missing cost fixes

[dixitaniket]

• provider tests for chat and image models
• image url retry for providers (glm image)

[Copilot]

Pull request overview

This PR fixes cases where successful inferences were missing token-usage metadata (and therefore missing OpenGradient cost/settlement data), with a particular focus on xAI’s LangChain response shapes and streaming behavior.

Changes:

• Expand extract_usage() to normalize usage from both LangChain’s usage_metadata and raw provider payloads under response_metadata (e.g., token_usage / usage).
• Ensure xAI non-streaming requests don’t accidentally use a streaming-configured cached model, and ensure xAI streaming requests explicitly request usage when applicable.
• Add targeted regression tests covering xAI streaming/non-streaming and the new extract_usage() fallbacks.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated no comments.

File	Description
tests/test_opengradient_field.py	Adds regression tests ensuring xAI streaming/non-streaming paths propagate usage/cost behavior correctly.
tee_gateway/test/test_tee_core.py	Adds unit tests for llm_backend.extract_usage() fallback behavior when LangChain omits usage_metadata.
tee_gateway/llm_backend.py	Extends extract_usage() to support additional metadata shapes and normalize reasoning tokens.
tee_gateway/controllers/chat_controller.py	Adjusts xAI streaming/non-streaming invocation kwargs and uses extract_usage() for streaming usage accumulation; logs when usage is missing.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 9 out of 9 changed files in this pull request and generated no new comments.

Suppressed comments (2)

tee_gateway/controllers/chat_controller.py:571

• This comment claims every provider is constructed with stream_usage=True, but Google models (ChatGoogleGenerativeAI) are constructed without that flag (see tee_gateway/llm_backend.py around the provider == "google" branch). Updating the comment avoids misleading future readers about why usage is available on stream chunks.

                    # Terminal usage needs no per-call opt-in: every provider
                    # is constructed with stream_usage=True (see
                    # get_chat_model_cached).

tee_gateway/test/test_provider_usage_integration.py:58

• In this live integration test list, glm-5.2 is labeled as provider Z.ai and tied to ZAI_API_KEY, but the model registry routes glm-5.2 through the ByteDance/ModelArk client (provider bytedance; see tee_gateway/model_registry.py:591-596). This mismatch is misleading and makes it harder to reason about which key is actually required for the test case.

    _ProviderCase("Z.ai", "glm-5.2", "ZAI_API_KEY"),