interactive-diff-review

Interactive Diff Review

Walk the user through PR changes: present an overview, step through files one at a time, collect decisions on findings, show a summary, and post to GitHub when ready.

File I/O: Use Bash with heredocs or >> for session/temp files — not Write, Read, or Edit tools.

Inputs

• pr_number: the PR number
• session_file: path to the session file (contains Why, Findings, etc.)
• repo: owner/repo

Read the session file to extract:

• Why: summary from ## Why
• threshold: the threshold: <N> line at the top
• files: non-skipped files from ## Findings (ordered: modified first, then new, then deleted)
• findings: map of file -> [{ line, severity, confidence, source, description, code, in_diff }] from entries where skip: false under ## Findings. The code and in_diff fields are the line-text anchor used by classify-and-verify.py at post time; interactive-diff-review surfaces them in Phase 3 to flag uncertain anchors before the user queues a comment.
• below_threshold: findings from ## Below Threshold — not shown during the walkthrough, but the orchestrator may reference them in reporting
• skipped: files with skip: true

Phase 1: Overview

Show:

1. Why: The summary from the session file
2. Changes: A markdown table:

| File | What changed |
|------|--------------|
| src/resolvers/foo.ts | Simplified resolver to delegate to shared method |
| packages/lib/client.ts | New static method with extracted logic |

1. Skipped files: List with reasons. If none: "None — all files contain meaningful changes."
2. Findings summary: Count and severity breakdown ONLY: "Found N issues across M files (X high, Y medium, Z low)."

Do NOT list individual findings here. They are presented one at a time during the walkthrough.

Then ask: "Does this match your understanding of the PR? Any corrections before we dive into the file-by-file review?"

If corrections: append to ## User Context in session file. Re-evaluate findings that the correction might invalidate.