code-review-flow

Code Review Flow

Wrapper around superpowers:requesting-code-review that avoids permission prompts by using information already in context.

Usage

When user requests code review:

1. Check conversation context for commit SHAs (they're usually visible in recent git output)
2. If not in context, run separate git commands (not compound commands)
3. Invoke a general-purpose reviewer with the template

Getting Git SHAs Without Prompts

Option 1: Use context (BEST - no Bash needed)

• Base SHA: Look at gitStatus, recent git log output, or known commit reference
• Head SHA: Shown in your own git commit output or git log

Option 2: Run git commands separately (already allowed in most projects)

# Run these as SEPARATE Bash calls, not chained with &&
git merge-base origin/main HEAD  # actual branch point, immune to main advancing
git rev-parse HEAD

DON'T do this (triggers permission prompts):

# ❌ Compound command with variable assignment
BASE_SHA=$(git merge-base origin/main HEAD) && HEAD_SHA=$(git rev-parse HEAD) && echo "BASE=$BASE_SHA HEAD=$HEAD_SHA"

Invoking Code Reviewer

Once you have the SHAs, invoke the code-reviewer subagent:

Task tool with subagent_type: general-purpose

Prompt template:
# Code Review Agent

You are reviewing code changes for production readiness.

**Your task:**
1. Review [WHAT_WAS_IMPLEMENTED]
2. Compare against [PLAN_OR_REQUIREMENTS]
3. Check code quality, architecture, testing
4. Categorize issues by severity
5. Assess production readiness

## What Was Implemented

[DESCRIPTION - Brief summary of what was built]

## Requirements/Plan

[PLAN_REFERENCE - Link to issue, plan doc, or inline description of requirements]

## Git Range to Review

**Base:** [BASE_SHA]
**Head:** [HEAD_SHA]

```bash
git diff --stat [BASE_SHA]..[HEAD_SHA]
git diff [BASE_SHA]..[HEAD_SHA]

[... rest of code-reviewer template from superpowers:requesting-code-review/code-reviewer.md]

## Posting Review Results

**After review completes, check if PR exists:**

```bash
gh pr list --head $(git branch --show-current) --json number,url

If no PR exists:

• Display review results to user in conversation
• User can create PR later or continue with local fixes

If a PR exists, post findings as inline diff-line comments, batched into a single PR review. Each finding that carries a file:line anchors to the exact diff line so the author reads it right where the change is, instead of mapping a wall of text back to the code. This is the canonical inline-review protocol; /code-review-intense-flow and /request-review defer to it. gh pr comment / gh pr review --body can only post PR-level text — line-anchored comments are reachable only through the REST API via gh api.

Attribution: the review summary body (not each individual inline finding) must end with the attribution footer from docs/attribution.md, using the model version resolved at runtime.

Inline review protocol

1. Resolve the head SHA robustly — never assume local HEAD matches the PR head:

   gh pr view <n> --json headRefOid -q .headRefOid
   

2. Author review.json — PRIMARY: write it directly with your file-writing tool. You produce the JSON, so there is no shell quoting, heredoc delimiter, or apostrophe to get wrong; direct authoring removes the shell from JSON assembly entirely and is immune to every injection class below. Every string field — the summary body and each comments[].body — is JSON you write, so un-anchorable or attacker-influenced finding text is inert data, never shell input. Shape to write:

   {
     "commit_id": "<head-sha>",
     "event": "COMMENT",
     "body": "Automated review — inline findings below. <un-anchorable findings / overall assessment here>\n\n---\n🤖 Review by [Claude Code](https://claude.com/claude-code) · Opus 4.8",
     "comments": [
       { "path": "go/web/foo.go", "line": 42, "side": "RIGHT",
         "body": "**[Important]** This nil check can move above the loop." },
       { "path": "go/web/bar.go", "start_line": 10, "start_side": "RIGHT", "line": 14, "side": "RIGHT",
         "body": "**[Minor]** This block can be simplified." }
     ]
   }
   

   The head SHA still comes from gh pr view <n> --json headRefOid -q .headRefOid; write the file into a mktemp -d workdir, then POST it (step 3).

