Pr Swarm Grade

• Read ${REVIEW_DIR}/_state.json → extract pr_number, head_sha, pr_title, agents, compiled.
• If _state.json doesn't exist → STOP: "No review found for PR #${PR_NUMBER}. Run /pr-swarm first."
• If compiled is false → STOP: "Review for PR #${PR_NUMBER} hasn't been compiled yet."

1c. Read compiled report

• Read ${REVIEW_DIR}/compiled-report.md.
• Parse into a structured list of findings. Each finding has:
  • Number (sequential)
  • Category: Must Fix, Suggestions, or Nitpicks — from section header
  • Location: file:line reference
  • Title: short description
  • Agents: which review agents flagged it
  • Description: full finding text

1d. Get fix commits

HEAD_SHA=<head_sha from _state.json>
git log ${HEAD_SHA}...HEAD --oneline

• If no commits → findings default to MISS, but fixes comment can still yield SKIP-OK/SKIP-BAD.
• Store the commit list.

1e. Get the diff

git diff ${HEAD_SHA}...HEAD

• If PR is merged, check if HEAD_SHA is reachable. If not, try merge commit.
• Store full diff for cross-referencing.

1f. Find "Review Fixes Applied" comment

gh pr view ${PR_NUMBER} --json comments --jq '.comments[].body'

• Look for comment starting with ## Review Fixes Applied.
• Parse Fixed and Skipped lists.
• If no such comment → warn: "No 'Review Fixes Applied' comment found. Grading from diff only."

Phase 2 — Cross-Reference

For each finding from the compiled report, determine a verdict:

Verdict Criteria

Cross-Reference Process

For each finding:

1. Check the diff: Does it contain changes to the referenced file(s) and line range(s)?
2. Check the fixes comment: Is this finding in "Fixed" or "Skipped"?
3. Check for dishonesty: If fixes comment claims "Fixed" but diff doesn't support it → Dishonest Claim.
4. Check for regressions: If diff changes relevant code but introduces new problems → REGRESSION.
5. Assess fix quality (PASS and PARTIAL only): Critique the approach, assign quality tag:

Quality Tags (PASS and PARTIAL only)

Flags

• Dishonest Claim: Fixes comment says fixed, but diff contradicts.
• Proactive Fix: Something fixed that wasn't in findings (bonus credit).
• Undocumented Skip: Finding not in diff AND not in fixes comment.

Phase 3 — Grade

Per-Finding Scoring

Category Weights

Score Calculation

weighted_score = sum(finding_score * category_weight) for each finding
max_possible = sum(100 * category_weight) for each finding
raw_percentage = (weighted_score / max_possible) * 100

Modifiers (applied to final percentage)

Letter Grade Scale

Final score clamped to 0-100 after modifiers.

Phase 4 — Report Card

Write to ${REVIEW_DIR}/grade-report.md and present in conversation. Do NOT post to PR.

Output Format

## Grade Report — PR #${PR_NUMBER}: ${PR_TITLE}

**Overall: ${LETTER_GRADE} (${SCORE}/100)**

| Category     | Addressed | Total | Score |
|--------------|-----------|-------|-------|
| Must Fix     | X/Y       | Y     | Z%    |
| Suggestions  | X/Y       | Y     | Z%    |
| Nitpicks     | X/Y       | Y     | Z%    |

### Must Fix
1. **${VERDICT}** · ${QUALITY} — `${location}` — ${title}
   _What changed:_ ${factual description}
   _Feedback:_ ${opinionated quality assessment}

### Suggestions
...

### Nitpicks
...

### Flags
- **Dishonest claim**: Finding #${N} claimed fixed but diff shows ${what}.
- **Proactive fix**: ${description}.
- **Undocumented skip**: Finding #${N} not addressed and not mentioned.

(Omit Flags section if none.)

### Summary
${2-3 sentence assessment: quality, strengths, gaps, recommendation}

Formatting Rules

• Number findings matching compiled report numbering.
• Use exact verdict labels: PASS, PARTIAL, MISS, REGRESSION, SKIP-OK, SKIP-BAD.
• "Addressed" counts PASS + PARTIAL + SKIP-OK.
• Every PASS/PARTIAL has two lines:
  • _What changed:_ — factual, 1 sentence.
  • _Feedback:_ — opinionated, 1-2 sentences. Judgment, not description.
• Quality tags appear after verdict for PASS/PARTIAL only.
• _Feedback:_ must express judgment. Bad: "Used a typed setter map." Good: "Clean approach — eliminates unsafe cast without over-engineering."

Edge Cases

Hard Rules

• Read-only. Do not modify files, post comments, or push code.
• Grade every finding. No finding from compiled report should be skipped.
• Be honest. Grade what the diff shows, not what the fixes comment claims.
• Show your work. Each verdict must reference specific code changes (or lack thereof).