Agent Model Router & Cost Optimizer

ReformCode routes specialist-agent work through a deterministic model router before the agent starts. The router chooses the best executable model for architecture, implementation, review, testing, security, and fix-forward work using quality, latency, cost, context size, provider health, and historical acceptance outcomes.

Why It Exists

Model choice is a product-quality decision, not a preference dropdown. A cheap model can be excellent for a small review but wrong for a risky implementation. A powerful model can be wasteful for deterministic security checks. The router makes that tradeoff explicit and auditable.

Routing Signals

• Task fit: Each candidate declares whether it supports architect, implementer, reviewer, tester, security, or fix-forward work.
• Execution mode: Tool-using agents are constrained to models with the current agent tool-loop adapter. This prevents routing to a model that looks good on paper but cannot safely execute the workspace loop.
• Quality score: Candidate quality is scored per task, with higher weight for complex implementation and fix-forward tasks.
• Latency: Median latency and optional latency budgets influence fast-path decisions.
• Cost: Estimated input/output token cost is converted into credit estimates and used by economy/balanced routing.
• Context size: Large prompts and workspaces route toward long-context models when the task does not require the current tool-loop adapter.
• Provider health: Unhealthy providers are demoted while still preserving an escape route if every configured provider is unhealthy.
• Historical outcomes: Telemetry can lift models that have produced accepted outcomes for similar agent roles.

Runtime Behavior

Every orchestrated agent pipeline now computes model-routing decisions before execution. Agent results carry:

• selected model ID
• selected provider
• routing strategy
• estimated credit cost
• fallback model IDs
• selection rationale
• warnings such as provider health, budget pressure, or context overflow

Security work uses the deterministic audit engine instead of spending model credits. Tool-using agent roles currently route to executable Anthropic tool-loop models; analysis-only routing can rank Gemini and OpenAI candidates for long-context and historical-outcome scenarios.

Continuous Evaluation

The continuous evaluation CI model-router suite now checks real routing behavior instead of placeholder provider ordering:

• complex tool-loop implementation chooses an executable high-quality model
• large-context architecture work chooses a long-window provider
• historical acceptance outcomes can lift a cheaper accepted reviewer model
• all-unhealthy provider states still preserve a fallback escape route

Operator Notes

• Keep MODEL_ROUTER_CANDIDATES aligned with real provider adapters.
• Do not mark a model as agent_tool_loop capable until the runner can execute its tool-call protocol safely.
• Add evaluation fixtures before changing weights or candidate quality baselines.
• Treat cost savings as a win only when acceptance and review quality stay healthy.