[adambalogh]

Why

The gateway accepted a reasoning_effort field on chat requests but never sent it on to the provider, so every model ran at its provider's default. DeepSeek V4 thinks at effort high by default and opened every reply with a long reasoning pass, and no client could turn that down.

An earlier revision of this PR turned DeepSeek's thinking off with a per-model registry flag. This revision drops that flag: the client now chooses per request with the standard OpenAI field, and the gateway translates it once per provider, with no lists of models.

What

• Request: reasoning_effort is parsed, checked against REASONING_EFFORTS (none, minimal, low, medium, high, xhigh, max; anything else is a 400) and passed to get_chat_model_cached, where it is part of the cache key.
• Translation, once per provider:
  • OpenAI, xAI, Z.ai, abliteration.ai, ModelArk: sent as-is. On OpenAI's Responses API path, langchain moves it into reasoning.effort.
  • ModelArk "none": sends thinking: {"type": "disabled"} instead, since ModelArk rejects reasoning_effort alongside that switch.
  • OpenRouter: reasoning: {"effort": ...}. OpenRouter lists DeepSeek V4 Pro 0813 as accepting none/low/high/max, with reasoning not mandatory.
  • Anthropic, Google: not translated yet, so the field is ignored there.
• Omitted sends nothing, so every request that doesn't set the field reaches its provider exactly as before.
• A value the gateway knows but the model doesn't support (for example low on a model that only offers high/max) comes back as the provider's 4xx, passed through unchanged.
• Signed request hash: the field is included only when sent, like aspect_ratio and resolution.
• OpenAPI: the old enum (low/medium/high, default medium) would have rejected none at connexion's body validation, which the OHTTP inner request also goes through. The enum now lists every accepted value and has no default. The request model's own "medium" default is removed too.

Compatibility

A client that already sent reasoning_effort (it was silently dropped until now) will see it reach the provider, and will find it in the signed request hash. A verifier that rebuilds the canonical request on its own has to add the field when it sends it. chat-app does that in its companion PR.

Testing

• pytest tee_gateway/test/ tests/test_pricing.py tests/test_opengradient_field.py: 512 passed
• New tests:
  • each provider's translation (ModelArk none vs a level, OpenRouter, OpenAI) and omitted → nothing sent;
  • an unknown value is a 400;
  • the value reaches the model factory;
  • the hash covers the field only when sent;
  • "none" passes schema validation. Run against the old openapi.yaml, this test fails.
• ruff check, ruff format --check, mypy: clean
• Not tried against live providers (no keys here). After deploy, a DeepSeek request with reasoning_effort: "none" should come back with reasoning_tokens: 0.

Needs a new enclave image and PCR measurements and a redeploy. Deploy this before any client starts sending the field: an old gateway leaves it out of the signed hash, so a client that includes it would fail verification.

🤖 Generated with Claude Code

https://claude.ai/code/session_01SQvDLwdbHbxxCg4dGzeyh2