review-loop

Review Loop

Run /deep-review repeatedly, fixing issues each round. Progressively kills bugs, improves code quality, and surfaces improvement opportunities.

Input Parsing

Parse arguments to extract:

• count (number): How many review rounds. Default: 3
• --aggressive: Fix almost everything. Ask user only for truly base-changing decisions (e.g., "should we switch frameworks?"). Use during prototyping, massive migration, or greenfield work
• --defensive (default): Handle results carefully. Fix only clear bugs and convention violations. Use when the project is live and stability matters
• --stay (default): Apply fixes directly to the current branch
• --as-pr: Create a branch + draft PR, then apply fixes there. Follow the /x-as-pr workflow
• Model flags (-haiku / --haiku, -so / --sonnet, -op / --opus): Forwarded verbatim to /deep-review (or the backend -review variant if a backend flag is also set). Sets the Claude model for subagent reviewers. Pick at most one. Default: -op (matches /deep-review's default).
• -co or --codex: Use /codex-review instead of /deep-review for the review process. Codex review uses OpenAI Codex CLI for faster, cheaper reviews
• -gco or --github-copilot: Use /gco-review (GitHub Copilot CLI) instead of /deep-review for the review process. Mutually exclusive with -co and -gcoc
• -gcoc or --github-copilot-cheap: Same as -gco but forces the free gpt-4.1 model (skips the Premium opus attempt). Uses /gcoc-review. Mutually exclusive with -co and -gco

Workflow

Step 1: Setup

Determine the working mode from parsed arguments.

If --as-pr:

1. Invoke /x-as-pr <description> to create a branch and draft PR
2. Record the PR number as PR_NUM
3. All subsequent work happens on the new branch

If --stay:

1. Work directly on the current branch

Step 2: Review Loop (repeat N times)

For each round (1 to N):

2a: Run review

If --codex is set, invoke /codex-review using the Skill tool. /codex-review silently falls back to Claude Code reviewers if codex is rate-limited — no special handling needed here. If --github-copilot is set, invoke /gco-review using the Skill tool. /gco-review silently falls back to Claude Code reviewers if Copilot is rate-limited — no special handling needed here. If --github-copilot-cheap is set, invoke /gcoc-review using the Skill tool. /gcoc-review silently falls back to Claude Code reviewers if Copilot is rate-limited — no special handling needed here. Otherwise, invoke /deep-review. Wait for all reviewers to complete.

2b: Categorize findings

Sort all findings into:

• Must-fix: Bugs, logic errors, broken functionality, convention violations, security issues
• Should-fix: Code quality, DRY violations, missing types, accessibility gaps
• Consider: Refactoring opportunities, architectural improvements, nice-to-have enhancements

2c: Apply fixes based on strategy

If --aggressive:

• Fix everything in must-fix and should-fix automatically
• Fix consider items too, UNLESS the change is truly base-changing (framework switch, major API redesign). In that case, ask the user

If --defensive:

• Fix must-fix items automatically
• Fix should-fix items only if clearly safe and low-risk
• Skip consider items entirely
• Report consider items to the user in the terminal for awareness

2d: Handle fix volume

• Small fixes (< 5 files, straightforward): Apply directly
• Large fixes (many files, complex refactoring): Use /x-wt-teams to spawn parallel agents. Group related fixes into topics

2e: Commit fixes

After applying fixes for this round:

1. Run typecheck / lint to verify
2. Commit with descriptive message: [scope] Fix review round N findings: <summary>

2f: Report round results

Tell the user what was found and fixed in this round. Be concise.

2g: Early exit

If a round finds 0 actionable issues, skip remaining rounds. Report "No issues found — stopping early."

Step 3: Finalize

If --as-pr:

1. Push changes to remote
2. Invoke /pr-revise to update the PR title and description
3. Report the PR URL

If --stay:

1. Report what was done across all rounds

Examples

Basic: 3 rounds, defensive, stay on branch

/review-loop

Aggressive migration review

/review-loop 5 --aggressive

Runs 5 rounds, fixes aggressively.

Careful review as a PR

/review-loop 3 --defensive --as-pr

Creates branch+PR, runs 3 defensive rounds, updates PR.

Codex-powered review

/review-loop 3 --codex

Uses /codex-review (OpenAI Codex CLI) instead of /deep-review for each round.

GitHub Copilot-powered review

/review-loop 3 --github-copilot

Uses /gco-review (GitHub Copilot CLI) instead of /deep-review for each round.

GitHub Copilot Cheap-powered review

/review-loop 3 --github-copilot-cheap

Uses /gcoc-review (GitHub Copilot CLI, forced to free gpt-4.1 model) instead of /deep-review for each round.

Quick single round

/review-loop 1 --aggressive

Important Notes

• Default review uses /deep-review with 3 Opus reviewers in PR mode (or 6 in full project mode)
• With --codex, review uses /codex-review (OpenAI Codex CLI) instead — faster and cheaper. If codex is rate-limited, /codex-review silently falls back to Claude Code reviewers (no workflow interruption)
• With --github-copilot, review uses /gco-review (GitHub Copilot CLI) instead. If Copilot is rate-limited, /gco-review silently falls back to Claude Code reviewers (no workflow interruption)
• With --github-copilot-cheap, review uses /gcoc-review (GitHub Copilot CLI, forced gpt-4.1) instead. If Copilot is rate-limited, /gcoc-review silently falls back to Claude Code reviewers (no workflow interruption)
• Later rounds often find fewer issues as earlier rounds fixed the low-hanging fruit
• The --aggressive vs --defensive distinction controls the threshold for automatic fixes, not the review depth
• Always run typecheck between rounds to catch regressions from fixes
• If a round finds 0 actionable issues, skip remaining rounds early
• Do NOT create GitHub issues for review findings — findings are reported in the terminal only. Issues created for review findings tend to linger forever since they are not urgent enough to fix immediately