fix(eval): strip restricted temperature for default-only models in mocks

[cotovanu-cristian]

Summary

The eval LLM mock generator forwarded the simulation strategy's temperature verbatim to the LLM Gateway. For models that only accept the default temperature (gpt-5 family, o-series reasoning models), an explicit temperature=0 makes the gateway return HTTP 400 (Unsupported value: 'temperature' does not support 0 with this model. Only the default (1) value is supported.). That error collapses through UiPathMockResponseGenerationError → AgentRuntimeError into an Unknown-category runtime failure.

This normalizes completion kwargs at the structured-output funnel (generate_structured_output), dropping the restricted temperature for those models before calling chat_completions. Both the LLM mocker and the input mocker route through this funnel, so the path is covered once. Standard models are unaffected.

Root cause

• _llm_mocker.py builds completion_kwargs from ModelSettings and forwards them to chat_completions, which always sends temperature in the request body.
• gpt-5 / o-series reject any explicit temperature → HTTP 400.
• Representative job key: 9dc7655c-241d-4a9d-a190-1545df616acc.

Changes

• _structured_output.py: add _model_supports_only_default_temperature + _normalize_completion_kwargs, applied once in generate_structured_output.
• test_structured_output.py: parametrized regression tests for gpt-5/o-series (temperature stripped) and a guard that standard models keep it.

Not included

Preserving the underlying HTTP error category on UiPathMockResponseGenerationError was scoped out: it requires new category plumbing on a shared exception type used across the mocker modules and the runtime's category logic (not localized). The temperature fix removes the miscategorized failures at the source.

Testing

ruff check + format clean; mypy clean on changed files; pytest tests/cli/eval/mocks/ = 65 passed.

Refs: SRE-607465, PC-4769

[Copilot]

Pull request overview

This PR fixes eval-time LLM mock generation failures by normalizing completion_kwargs in the shared structured-output funnel so that models which only allow the gateway’s default temperature (notably gpt-5 and o* reasoning models) do not receive an explicit temperature parameter that triggers HTTP 400s.

Changes:

• Added model detection + normalization logic in generate_structured_output() to drop temperature for default-only-temperature models before calling chat_completions.
• Added regression tests ensuring temperature is stripped for gpt-5 / o1 / o3 models and preserved for standard models.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
packages/uipath/src/uipath/eval/mocks/_structured_output.py	Normalizes completion_kwargs to omit temperature for models that reject explicit temperature values.
packages/uipath/tests/cli/eval/mocks/test_structured_output.py	Adds regression coverage verifying stripping behavior for default-only models and non-stripping for standard models.

---

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

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
100.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud