Implement model routing for simple versus reasoning-heavy agent workflows

[michaelmwu]

Parent: #257

Goal

Define and implement a model routing strategy so the system uses stronger reasoning models where they help, while keeping simple operations fast and inexpensive.

Scope

• Use smaller/faster models for:
  • intent classification
  • simple field extraction
  • simple task creation
  • summaries
• Use stronger reasoning models for:
  • ambiguous English requests
  • multi-step plans
  • cross-tool orchestration
  • service/code modification workflows
  • PR drafting
  • tradeoff explanations before risky actions
• Route based on complexity, risk, and workflow type.
• Keep permission decisions outside the model entirely.

Acceptance criteria

• The orchestrator can select a model tier based on request class.
• High-risk or multi-step workflows use the stronger reasoning tier for planning.
• Permission checks are deterministic and unchanged by model tier.
• Routing decisions are logged for debugging and cost analysis.

[michaelmwu]

Implemented in PR #264. The agent has fast/strong/reasoning tier routing, OpenAI-compatible provider fallback configuration, deterministic policy independent of model choice, model metadata in plans/audit/dashboard reporting, and live eval coverage for planner model comparison.