Model router

Model Router

Classify each task by difficulty and delegate it to the optimal model + reasoning effort, balancing cost and quality.

Procedure

1. Classify — Assign the task a tier (T0–T3) using the rubric below. Classify multi-task requests item by item. When unsure, round up one tier.
2. Pick the delegation target — Look up the current platform's target in references/routing.md. Platform-specific delegation mechanics are in references/platforms.md.
3. Delegate — Write a self-contained prompt: background, target file paths, and completion criteria.
4. Verify — Always check the result. If the delegate returned ESCALATE: or verification failed, escalate one tier and retry exactly once. If it fails again, take the task over yourself (main loop).
5. Report — End with one line telling the user which tier/model executed the task.

Difficulty rubric

Tier	Criteria	Examples
T0 Lite	Single file, mechanical, one correct answer	Typo fixes, renames, import cleanup, boilerplate, file lookups/summaries, formatting
T1 Standard	Clear scope, a few files, follows existing patterns	Reproducible bug fixes, small features, adding tests, in-module refactors, docs updates
T2 Deep	Cross-cutting, unknown cause, involves design decisions	Multi-module changes, hard debugging, performance work, concurrency, migrations
T3 Frontier	Ambiguous, high-risk, architecture-level	Tech selection, architecture design, clarifying vague requirements, large refactor planning

Hard rules (override the rubric)

• Tasks touching security, auth, billing, or data deletion are at least T2.
• An explicit model/target choice from the user always wins (this skill's classification is the default behavior).
• Don't spend time on classification. If you can't decide instantly, go one tier up.
• Delegation prompts for T0/T1 must include: "If this turns out harder than expected, stop and reply only with ESCALATE: <reason>."

Cross-model consultation (optional, T2/T3)

For design decisions or reviews where a second opinion from a third-party model (Kimi / Qwen / DeepSeek, etc.) helps, consult via OpenRouter:

echo "<question or code to review>" | node scripts/consult.mjs --model moonshotai/kimi-k2.7-code
node scripts/consult.mjs --models qwen        # search available model IDs (no API key needed)

If OPENROUTER_API_KEY is not set, the script runs as a dry-run (prints the request without sending). Treat consultation results as input, not verdicts — make the final call yourself.