   FALLBACK for shell-only contexts without a file-writing tool: build it with jq. No finding or assessment text may transit a double-quoted shell word OR sit as a bare literal inside the single-quoted jq program — bash expands $(...), backticks, and $var inside double quotes before jq runs, and an apostrophe in a jq-program literal (e.g. "doesn't") breaks the single-quoted bash arg. This covers the summary body field EXACTLY as much as each comments[].body: write every body — summary included — to its own file with a single-quoted heredoc (<<'BODY' suppresses ALL shell expansion regardless of content), then pull them in raw with jq --rawfile, so body: $summary. This holds for every body without exception — not even a short fixed boilerplate summary may stay --arg/literal. The only --arg value is the head SHA (safe — it comes from gh pr view, not from finding text). Structural fields (path, line, side, event) are author-controlled literals inside the jq program. Pick a heredoc delimiter (e.g. BODY) that does not appear at column 0 inside any finding text; if it might, prefer the primary direct-authoring path (or choose an unlikely per-run delimiter).

   HEAD_SHA=$(gh pr view <n> --json headRefOid -q .headRefOid)
   
   # Private temp dir; auto-cleaned. Never a fixed /tmp path.
   WORKDIR="$(mktemp -d "${TMPDIR:-/tmp}/review.XXXXXX")"
   trap 'rm -rf "$WORKDIR"' EXIT
   
   # Write EVERY body — the summary AND each finding — with a SINGLE-QUOTED heredoc so backticks /
   # $(...) / $var / apostrophes in the text are written literally and never executed or broken.
   cat > "$WORKDIR/summary.md" <<'BODY'
   Automated review — inline findings below. <un-anchorable findings / overall assessment here>
   
   ---
   🤖 Review by [Claude Code](https://claude.com/claude-code) · Opus 4.8
   BODY
   cat > "$WORKDIR/b1.md" <<'BODY'
   **[Important]** This nil check can move above the loop.
   BODY
   cat > "$WORKDIR/b2.md" <<'BODY'
   **[Minor]** This block can be simplified.
   BODY
   
   # Assemble with jq. Only the SHA is an --arg; the summary and every finding body come in raw
   # via --rawfile — no body is ever a literal inside the jq program.
   jq -n \
     --arg commit "$HEAD_SHA" \
     --rawfile summary "$WORKDIR/summary.md" \
     --rawfile b1 "$WORKDIR/b1.md" \
     --rawfile b2 "$WORKDIR/b2.md" \
     '{
        commit_id: $commit,
        event: "COMMENT",
        body: $summary,
        comments: [
          { path: "go/web/foo.go", line: 42, side: "RIGHT", body: $b1 },
          { path: "go/web/bar.go", start_line: 10, start_side: "RIGHT", line: 14, side: "RIGHT", body: $b2 }
        ]
      }' > "$WORKDIR/review.json"
   

3. POST once — one review carrying every inline comment, so the author gets a single notification, not N:

   gh api repos/{owner}/{repo}/pulls/<n>/reviews --method POST --input "$WORKDIR/review.json"
   

4. event: "COMMENT" always — this preserves the "never self-approve" rule for code-review-flow and code-review-intense-flow. Solo-developer workflows don't allow self-approval.

Anchoring rule and fallback

• Each finding maps to a changed line inside a diff hunk (or the nearest changed line in the same hunk). Use side: "RIGHT" for the new version, LEFT for the deleted side; a multi-line span adds start_line + start_side.
• A finding that cannot be tied to a diff line (architectural, file-spanning, or about unchanged context) goes into the review body summary instead. Nothing is dropped — every finding lands either as an inline comment or in the summary.

Gotchas

1. Line must be inside the diff. Commenting on unchanged context outside a hunk returns 422. Map findings to changed lines; route the rest to the summary body.
2. SHA discipline. Always resolve headRefOid via gh pr view; a stale or local SHA returns 422.
3. Batch, don't spray. One pulls/<n>/reviews POST with a comments[] array — not a loop of single pulls/<n>/comments POSTs.
4. Never place finding text in a double-quoted shell word — and never as a bare literal inside the jq program. This rule covers the summary body field EXACTLY as much as each comments[].body. Bash expands $(...), backticks, and $var inside double quotes before jq runs, so attacker-controlled diff text quoted into a finding becomes command execution; and an apostrophe in a single-quoted-jq literal (e.g. the assessment saying "doesn't") breaks the bash arg. Keep every body — summary included — out of the shell: author review.json directly with your file-writing tool (primary), or in the shell fallback write each body to a file with a single-quoted heredoc (<<'BODY') and pull it in with jq --rawfile (body: $summary). This is absolute: there is no case where a summary or finding body uses --arg or a double-quoted shell word, not even a short fixed boilerplate body. Only the head SHA (--arg commit "$HEAD_SHA", from gh pr view) uses --arg.
5. JSON assembly. Prefer authoring review.json directly with your file-writing tool. In the shell fallback, build it with jq into a mktemp -d workdir (auto-cleaned via trap 'rm -rf "$WORKDIR"' EXIT), pulling the summary AND every finding body in via --rawfile (never a literal for either); don't inline a heredoc into jq and don't use a fixed temp path.
6. Replies vs. new threads. To follow up on an existing thread, POST to pulls/<n>/comments/{comment_id}/replies rather than opening a new one.

