Comprehensive PR Review

Run a comprehensive pull request review by spawning general-purpose subagents with inlined reference instructions. Each subagent gets a fresh context window with its specialized review instructions, the diff, and project context.

Review Aspects (optional): "$ARGUMENTS"

1. Determine Review Scope

• Check git status to identify changed files
• Parse arguments to see if user requested specific review aspects
• Default: Run all applicable reviews

2. Available Review Aspects

Aspect	Reference File	Purpose
code	code-reviewer-instructions.md	General code review for project guidelines
tests	pr-test-analyzer-instructions.md	Test coverage quality and completeness
comments	comment-analyzer-instructions.md	Code comment accuracy and maintainability
errors	failure-finder-instructions.md	Silent failures and error handling
types	type-design-analyzer-instructions.md	Type design and invariants
simplify	code-simplifier-instructions.md	Code clarity and maintainability
all	(all applicable)	Run all reviews (default)

3. Identify Changed Files

• Run git diff --name-only to see modified files
• Check if PR already exists: gh pr view
• Identify file types and what reviews apply

Error handling:

• If git diff fails (not a git repo): Report error clearly and stop
• If gh pr view fails with "no PR found": Expected for pre-PR reviews, continue with git diff
• If gh pr view fails with auth error: Note that GitHub CLI authentication is needed
• If no changed files found: Report "No changes detected" and stop

4. Determine Applicable Reviews

Based on changes:

• Always applicable: code (general quality)
• If test files changed: tests
• If comments/docs added: comments
• If error handling changed: errors
• If types added/modified: types
• After passing review: simplify (polish and refine)

5. Read Reference Instructions

Read each applicable reference file into a variable for inlining into subagent prompts:

code_instructions      = Read("./references/code-reviewer-instructions.md")
test_instructions      = Read("./references/pr-test-analyzer-instructions.md")
comment_instructions   = Read("./references/comment-analyzer-instructions.md")
errors_instructions    = Read("./references/failure-finder-instructions.md")
types_instructions     = Read("./references/type-design-analyzer-instructions.md")
simplify_instructions  = Read("./references/code-simplifier-instructions.md")

Only read files for applicable review aspects. Also read:

• git diff output into diff_content
• CLAUDE.md (if exists) into project_context

6. Resolve Model Profile

MODEL_PROFILE=$(node scripts/kata-lib.cjs read-config "model_profile" "balanced")
WORKTREE_ENABLED=$(node scripts/kata-lib.cjs read-config "worktree.enabled" "false")

Default to "balanced" if not set.

Model lookup table:

Agent	quality	balanced	budget
code-reviewer	opus	sonnet	sonnet
test-analyzer	sonnet	sonnet	haiku
comment-analyzer	sonnet	sonnet	haiku
failure-hunter	sonnet	sonnet	haiku
type-analyzer	sonnet	sonnet	haiku
code-simplifier	sonnet	sonnet	haiku

7. Launch Review Agents

Spawn general-purpose subagents via parallel Task calls. Each agent receives its reference instructions inlined as <agent-instructions>, the diff, and project context.

Task call pattern:

Task(
  prompt="<agent-instructions>\n{instructions_content}\n</agent-instructions>\n\nReview the following changes:\n\n<diff>\n{diff_content}\n</diff>\n\n<project-context>\n{project_context}\n</project-context>",
  subagent_type="general-purpose",
  model="{resolved_model}",
  description="PR review: {aspect}"
)

Example parallel launch (3 agents):

Task(prompt="<agent-instructions>\n{code_instructions}\n</agent-instructions>\n\nReview these changes:\n<diff>\n{diff_content}\n</diff>\n<project-context>\n{project_context}\n</project-context>", subagent_type="general-purpose", model="{code_model}", description="PR review: code")
Task(prompt="<agent-instructions>\n{test_instructions}\n</agent-instructions>\n\nReview these changes:\n<diff>\n{diff_content}\n</diff>\n<project-context>\n{project_context}\n</project-context>", subagent_type="general-purpose", model="{test_model}", description="PR review: tests")
Task(prompt="<agent-instructions>\n{errors_instructions}\n</agent-instructions>\n\nReview these changes:\n<diff>\n{diff_content}\n</diff>\n<project-context>\n{project_context}\n</project-context>", subagent_type="general-purpose", model="{errors_model}", description="PR review: errors")

All run in parallel. Task tool blocks until all complete.

