name: razor
description: "Minimize local diff size against merge base for uncommitted branch work using only behavior-preserving, scope-preserving architectural reductions."

Razor

Reduce local diff size with merge-base-anchored architectural planning and behavior-preserving execution only.

Core Iteration Directive

On every iteration, follow this exact directive:

"Deeply analyze this implementation across the entire PR (not file-by-file in isolation), and create a detailed plan called TRIM_CODEX.MD with the most impactful architectural ideas to dramatically reduce the total size of this PR, without reducing scope or behavior. Comments do not count."

Non-negotiables

• Always anchor to merge base: MB=$(git merge-base <base-ref> HEAD) (origin/latest by default).
• Use unified working-tree diff only: git diff "$MB".
• Include untracked code-file lines in insertion totals.
• Only propose/apply behavior-preserving, scope-preserving reductions.
• Never generate functionality-loss or scope-reduction options.
• Do not remove/trim comments or doc-comments. Comment lines do not count toward reduction targets.
• Keep optimization priority:
• Primary: reduce non-comment insertions.
• Secondary: reduce non-comment total churn.
• Optimize at PR architecture level first:
• Identify duplication and redundant abstractions spanning multiple files/components.
• Prefer ideas that remove whole cross-file patterns or flows, not local line edits.
• Treat file-local cleanups as fallback only after high-impact PR-wide ideas are exhausted.
• Read every line of every file in the diff on every iteration.
• Update TRIM_CODEX.MD every iteration before execution, then refresh after re-measurement.
• Run formatting on touched files and run targeted validation each pass.
• Never use destructive git commands.

Workflow

1. Establish base and merge base.
2. Measure unified diff (status, numstat, name-status, untracked files, untracked code lines).
3. Compute baseline metrics:
4. insertions
5. deletions
6. total churn
7. comment-line additions (reported separately, not targeted)
8. Full-diff read (mandatory):
9. For each file in git diff --name-status "$MB", read full current file contents.
10. For modified files, also review git diff "$MB" -- <file> hunks.
11. Deep architectural analysis:
12. Identify structural duplication and avoidable abstraction layers across the full PR surface.
13. Map cross-file data/control flow to find consolidation cuts that remove entire repeated paths.
14. Rank ideas by expected PR-wide churn reduction; prioritize multi-file high-leverage cuts.
15. Write/update TRIM_CODEX.MD with:
16. Goal and invariants.
17. Merge-base baseline and divergence.
18. Runtime/test hotspot ranking.
19. Architectural diagnosis of churn drivers.
20. Ranked behavior-preserving architectural ideas with explicit PR-wide scope (estimated insertion/churn savings, risk, validation notes).
21. Clear separation of PR-wide architectural cuts vs local cleanup (local cleanup only when no higher-impact architectural cuts remain).
22. Ordered execution sequence and validation gates.
23. Iteration history.
24. Apply viable behavior-preserving ideas.
25. Format touched files.
26. Run targeted validation.
27. Re-measure and refresh TRIM_CODEX.MD with before/after metrics and applied/skipped status.
28. Rerun fresh planning on the updated tree.

Output contract

Always return:
1. Base ref and merge-base SHA.
2. Before stats and after stats (insertions, deletions, churn).
3. Primary and secondary percent reductions.
4. Comment-line additions count (reported separately).
5. Ranked behavior-preserving ideas and applied changes.
6. Confirmation TRIM_CODEX.MD was refreshed this iteration.
7. Confirmation every diff file was fully read this iteration.
8. Explicit statement whether any additional behavior-preserving, scope-preserving reductions remain after fresh rerun.