nhouseholder/smart-model-router

Smart Model Router — Right Model for the Job

How It Works

The smart-model-router.py hook (UserPromptSubmit) classifies each prompt and recommends a model tier. If the current model matches, nothing happens (zero overhead). If mismatched, it injects a recommendation.

Classification Tiers

Sonnet Tier (fast, cheap — ~5% of prompts)

Simple tasks where Opus is overkill:

• Single-file formatting, renaming, or trivial edits
• "What does X do?" questions about visible code
• Running a command and reporting output
• Git status, git log, simple git operations
• Reading/summarizing a file
• Answering factual questions about the codebase

Signals: Short prompt (<50 words), no multi-step reasoning, no domain expertise, no creative work, single tool call expected.

Opus Tier (standard — ~85% of prompts)

Most development work:

• Multi-file changes, feature implementation
• Debugging, architecture decisions
• Code review, refactoring
• Planning, spec writing
• Any domain-specific logic (betting, finance, etc.)
• Creative design work
• Multi-step tasks

Signals: Requires reasoning, judgment, or multi-step execution.

Opus 1M Tier (extended context — ~10% of prompts)

Tasks requiring massive context:

• Full codebase audits (/site-review, /site-audit, /audit)
• Reviewing 10+ files simultaneously
• Large refactors spanning many files
• Comparing entire codebases or large diffs
• /site-redesign (multi-phase agent pipeline)
• Understanding complex data flows across many modules
• Sessions that have been running long and context is filling up

Signals: Task explicitly involves "all files", "entire codebase", "full audit", or references 10+ files. Also triggered when the conversation has been running long and nearing context limits.

Model IDs

| Tier | Model Picker | Model ID | |---|---|---| | Sonnet | Sonnet 4.6 | claude-sonnet-4-6 | | Opus | Opus 4.6 | claude-opus-4-6 | | Opus 1M | Opus 4.6 (1M) | claude-opus-4-6 with extended context |

Recommendation Behavior

• Current = Opus, Task = Sonnet: "This is a simple task — Sonnet would handle it faster. Switch via model picker if you want to save tokens."
• Current = Sonnet, Task = Opus: "This task needs Opus-level reasoning. Switch to Opus via model picker."
• Current = Opus, Task = 1M: "This task involves large context — consider switching to Opus 1M via model picker for full codebase visibility."
• Match: Silent. No recommendation injected.

Important Limitations

• Hooks cannot programmatically switch models — the user must use the model picker
• The recommendation is injected as additionalContext on UserPromptSubmit
• Classification is heuristic — when uncertain, defaults to current model (no recommendation)
• Never recommends downgrading mid-task (only at task boundaries)