Fixing Review Issues

When the review finds issues, fix them automatically rather than just reporting:

1. Fix all Critical, Important, AND Minor issues found
2. Exception: If the diff is over 500 lines, fix Critical and Important issues in the branch but create GitHub issues for Minor ones so they don't get lost
3. If a Minor issue seems wrong or counterproductive, push back rather than blindly implementing — but default to fixing since it's less overhead than a follow-up issue
4. Commit fixes with a clear message referencing the review. Every such commit ends with a blank line then the Co-authored-by trailer from docs/attribution.md (display name = the model running at runtime), e.g. Co-authored-by: Claude Opus 4.8 <[email protected]>
5. Re-run the review cycle on the new commits (update HEAD_SHA and review again)
6. Repeat until the review passes clean

After Review Passes

Once the review finds no remaining issues:

1. Post the final clean review to the PR (if one exists)
2. Check if /monitor-ci slash command exists in the current project's available skills
3. If /monitor-ci exists, invoke it to monitor CI status
4. If /monitor-ci does not exist, fall back to /poll-ci (the generic gh-based CI poller) to monitor CI status
5. If neither command is available, inform the user the review is complete

Example

User: "request a code review using superpowers"

Step 1: Check context for SHAs
- Recent git commit showed: [fix-1065 d0e856b8]
- git log showed base: 4f940124

Step 2: Invoke code-reviewer
Task(general-purpose):
  WHAT_WAS_IMPLEMENTED: Tag sorting fix with case-insensitive handling
  PLAN_OR_REQUIREMENTS: Issue #1065 - sort tags by distance value
  BASE_SHA: 4f940124
  HEAD_SHA: d0e856b8
  DESCRIPTION: Added parseDistanceTag() and case-insensitive regex

Step 3: Review found 1 major issue (missing nil check) and 2 minor issues
- Diff is 180 lines (under 400) → fix all issues
- Commit fixes: [fix-1065 a1b2c3d4] (commit ends with the `Co-authored-by` trailer from `docs/attribution.md`)

Step 4: Re-run review with updated HEAD
Task(general-purpose):
  BASE_SHA: 4f940124
  HEAD_SHA: a1b2c3d4
  ... (same params, new HEAD)

Step 5: Review passes clean

Step 6: Check for PR
$ gh pr list --head fix-1065 --json number
[{"number": 123}]

Step 7: Post clean review to PR as a single review (inline anchored comments + summary body)
$ HEAD_SHA=$(gh pr view 123 --json headRefOid -q .headRefOid)
$ WORKDIR="$(mktemp -d "${TMPDIR:-/tmp}/review.XXXXXX")"; trap 'rm -rf "$WORKDIR"' EXIT
$ cat > "$WORKDIR/body.md" <<'BODY'
Automated review — passes clean, no remaining issues.

---
🤖 Review by [Claude Code](https://claude.com/claude-code) · Opus 4.8
BODY
$ jq -n --arg commit "$HEAD_SHA" --rawfile body "$WORKDIR/body.md" \
    '{commit_id: $commit, event: "COMMENT", body: $body, comments: []}' > "$WORKDIR/review.json"
$ gh api repos/{owner}/{repo}/pulls/123/reviews --method POST --input "$WORKDIR/review.json"
✓ Review posted to PR #123 (event: COMMENT — never self-approve)

Step 8: Check for /monitor-ci (pick the first that applies)
- /monitor-ci exists → invoke it
- else /poll-ci exists → fall back to /poll-ci
- else → inform user review is complete

Integration with Permissions

This skill works because:

• Bash(git rev-parse:*) is typically already allowed
• Separate commands don't need complex shell parsing
• Reading conversation context needs no permissions

Benefits

✅ No permission prompts during code review ✅ User can leave window while review runs ✅ Faster workflow - no blocking on permissions ✅ Uses information already in context when available ✅ Posts review to PR when one exists - keeps discussion centralized

When to Use

• After completing a task/feature
• Before merging to main
• After fixing complex bugs
• When stuck (get fresh perspective)

Related Skills

• superpowers:requesting-code-review - The skill this wraps
• superpowers:receiving-code-review - How to handle review feedback
• /code-review-intense-flow - Heavyweight variant that also fans out to specialized reviewers (security, frontend, seo, geo, playwright) based on the diff