Declarative multi-model pipeline tool for coding agents that routes planning, implementation, and review to different models with automatic handoff and cost tracking

Developers manually chain models by role -- Gemini for architecture planning, DeepSeek for implementation, Claude for code review -- because each model has a different strength-to-cost profile per task type. This is entirely manual: copy output from model A, paste into model B, copy result, paste into model C. No tool lets developers declare a three-stage pipeline (planner model, executor model, reviewer model) with automatic context handoff, per-stage cost attribution, and output diff between stages. OpenRouter handles provider routing but has no task-role mapping, no staged handoff, and no output comparison across stages. The 2024 RouteLLM framework (244 HN points, 36 comments) solved cost-vs-quality routing for single requests; the multi-role pipeline for coding workflows -- where stages have different latency, cost, and quality requirements -- is un-served. Active developer practice of manual model chaining documented in HN threads (June 2026) confirms the behavior exists without tooling.