jackieziao/llmroutingrepo/select-llm-model

Select LLM Model

Select one configured model and, when requested, send the task to that model. Never print, return, log, or commit an API key.

Workflow

1. First obtain the caller's JSON config containing the sol, terra, and

luna entries shown in model-api-keys.example.json. Do not route or execute a task until the config is available and valid.

1. Prefer api_key_env references. If the config contains direct api_key

values, use them without displaying them and warn the caller to keep the file untracked. Never ask the caller to paste a secret into chat.

1. Extract the task priority (quality, balanced, speed, or cost) and

complexity. Infer missing values; ask only if a missing constraint could materially reverse the route.

1. Run scripts/select_model.py --config <path> --task <task> to select a model.

Add --execute only when the caller asked the skill to run the task through the API; this sends data and incurs API usage.

1. Return the selected model and result. Do not claim that the current Codex

session changed models: execution is a separate Responses API call.

Routing policy

• gpt-5.6-sol: difficult architecture, deep debugging, security, complex

algorithms, high-risk review, research synthesis, or maximum-quality work.

• gpt-5.6-terra: normal implementation, refactoring, tests, documentation,

code review, and balanced daily engineering. This is the default.

• gpt-5.6-luna: extraction, classification, formatting, short summaries,

boilerplate, and other low-risk or cost-sensitive work.

Explicit priority overrides inferred complexity, except that high-complexity tasks continue to use Sol. The fallback order is Sol to Terra, Terra to Sol, and Luna to Terra.

Config and command

Use three environment variables so secrets stay outside the config:

{
  "models": {
    "sol": {"model": "gpt-5.6-sol", "api_key_env": "OPENAI_SOL_API_KEY"},
    "terra": {"model": "gpt-5.6-terra", "api_key_env": "OPENAI_TERRA_API_KEY"},
    "luna": {"model": "gpt-5.6-luna", "api_key_env": "OPENAI_LUNA_API_KEY"}
  }
}

python scripts/select_model.py \
  --config model-api-keys.json \
  --task "Implement a typed REST endpoint and tests" \
  --priority balanced \
  --execute \
  --format text

Without --execute, the helper selects a route offline and does not resolve or use the keys. The JSON selection schema is:

{"model":"gpt-5.6-terra","reason":"...","fallback":"gpt-5.6-sol"}