Stage: Plan Model: Opus (strong reasoning) Task: Scope the payment retry logic in read-only mode. List files to touch and open questions. Don't write code.
Stage: Code Model: Sonnet (fast, code-tuned) Task: Implement the plan above. Keep the diff under 200 lines.
Stage: Review Model: GPT-5.5 (didn't write this code) Task: Review the diff for correctness and missed edge cases. List every issue, even minor ones.
Stage: Test Model: GPT-5.5 Task: Write tests for the edge cases found during review. Don't just cover the happy path.