Requesting Code Review
Dispatch ed3d-plan-and-execute:code-reviewer subagent to catch issues before they cascade.
Core principle: Review early, review often. Fix ALL issues before proceeding.
Session Isolation
If the calling context provides a SCRATCHPAD_DIR, pass it to code-reviewer.
This prevents collisions when multiple planning/execution sessions run in parallel. The SCRATCHPAD_DIR is a namespaced temp directory (e.g., /tmp/plan-2025-01-24-feature-a7f3b2/) that the code-reviewer uses for any scratch files.
When to Request Review
Mandatory:
- After each task in plan execution
- After completing major feature
- Before merge to main
Optional but valuable:
- When stuck (fresh perspective)
- Before refactoring (baseline check)
- After fixing complex bug
The Review Loop
The review process is a loop: review → fix → re-review → until zero issues.
┌──────────────────────────────────────────────────┐
│ │
│ Dispatch code-reviewer │
│ │ │
│ ▼ │
│ Issues found? ──No──► Done (proceed) │
│ │ │
│ Yes │
│ │ │
│ ▼ │
│ Dispatch bug-fixer │
│ │ │
│ ▼ │
│ Re-review with prior issues ◄──────────────────┘
│
└──────────────────────────────────────────────────┘Exit condition: Zero issues, or issues accepted per your workflow's policy.
Step 1: Initial Review
Get git SHAs:
BASE_SHA=$(git rev-parse HEAD~1) # or commit before task
HEAD_SHA=$(git rev-parse HEAD)Dispatch code-reviewer subagent:
<invoke name="Task">
<parameter name="subagent_type">ed3d-plan-and-execute:code-reviewer</parameter>
<parameter name="description">Reviewing [what was implemented]</parameter>
<parameter name="prompt">
Use template at requesting-code-review/code-reviewer.md
WHAT_WAS_IMPLEMENTED: [summary of implementation]
PLAN_OR_REQUIREMENTS: [task/requirements reference]
BASE_SHA: [commit before work]
HEAD_SHA: [current commit]
DESCRIPTION: [brief summary]
SCRATCHPAD_DIR: [session-isolated temp dir, or omit if not applicable]
</parameter>
</invoke>Code reviewer returns: Strengths, Issues (Critical/Important/Minor), Assessment
Step 2: Handle Reviewer Response
If Zero Issues
All categories empty → proceed to next task.
If Any Issues Found
Regardless of category (Critical, Important, or Minor), dispatch bug-fixer:
<invoke name="Task">
<parameter name="subagent_type">ed3d-plan-and-execute:task-bug-fixer</parameter>
<parameter name="description">Fixing review issues</parameter>
<parameter name="prompt">
Fix issues from code review.
Code reviewer found these issues:
[list all issues - Critical, Important, and Minor]
Your job is to:
1. Understand root cause of each issue
2. Apply fixes systematically (Critical → Important → Minor)
3. Verify with tests/build/lint
4. Commit your fixes
5. Report back with evidence
Work from: [directory]
Fix ALL issues — including every Minor issue. The goal is ZERO issues on re-review.
Minor issues are not optional. Do not skip them.
</parameter>
</invoke>After fixes, proceed to Step 3.
Step 3: Re-Review After Fixes
CRITICAL: Track prior issues across review cycles.
<invoke name="Task">
<parameter name="subagent_type">ed3d-plan-and-execute:code-reviewer</parameter>
<parameter name="description">Re-reviewing after fixes (cycle N)</parameter>
<parameter name="prompt">
Use template at requesting-code-review/code-reviewer.md
WHAT_WAS_IMPLEMENTED: [from bug-fixer's report]
PLAN_OR_REQUIREMENTS: [original task/requirements]
BASE_SHA: [commit before this fix cycle]
HEAD_SHA: [current commit after fixes]
DESCRIPTION: Re-review after bug fixes (review cycle N)
SCRATCHPAD_DIR: [session-isolated temp dir, or omit if not applicable]
PRIOR_ISSUES_TO_VERIFY_FIXED:
[list all outstanding issues from previous reviews]
Verify:
1. Each prior issue listed above is actually resolved
2. No regressions introduced by the fixes
3. Any new issues in the changed code
Report which prior issues are now fixed and which (if any) remain.
</parameter>
</invoke>Tracking prior issues:
- When re-reviewer explicitly confirms fixed → remove from list
- When re-reviewer doesn't mention an issue → keep on list (silence ≠ fixed)
- When re-reviewer finds new issues → add to list
Loop back to Step 2 if any issues remain.
Handling Failures
Operational Errors
If reviewer reports operational errors (can't run tests, missing scripts):
- STOP - do not continue
- Report to human
- When told to continue, re-execute same review
Timeouts / Empty Response
Usually means context limits. Retry with focused scope:
First retry: Narrow to changed files only:
FOCUSED REVIEW - Context was too large.
Review ONLY the diff between BASE_SHA and HEAD_SHA.
Focus on: [list only files actually modified]
Skip: broad architectural analysis, unchanged files, tangential concerns.
WHAT_WAS_IMPLEMENTED: [summary]
PLAN_OR_REQUIREMENTS: [reference]
BASE_SHA: [sha]
HEAD_SHA: [sha]Second retry: Split into multiple smaller reviews (one per file or logical group).
Third failure: Stop and ask human for help.
Quick Reference
| Situation | Action |
|---|---|
| Zero issues | Proceed |
| Any issues | Fix, re-review (or accept per workflow) |
| Operational error | Stop, report, wait |
| Timeout | Retry with focused scope |
| 3 failed retries | Ask human |
Red Flags
Never:
- Skip review because "it's simple"
- Proceed with ANY unfixed issues (Critical, Important, OR Minor)
- Argue with valid technical feedback without evidence
- Rationalize skipping Minor issues ("they're just style", "we can fix later")
Minor issues are NOT optional. The code reviewer flagged them for a reason. Fix all of them. "Minor" means lower severity, not "ignorable."
If reviewer wrong:
- Push back with technical reasoning
- Show code/tests that prove it works
- Request clarification on unclear feedback
Integration
Called by:
- executing-an-implementation-plan (after each task)
- finishing-a-development-branch (final review)
- Ad-hoc when you need a review
Template location: requesting-code-review/code-reviewer.md