power-platform-developer-suite: Skill for Claude Code

Review

Code review with structural bias prevention. The reviewer gets NO implementation context — no plan, no task descriptions, no "what we were trying to do." They get ONLY: the diff, the file contents, the constitution, and relevant spec ACs. They read code and find bugs.

Why This Exists

Evidence from vscode-extension-mvp worktree:

• Biased review (reviewer had implementation context): found 1 issue
• Independent review (fresh session, no context): found 30 issues

Same code. 30x difference. Reviewer bias is measurable and severe.

Input

$ARGUMENTS = optional scope (e.g., src/PPDS.Cli/Tui/ to limit review). Default: review all staged/recent changes.

Process

Step 1: Gather Review Material

# Ensure local main ref is current (worktrees can be stale)
git fetch origin main --quiet

git diff main...HEAD --stat
git diff main...HEAD

If $ARGUMENTS specifies a scope, filter the diff to those paths only.

Step 2: Load Spec Context (NO implementation context)

• Read specs/CONSTITUTION.md
• Map changed files to specs by grepping all specs/*.md files for **Code:** frontmatter lines. Match changed file paths against code path prefixes to find governing specs.
• Read each relevant spec — extract ONLY the ## Acceptance Criteria section
• Do NOT read any plan files, task descriptions, or implementation notes

Step 2b: Load QA Findings for Dedup

If .qa-findings.json exists at the repo root (written by /qa), include it in each reviewer subagent prompt as an "Already found by QA" appendix. QA findings are permitted reviewer metadata — the sanctioned exception to Step 3's isolation rule: they reveal prior findings, not implementation intent, and reviewers may use them ONLY for deduplication. Reviewers should NOT re-report QA findings that were fixed (fixed: true) unless the fix introduced a new problem. If the file is absent, skip this step.

Step 3: Dispatch Impartial Reviewer

For large diffs (>10 files), use per-file chunking instead of a single subagent: