Model Request Profiles

The OpenAI adapter reads ModelInfo.request_profile from the same registry used for pricing and output limits. Model names and reasoning levels are data; the adapter does not infer capabilities from model-name prefixes. Exact model keys and explicit registry aliases are supported. Register snapshot names explicitly.

Updating a model

Edit src/tnh_scholar/runtime_assets/registries/providers/openai.jsonc, or supply a provider registry under registries/providers/openai.jsonc in the workspace or user configuration root. Provider files are selected in workspace, user, then bundled order; a higher-precedence file replaces the entire provider file. Do not expect new bundled models to appear through a stale local provider file.

Add the model's identity, limits, capabilities, pricing, and request profile:

"request_profile": {
  "reasoning_efforts": ["low", "medium", "high", "xhigh", "max"],
  "default_reasoning_effort": null,
  "temperature": "never"
}

• reasoning_efforts lists literal API values. New levels require only a registry edit. Unknown levels fail before provider dispatch and show supported values.
• default_reasoning_effort is an application default, used when neither CLI nor configuration selects an effort. It must belong to reasoning_efforts. Null omits the field and uses the provider default.
• temperature accepts always, never, or without_reasoning. The last mode permits temperature for literal none. Set temperature_without_effort: true only when the provider also accepts temperature with the reasoning field omitted.
• Profiles without reasoning support use an empty effort list. Older registries without profiles default to no explicit reasoning and temperature allowed; reasoning models in such files must be updated with an explicit profile.
• auto is reserved for omitting the API field. none and max are literal provider values. The implementation never silently downgrades an effort.

Existing models can have their complete request_profile replaced in registries/overrides/openai.jsonc, alongside existing pricing overrides:

{
  "schema_version": "1.0",
  "provider": "openai",
  "models": {
    "gpt-6-astra": {
      "request_profile": {
        "reasoning_efforts": ["low", "medium", "high", "xhigh", "max"],
        "default_reasoning_effort": "high",
        "temperature": "never"
      }
    }
  }
}

Overrides modify existing models; new models belong in the provider file. Restart the process after changing registries because the loader caches them. Validate with tests/gen_ai_service/test_model_request_profiles.py and the provider, registry, safety, and CLI suites. The regression suite includes a new model with a new effort defined solely in a temporary registry.

Astra and pricing

Astra's bundled profile targets the existing Chat Completions text/structured output transport. Its function-calling flag is false here because Astra tool calling requires Responses. New endpoints or new parameter shapes still require transport implementation; registry entries cannot create those capabilities.

Prices are dollars per 1K tokens. Astra standard rates are 0.01 input, 0.05 output, 0.001 cached input, and 0.0125 cache-write input. Before dispatch, the estimator conservatively uses the larger input/cache-write rate. Above 272,000 input tokens, long_context multiplies all input/cache rates by 2 and output rates by 1.5. This covers an individual request, not retry totals or an exact invoice. Other models' pricing was not refreshed in this change; the registry's overall freshness date is intentionally unchanged.

Sources checked September 19, 2026:

• Astra model specification
• Astra migration guidance
• GPT-5.4 parameter compatibility
• GPT-5.5 reasoning levels

See tnh-gen CLI reference for reasoning precedence and the changed max, none, and auto semantics.