Agent failure handling:

• If agent completes: Include results in aggregation
• If agent times out: Report "[aspect] review timed out - consider running independently"
• If agent fails: Report "[aspect] review failed: [error reason]"
• If one agent fails, STILL proceed with remaining agents
• Never silently skip a failed agent - always report its status

8. Aggregate Results

After agents complete, summarize:

• Critical Issues (must fix before merge)
• Important Issues (should fix)
• Suggestions (nice to have)
• Positive Observations (what's good)

Edge cases:

• If no issues found: "All Checks Passed" summary
• If agents conflict: Note the disagreement and let user decide
• If agent output malformed: Note "[aspect] output could not be parsed"
• Always include count of agents completed vs failed

9. Provide Action Plan

Organize findings:

# PR Review Summary

## Critical Issues (X found)

- [aspect]: Issue description [file:line]

## Important Issues (X found)

- [aspect]: Issue description [file:line]

## Suggestions (X found)

- [aspect]: Suggestion [file:line]

## Strengths

- What's well-done in this PR

## Recommended Action

1. Fix critical and important issues
2. Consider suggestions
3. Re-run review after fixes

10. Handle Review Findings

Route based on review results:

Findings	Route
Critical issues found	Route A (must address)
Important issues, no critical	Route B (should address)
Suggestions only	Route C (optional)
No issues	Route D (clean) → step 11

---

Route A: Critical issues found

Use AskUserQuestion:

• header: "Critical Issues"
• question: "{N} critical issues found. How do you want to handle them?"
• options:
  • "Fix all issues" — fix critical, important, and suggestions
  • "Fix critical only" — fix critical, add rest to backlog
  • "Add all to backlog" — create issues for everything, fix nothing now
  • "Skip" — continue without addressing

Route B: Important issues, no critical

Use AskUserQuestion:

• header: "Review Findings"
• question: "{N} important issues found. How do you want to handle them?"
• options:
  • "Fix all issues" — fix important and suggestions
  • "Fix important only" — fix important, add suggestions to backlog
  • "Add to backlog" — create issues, fix nothing now
  • "Skip" — continue without addressing

Route C: Suggestions only

Use AskUserQuestion:

• header: "Suggestions"
• question: "{N} suggestions found. Address them?"
• options:
  • "Fix suggestions" — apply suggested improvements
  • "Add to backlog" — create issues for later
  • "Skip" — continue without addressing

---

Handling "Fix" paths:

1. Apply fixes to the identified files
2. Stage and commit: fix: address PR review findings
3. If PR exists, push to branch

Handling "Add to backlog":

1. Create GitHub issues for each finding via gh issue create
2. Store TODOS_CREATED count for output

After handling (or Route D): Proceed to step 11.

11. Offer Next Actions

Resolve phase context for next-up routing:

# Current phase from STATE.md
CURRENT_PHASE=$(grep -oP 'Current Phase:\s*\K\d+' .planning/STATE.md 2>/dev/null || echo "")
NEXT_PHASE=$((CURRENT_PHASE + 1))

# Check if more phases remain
TOTAL_PHASES=$(grep -c '^## Phase' .planning/ROADMAP.md 2>/dev/null || echo "0")

{If PR exists from step 3:}

Use AskUserQuestion:

• header: "Next Step"
• question: "Review complete. What next?"
• options:
  • "Merge PR" — merge and clean up branch
  • "Verify work" — conversational acceptance testing
  • "Done" — stop here

{If no PR exists:}

Use AskUserQuestion:

• header: "Next Step"
• question: "Review complete. What next?"
• options:
  • "Verify work" — conversational acceptance testing
  • "Done" — stop here

---

If user chose "Merge PR":

gh pr merge "$PR_NUMBER" --merge

Then update local state:

if [ "$WORKTREE_ENABLED" = "true" ]; then
  # Bare repo layout: update main/ worktree, reset workspace/ to workspace-base
  git -C main pull
  bash "scripts/manage-worktree.sh" cleanup-phase workspace "$PHASE_BRANCH"
else
  git checkout main && git pull
fi

{If TODOS_CREATED: Backlog: {N} issues created from review findings}

Show: PR: #{pr_number} — merged ✓

Then show post-merge next up:

───────────────────────────────────────────────────────────────

▶ Next Up

{If NEXT_PHASE <= TOTAL_PHASES:}

Phase {Z+1}: {Name} — {Goal from ROADMAP.md}

/kata-discuss-phase {Z+1} — gather context and clarify approach

/clear first → fresh context window

───────────────────────────────────────────────────────────────

Also available:

• /kata-plan-phase {Z+1} — skip discussion, plan directly
• /kata-execute-phase {Z+1} — skip to execution (if already planned)

{If last phase (NEXT_PHASE > TOTAL_PHASES):}

Audit milestone — verify requirements, cross-phase integration

/kata-audit-milestone

/clear first → fresh context window

───────────────────────────────────────────────────────────────

Also available:

• /kata-complete-milestone — skip audit, archive directly

{If no phase context:}

Walk through deliverables — conversational acceptance testing

/kata-verify-work

/clear first → fresh context window

───────────────────────────────────────────────────────────────

---

If user chose "Verify work":

{If TODOS_CREATED: Backlog: {N} issues created from review findings}

───────────────────────────────────────────────────────────────

▶ Next Up

Walk through deliverables — conversational acceptance testing

/kata-verify-work {Z}

/clear first → fresh context window

───────────────────────────────────────────────────────────────

---

If user chose "Done": Stop.

{If TODOS_CREATED: Backlog: {N} issues created from review findings}