test(together_ai): move request-shape checks to the mapped file, drop the live ones

[yuneng-berri]

TLDR

Problem this solves:

• Three tests pinned a Together model that left serverless
• The live API now answers 400 model_not_available

How it solves it:

• Delete all three: e2e already covers the live behavior
• Assert the two request shapes in the mapped transformation file

User Flow

This is a test-only change. No runtime behavior changes, so there is no end user flow to describe

Relevant issues

Affected release

Linear ticket

Pre-Submission checklist

• I have added meaningful tests
• The handful of test files covering my change pass locally
• My PR passes all required CI/CD checks
• My PR's scope is as isolated as possible; it only solves 1 specific problem
• I have received a Greptile Confidence Score of at least 4/5 before requesting a maintainer review

Screenshots / Proof of Fix

This PR only touches test files, so there is nothing to exercise against a live proxy: runtime behavior is identical before and after. The failure was the live vendor call itself, so the proof is the red run, plus a mutation check that the replacement assertions actually bite

Before (c64b821)

the three tests fail against Together

1. uv run pytest tests/local_testing/test_completion.py::test_completion_custom_provider_model_name tests/local_testing/test_completion.py::test_customprompt_together_ai tests/local_testing/test_completion.py::test_completion_together_ai_stream -q
2. Output:

E  litellm.BadRequestError: Together_aiException - {
E    "error": {
E      "message": "Unable to access non-serverless model openai/gpt-oss-20b. Please visit
E                  https://api.together.ai/models/openai/gpt-oss-20b to create and start a
E                  new dedicated endpoint for the model.",
E      "type": "invalid_request_error",
E      "code": "model_not_available"
E    }
E  }
3 failed

After (8c046e1)

the three are gone and the file still collects

1. uv run pytest tests/local_testing/test_completion.py --collect-only -q
2. Output:

127 tests collected in 0.08s

the mapped file carries the two request-shape assertions

1. uv run pytest tests/test_litellm/llms/together_ai/chat/test_together_ai_chat_transformation.py -q
2. Output:

85 passed, 2 warnings in 1.34s

the new prefix assertion catches an over-strip

1. Temporarily change "model": model, to "model": model.split("/")[-1], in litellm/llms/openai/chat/gpt_transformation.py
2. uv run pytest tests/test_litellm/llms/together_ai/chat/test_together_ai_chat_transformation.py -q
3. Output:

FAILED ...::test_completion_routes_through_together_chat_config
FAILED ...::test_only_the_provider_prefix_is_stripped_from_a_slashed_model_name
2 failed, 83 passed

4. Revert the edit

Type

✅ Test

Caveats (if any)

Medium

• test_custom_role_wrappers_never_reach_the_request pins roles being ignored
  • Together goes through the OpenAI compatible path, so the wrappers never reach the request
  • The kwarg is dropped with no error, which someone may want to change
  • Pinning it means a deliberate change gets noticed instead of passing silently

Low

• Streaming now rides on e2e only
  • tests/e2e/llm_translation/test_together_ai_e2e.py covers it three ways against the live API
  • That file selects its model from the cost map, so it cannot rot the way the deleted tests did
  • It runs on the e2e lane rather than per PR, so a Together streaming break surfaces there, not in the fast suite

Final Attestation

• The tests check the right things, including the edge cases, and regressions in the respective real-world customer use-cases are not possible after this PR

[greptile-apps]

Greptile Summary

This test-only PR removes three brittle Together AI live tests pinned to an unavailable model and moves request-shape coverage into the mocked transformation suite.

• Adds cost-map-driven selection for remaining Together live tests.
• Uses a typed, immutable model for validating relevant cost-map fields.
• Verifies that provider prefixes and custom role wrappers are handled correctly without making network requests.

Confidence Score: 5/5

The PR appears safe to merge because the remaining live tests select eligible models dynamically and the moved request-shape assertions run through a mocked transport.

No new actionable issues remain, and all previous findings were resolved by deleting the affected helpers or replacing the coarse type with a concrete validated model.

Important Files Changed

Filename	Overview
tests/_live_test_helpers.py	Adds typed selection of the cheapest active Together chat model matching requested capabilities.
tests/llm_translation/test_together_ai.py	Replaces a retired pinned model with capability-aware dynamic model selection.
tests/local_testing/test_completion.py	Removes three brittle live Together tests whose request-shape behavior is covered by mocked tests or e2e coverage.
tests/local_testing/test_text_completion.py	Replaces the retired Together model pin with cost-map-driven model selection.
tests/test_litellm/llms/together_ai/chat/test_together_ai_chat_transformation.py	Adds mocked request-shape coverage using LiteLLM's existing HTTP handler and httpx MockTransport.

_{Reviews (3): Last reviewed commit: "refactor(test): validate cost-map entrie..." | Re-trigger Greptile}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[yuneng-berri]

@greptile

[yuneng-berri]

@greptile