Effective parameter assessment and temperature policy

C9-B adds pure assessModelParameters({ vendor, model, endpoint, reasoningEffort? }) on root and browser-safe /shared. It assesses actual effective reasoning, not a model-name-only temperature flag. Named vendor is mandatory. Custom/unknown endpoints and models remain unknown. Supported/unsupported/unknown are distinct. Missing effort uses only documented family defaults; elsewhere it remains unknown.

The additive opt-in parameterPolicy: { temperature: 'omit-unsupported', observe? } applies to Agent configuration, managed run/stream, direct calls and provider requests. temperature: 'strict' explicitly selects strict assessment, including native counting. With no policy selected, legacy generation errors/filtering and native-count behavior remain unchanged. No API migration is required.

const agent = Agent.create({
  connector: 'openai-main', model: 'gpt-6-sol', temperature: 0,
  thinking: { enabled: true, effort: 'high' },
  parameterPolicy: {
    temperature: 'omit-unsupported',
    observe: decision => hostDiagnostics(decision),
  },
});
await agent.run('Do the task');

“Provider-neutral temperature” means the normalized temperature option, including zero; it does not refer to a special neutral numeric value. The policy omits that field only when unsupported by the selected endpoint/model/effective effort. Supported zero, other supported values, unknown values and absence are preserved. Saved configuration is never mutated. Raw vendorOptions.temperature, top_p, top_logprobs and unsupported include controls keep strict validation. The policy is not a generic sanitizer and does not repair invalid reasoning or other options.

OpenAI Responses generation, stream, native count and batch evaluate the same request facts after overrides, reasoning defaults and explicit conversation configuration_update items. Unified enabled thinking takes precedence over raw request reasoning; disabled thinking does not invent none. The latest native configuration update determines the selected effective effort without rewriting the cached request prefix. Actual route selection/fallback and per-call overrides are assessed at dispatch. A single fact primitive also serves the frozen legacy validation adapter; no second router/provider layer is created.

GPT-6 Sol/Luna and documented 6.1 Terra support sampling at none; Astra/6.1 Sol do not accept none. Active reasoning disallows sampling controls. Terra facts do not add a model to the registry or establish other capabilities. Other static registry facts remain conservative. OpenAI parameter guidance, checked 2026-10-28. Custom gateways do not inherit official endpoint facts.

Selected native counting validates constraints before projecting input-bearing fields; generation-only temperature need not be sent to the counting endpoint. Existing native replay restrictions still reject unsupported native item kinds, including configuration updates; policy selection does not broaden those ownership contracts. An ordinary legacy count still behaves as before. Counting success does not certify all dispatch, authentication or capacity constraints. Raw Responses WebSocket keeps its existing strict validator and has no normalized-temperature policy selection. Pure Chat Completions assessment does not add a Chat Completions transport. Helpers retain their existing explicit configuration/validation contracts.

observe receives a bounded detached decision (model, named vendor, endpoint, action, effective effort and closed reason code), never prompts or credentials. Failures/rejected observers are isolated. Decisions can occur for preparation/count as well as dispatch; they are not billable-call counters or completion proof. There is no global decision journal. Hosts own retention and presentation.

Policy callbacks are host authority. Runtime snapshots include them only in the host-local view; portable v2 exports omit the policy before traversal. Receiving hosts reconfigure it explicitly. Old package/session DTOs and default hooks/events remain unchanged. ICOS must keep its guard until explicitly adopting this contract; static model-only omission can remain conservative. AMOS has no policy selector, and its defaults are unchanged.

Run npm run example:parameter-policy for actual mocked Responses bodies, high reasoning omission, supported zero at none, explicit raw rejection and immutable saved temperature. It performs no network IO. Tests cover actual generation, streaming, count and batch builders, routing/fallback and native effort updates. Live testing remains paused by user instruction.