review-plan

name: review-plan description: Review an implementation plan using multiple AI models (GPT-4o, Gemini Flash) across 5 dimensions. Use when the user mentions "/review-plan", asks to review a plan, or after creating an implementation plan that would benefit from external validation.

Plan Review

Send an implementation plan to external AI models for structured review across 5 dimensions: completeness, blind spots, regression risk, test coverage, and hypothesis scope. Uses OpenAI (GPT-4o) and Google (Gemini Flash) only — no Anthropic tokens consumed. Feedback is deduplicated and ranked by severity.

1. Identify the Plan

• If the user provides plan text directly, use that.
• If a plan was just created in the current session (via plan mode), use that plan.
• Otherwise, check research/plans/ for the most recent plan file.
• If no plan is found, ask the user to provide one.

2. Gather Context

Read files mentioned in the plan to provide additional context to reviewers:

• Read up to 5 files referenced in the plan (max 200 lines each)
• Concatenate their contents as supplementary context

3. Check Provider Availability

Call list_review_providers to see which API keys are configured. Report to the user:

• Which providers are available (OpenAI, Google)
• If none are available, inform the user they need to set API keys

4. Auto-Select Budget

Based on plan size:

• Small plan (<50 lines): minimal (1 model per dimension, ~$0.01-0.03)
• Medium plan (50-200 lines): standard (2 models per dimension, ~$0.03-0.08)
• Large plan (>200 lines): thorough (3 model calls per dimension using gpt-4o, o3-mini, gemini-flash, ~$0.08-0.20)

5. Auto-Select Dimensions

Skip dimensions that don't apply:

• Skip hypothesis_scope for non-experiment plans (no hypothesis, no ML metrics)
• Skip regression_risk for plans that only create new files (no existing code changes)
• Always include completeness, blind_spots, and test_coverage

6. Execute Review

Call review_plan with:

• plan: the full plan text
• dimensions: the selected dimensions
• context: gathered file contents
• include_adrs: true (always inject relevant ADR context)
• budget: the auto-selected budget tier

7. Present Results

Format the output as a structured report:

Critical Issues

List any items with severity "critical" — these should be addressed before implementation.

Warnings

List items with severity "warning" — these are worth considering but may not block progress.

Suggestions

List items with severity "suggestion" — nice-to-have improvements.

Summary

• Number of items found per dimension
• Which providers contributed feedback
• Overall assessment: "Plan looks solid" / "Some concerns to address" / "Significant gaps identified"

For each item, show:

• Description: What the issue is
• Affected files: Which files are impacted
• Reasoning: Why this matters
• Corroborated by: Which models flagged this (items flagged by multiple models are more likely real issues)

8. Follow-Up

Ask the user what they'd like to do:

1. Address issues — update the plan to fix critical/warning items
2. Deeper review — re-run on a specific dimension with all providers
3. Proceed as-is — accept the plan and begin implementation

9. Wrap-Up

1. Save the plan to the research/plans/ directory like f"{timestamp}_{title_slug}.md" where title slug is a descriptive name from the content of the plan.