Model Selector

Model Selector

Control the preferred backbone LLM for compatible AI-agent workflows. Two modes: manual (user picks a specific model) or auto (dynamically selects the best model based on task category via keyword detection).

Commands

Execution

All commands are executed by running the CLI script:

python3 scripts/model_selector.py <command> [args]

The script path is relative to this skill's directory. A common installed-skill path is:

python3 /home/ubuntu/skills/model-selector/scripts/model_selector.py <command> [args]

When a user sends a command (e.g., /model-selector auto build a website), the active AI client should:

1. Parse the command and arguments from the user message.
2. Invoke the CLI script via shell with the appropriate command and arguments.
3. Read the script output and report the result to the user.
4. For errors, display the error message and suggest remediation.

Architecture

Persistent Files

Config (~/.model_selector_config.json):

{
  "enabled": true,
  "mode": "manual|auto",
  "manual_model": "model_id|null",
  "auto_category": "category|null",
  "last_selected_model": "model_id|null",
  "last_selected_reason": "human-readable explanation",
  "updated_at": "ISO-8601 timestamp"
}

Cache (~/.model_selector_cache.json):

{
  "cached_at": 1712345678,
  "cached_at_human": "ISO-8601",
  "total_models": 349,
  "frontier_models": [{"id": "...", "provider": "...", "context_length": 0, "price_per_m_output": 0}],
  "model_lookup": {"model_id": {"id": "...", "name": "...", "context_length": 0, "pricing_prompt": "0", "pricing_completion": "0"}}
}

Task Category Detection

Auto mode classifies tasks into six categories by word-boundary regex matching: coding, reasoning, research, creative, math, or general (fallback). Each category has a curated ranking of the best models based on benchmark data. No AI classification is used — keyword detection is deterministic, fast, and credit-free. Specialized categories always win ties over "general".

Agent Integration

At the start of each task, a compatible client should:

1. Read ~/.model_selector_config.json.
2. If enabled is true, note the last_selected_model and mode.
3. Use this model preference when making decisions about which LLM to invoke as the backbone.
4. If enabled is false, config is missing, or config is corrupted, use the client's default backbone model.

Other skills' model preferences remain unaffected.

Workflow

1. Parse command from user message.
2. Execute via the CLI script using shell.
3. Report result — which model was selected, why, detected category (if auto), and alternatives.
4. For errors: Respond with clear error message, usage help, and suggested remediation.

Error Handling

• Cache missing/corrupted: Auto-reinitialize on next use or prompt user to run refresh.
• Refresh fails: Keep stale cache, warn user, suggest retry.
• Unknown model set: Warn but accept (may be private/custom model).
• Invalid command: Show usage help with all supported commands.
• File permission errors: Suggest remediation steps.
• Config corrupted: Fall back to defaults, notify user.
• Malformed API data: Skip entries with missing IDs or non-numeric pricing gracefully.

Performance Targets

• Read-only commands (status, list, recommend): under 1 second (no network calls).
• Write commands (set, auto): under 5 seconds (may refresh cache if stale).
• Cache refresh: under 30 seconds.
• Cache auto-expires after 24 hours; refreshes on next write command.

Security

• Never expose OPENROUTER_API_KEY in outputs, logs, or error messages.
• Atomic file writes to prevent config/cache corruption.
• Preserve manual model settings unless explicitly changed by user.

Requirements

• OPENROUTER_API_KEY environment variable (for cache refresh and model validation).
• requests Python package.

Benchmark Reference

For detailed benchmark data, tier rankings, and cost analysis, read: references/model_benchmarks.md

Only read this file when the user asks about specific benchmark data, model comparisons, or pricing details.

Constraints

• Controls backbone LLM preference only — does not alter models used by other skills.
• Uses word-boundary regex task detection exclusively — no AI classification to avoid extra costs.
• Cache refresh rate-limited: max 1 per minute, auto-refresh every 24 hours.
• Manual model settings preserved unless explicitly changed by user.
• Graceful degradation: if disabled or config/cache unavailable, the active client uses its default model.