Context#

The strategy engine (lib/strategy-engine.ts) can select the optimal model for a task, but no agent natively calls selectModel() before running. This makes the strategy engine purely analytical — it does not affect actual routing.

Acceptance Criteria#

• POST /api/tasks/[id]/run calls selectModel() before dispatching
• Task profile is auto-classified from title + body + tags + assignee
• Preferred model (from task frontmatter model:) overrides auto-selection
• If selectModel() fails, fall back to agent's default model
• Routing decision is logged via routing-log.ts
• Agent model is set for single run (not permanently changed)
• Tests cover dispatch with and without strategy
• All 696+ tests pass

Design Reference#

See projects/model-strategy-engine/PROJECT.md — Gap 1 (Agent-Side Strategy Hook)

Key considerations#

• Use selectModel() from lib/strategy-engine.ts (already exists)
• Use classifyTaskProfile() to auto-classify (already exists)
• Use routing-log.ts (Phase 1) for logging decisions
• Dispatch pipeline is in lib/actions.ts (runTask action)
• Agent model override for single run: may need config patch + restart or use OpenClaw session API