult {
    NeedsReview,        // Logic/functionality change — full review
    Cosmetic,           // Formatting, whitespace, comments-only — skip
    ConfigChange,       // Config/env changes — lightweight review
    TestOnly,           // Test changes — review with different rules
    DeletionOnly,       // File deleted or lines-only removed — skip
    Generated,          // Auto-generated code — skip
}

async fn triage_file(
    diff: &UnifiedDiff,
    model: &ModelConfig,  // use weak/cheap model
) -> TriageResult {
    // Heuristic checks first (no LLM needed):
    // - All-whitespace changes → Cosmetic
    // - Deletion-only hunks → DeletionOnly
    // - Known generated file patterns → Generated
    // - Lock files, vendor dirs → Generated
    
    // LLM classification for ambiguous cases:
    // - "Is this a logic change or purely cosmetic?"
    // - Use cheap model (Haiku, GPT-4o-mini)
}

Heuristic-Only Triage (no LLM cost)

Many files can be triaged without any LLM call:

• Lock files (Cargo.lock, package-lock.json, yarn.lock)
• Generated code (.generated., _generated/)
• Binary files
• Deletion-only changes
• Whitespace-only changes
• Comment-only changes (parse for //, #, /* */ patterns)

Configuration

triage:
  enabled: true
  model: null  # null = use weak model, or specify explicitly
  skip_patterns:
    - "*.lock"
    - "*.generated.*"
    - "vendor/**"
  auto_approve:
    - deletion_only: true
    - whitespace_only: true
    - comment_only: true

Expected Impact

• 30-50% fewer files sent to the expensive review model
• Faster review times
• Less noise (no comments on formatting/lock file changes)
• Ties into multi-model routing (Multi-model routing: triage with cheap models, review with frontier #26) — triage uses cheap model

Priority

Medium — cost and noise reduction. Quick win that compounds with scale.

[haasonsaas]

Implementation Details (from codebase analysis)

Current State

• src/review/pipeline.rs skips binary/deleted files and files matching exclude_patterns
• Every non-excluded file gets the same expensive model — no classification step
• CodeRabbit uses a cheap model to classify NEEDS_REVIEW vs APPROVED (cosmetic)

Files to Modify

• src/review/pipeline.rs — add triage step before the per-file review loop (around line 259 where per-file processing starts)
• New: src/review/triage.rs — triage logic:
  • Heuristic (free): lock files, whitespace-only diffs, deletion-only hunks, comment-only changes → auto-skip
  • Optional LLM (cheap model): "Is this a logic change or cosmetic?" → NEEDS_REVIEW / APPROVED

Implementation

Start with heuristic-only triage (no LLM cost):

1. Lock files (Cargo.lock, package-lock.json, yarn.lock) → skip
2. Deletion-only changes → skip
3. Whitespace-only changes → skip
4. Files matching generated patterns (*.generated.*, vendor/**) → skip

Estimated Effort

Small — heuristic-only triage requires no LLM calls.

[haasonsaas]

Shipped in #44: triage_skip_deletion_only config, TriageOptions + SkipDeletionOnly, pipeline uses it in prepare_diff_analysis. Optional skip for deletion-only diffs before expensive review.