auto-draft-review

Auto Draft Review

Post analyzer findings as a pending GitHub review so the user can finalize on github.com. No walkthrough, no per-finding prompts. The user's first line ("Running auto review — ...") was already announced by the orchestrator; do not re-announce.

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

Inputs

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

Step 1: Fetch the diff

gh pr diff <PR_NUMBER> --repo <REPO> > $HOME/.local/state/review-assistant/pr-<PR_NUMBER>-diff.txt

Read the session file separately for metadata (Why summary, review_sha, threshold, skipped files). Findings themselves flow through the classifier in the next step — do not re-parse them by hand.

Step 2: Classify and verify findings

Run the shared classifier against the ## Findings section:

python3 "${CLAUDE_PLUGIN_ROOT}/scripts/classify-and-verify.py" \
  --diff "$HOME/.local/state/review-assistant/pr-<PR_NUMBER>-diff.txt" \
  --session "<session_file>" \
  --section findings \
  > $HOME/.local/state/review-assistant/pr-<PR_NUMBER>-classified.json

The script emits { inline, fallback, suspect, stats } where each bucket means:

• inline — safe to post as an inline review comment. line and side have been verified (or re-anchored within ±3 lines) against the finding's recorded code text. Use these verbatim. Each entry's body is the finding's description as written by the analyzer (paragraph breaks preserved). Severity, confidence, and source live on source_finding for the review-body summary, not in the inline body — keep it that way; meta-prefixes like **high** (confidence 87, source: bug-scan) add noise to anyone reading the PR on github.com.
• fallback — the line is outside every hunk for that file, or the finding was recorded with in_diff: false. Put these in the review body, not in comments[].
• suspect — the line IS inside a hunk, but the finding's code anchor didn't match the diff line or any of its ±3 neighbors. Also goes in the review body, with an explicit "line anchor uncertain" flag so the reviewer knows not to trust the number.