Regenerate Python SDK from updated openapi.yaml (schemas#71)

[inference-gateway-maintainer]

Context

Source-of-truth schema change: inference-gateway/schemas#71 (merged) added the
remaining additive OpenAI-compatible request parameters to
CreateChatCompletionRequest in openapi.yaml. This Python SDK generates from that
schema, so it needs a regenerate pass.

Upstream tracking issue: inference-gateway/schemas#72.

Work

• Pull / re-vendor the updated openapi.yaml from inference-gateway/schemas.
• Regenerate the Python SDK models.
• Verify the oneOf unions (stop, tool_choice, response_format) generate cleanly (these usually become Union[...] types).
• Sanity-check the public API surface for the new optional fields.

New / changed request fields

Field	Change
temperature	number 0-2, default 1 (polish)
top_p	number 0-1, default 1 (polish)
n	integer 1-128, default 1 (polish)
stop	oneOf [string, array of 1-4 strings] (polish)
max_tokens	now deprecated (use max_completion_tokens)
frequency_penalty	number -2..2, default 0
presence_penalty	number -2..2, default 0
seed	integer
logprobs	boolean, default false
top_logprobs	integer 0-20
response_format	oneOf [text, json_schema, json_object]
tool_choice	oneOf [string enum none/auto/required, named tool choice]
logit_bias	object mapping token id -> integer bias
user	string
reasoning_effort	enum minimal/low/medium/high

Codegen note

stop, tool_choice, and response_format are oneOf unions. Verify the generator
emits them cleanly (no dropped or flattened variants). This is exactly the class of
codegen quirk that scripts/check-reachable.js guards against upstream (schemas issue #31).

Refs inference-gateway/schemas#70, inference-gateway/schemas#71, inference-gateway/schemas#72.

[edenreich]

@claude Can you work on this?

[inference-gateway-maintainer]

Claude finished @edenreich's task in 18m 48s —— View job • feat/issue-49-20260618-1241 • [Create PR ➔](https://github.com/inference-gateway/python-sdk/compare/main...feat/issue-49-20260618-1241?quick_pull=1&title=Issue%20%2349%3A%20Changes%20from%20Claude&body=This%20PR%20addresses%20issue%20%2349%0A%0AGenerated%20with%20%5BClaude%20Code%5D(https%3A%2F%2Fclaude.ai%2Fcode)

---

Regenerate Python SDK from updated openapi.yaml (schemas#71)

Done. PR opened: #50

Completed:

• Re-vendored openapi.yaml from inference-gateway/schemas (1724 → 1977 lines) and regenerated inference_gateway/models.py (task generate)
• New CreateChatCompletionRequest fields: temperature, top_p, n, stop, frequency_penalty, presence_penalty, seed, logprobs, top_logprobs, response_format, tool_choice, logit_bias, user, reasoning_effort — plus max_completion_tokens & parallel_tool_calls; max_tokens now deprecated
• Verified the oneOf unions generate cleanly (no dropped/flattened variants):
  • stop: str | Stop | None
  • response_format: ResponseFormatText | ResponseFormatJsonSchema | ResponseFormatJsonObject | None
  • tool_choice: ChatCompletionToolChoiceOption | None = Literal['none','auto','required'] | ChatCompletionNamedToolChoice
• Serialization fix: new fields carry non-None spec defaults, so model_dump(exclude_none=True) would have leaked temperature=1, top_p=1, parallel_tool_calls=True, … onto every request. Switched both client methods to exclude_none=True, exclude_unset=True so only caller-specified fields are sent.
• Exported the new response_format / tool_choice helper models in __init__.py; added a README "Sampling and Output Controls" section
• Added a regression-guard test + a forward test; black/isort/mypy clean, 25 passed, 1 xfailed

Heads-up: incidental upstream regression

The re-vendored spec also re-added content to the required list of ChatCompletionStreamResponseDelta, reverting the fix for #29 (real streaming deltas omit content). The SDK's own streaming path is unaffected (raw SSEvents), but callers validating chunks against the model are. The #29 test is marked xfail(strict=True) so it auto-flags when the spec is corrected.

Follow-ups filed

• Upstream spec fix: inference-gateway/schemas#76
• Docs: inference-gateway/docs#245

Not done (intentional)

• The upstream spec content-required fix itself (tracked in schemas#76 — once merged + re-pulled, remove the xfail).
• Kept the explicit max_tokens client param as-is; max_completion_tokens is reachable via **kwargs and documented.
  · PR #50 · branch feat/issue-49-20260618-1241