feat(iflow): add support for glm-5, minimax-m2.5, qwen3-32b, tstars2.0, iflow-rome-30ba3b

[MasuRii]

Summary

Add support for 5 new iFlow models to maintain feature parity: glm-5, minimax-m2.5, qwen3-32b, tstars2.0, and iflow-rome-30ba3b. Each model is registered with appropriate thinking, reasoning, and preservation configurations.

Changes

• Add glm-5, minimax-m2.5, qwen3-32b, tstars2.0, and iflow-rome-30ba3b to HARDCODED_MODELS
• Add glm-5 to ENABLE_THINKING_MODELS and GLM_MODELS sets
• Add qwen3-32b to ENABLE_THINKING_MODELS set
• Add minimax-m2.5 to REASONING_SPLIT_MODELS set
• Update REASONING_PRESERVATION_MODELS_PREFIXES with glm-5 and tstars prefixes
• Sort HARDCODED_MODELS list alphabetically by model family for consistency

Files Changed

File	Type	Impact
src/rotator_library/providers/iflow_provider.py	Modified	Register 5 new models with correct thinking/reasoning configurations

Testing Instructions

1. Verify each new model appears in the model list endpoint
2. Test glm-5 — should send enable_thinking + clear_thinking=false (GLM-style)
3. Test minimax-m2.5 — should use reasoning_split instead of enable_thinking
4. Test qwen3-32b — should send enable_thinking
5. Test tstars2.0 and iflow-rome-30ba3b — standard chat completions

Tested right now:

Closes #129

---

Important

Add support for new iFlow models glm-5, minimax-m2.5, qwen3-32b, tstars2.0, and iflow-rome-30ba3b with appropriate configurations in iflow_provider.py.

• Models:
  • Add glm-5, minimax-m2.5, qwen3-32b, tstars2.0, iflow-rome-30ba3b to HARDCODED_MODELS in iflow_provider.py.
  • Add glm-5 to ENABLE_THINKING_MODELS and GLM_MODELS.
  • Add qwen3-32b to ENABLE_THINKING_MODELS.
  • Add minimax-m2.5 to REASONING_SPLIT_MODELS.
  • Update REASONING_PRESERVATION_MODELS_PREFIXES with glm-5 and tstars prefixes.
• Misc:
  • Sort HARDCODED_MODELS alphabetically by model family.

^{This description was created by for 87b894b. You can customize this summary. It will automatically update as commits are pushed.}

[mirrobot-agent]

Starting my review of the iFlow provider updates. I'll be checking the configuration for the 5 new models (glm-5, minimax-m2.5, qwen3-32b, tstars2.0, and iflow-rome-30ba3b) to ensure they're correctly registered and integrated. 🔍

[mirrobot-agent]

Overall Assessment

Nice work on expanding the iFlow model support! Adding these 5 new models (glm-5, minimax-m2.5, qwen3-32b, tstars2.0, and iflow-rome-30ba3b) with their specific thinking and reasoning configurations ensures the provider stays up-to-date with the latest offerings. The internal grouping of models by family in HARDCODED_MODELS is also a great improvement for maintainability.

Architectural Feedback

The approach of using specific sets for different thinking/reasoning behaviors (ENABLE_THINKING_MODELS, GLM_MODELS, REASONING_SPLIT_MODELS) is clean and matches the established patterns in the codebase.

Key Suggestions

• Reasoning Preservation: I've noted that qwen3-32b (and some existing models like deepseek-v3.2) are missing from the REASONING_PRESERVATION_MODELS_PREFIXES list. If these models return reasoning_content, adding them to this list is necessary for coherent multi-turn conversations.
• Prefix Simplification: Suggesting a move to more generic prefixes (e.g., glm- instead of version-specific ones) to reduce the need for future updates.

Nitpicks

• Alphabetical Sorting: The deepseek family is currently slightly out of place in the family-sorted list.

Questions for the Author

• Does iflow-rome-30ba3b also support any form of thinking or reasoning content, or is it strictly a standard completion model?
• Are there specific reasons why deepseek and qwen models are excluded from the reasoning preservation cache?

This review was generated by an AI assistant.

[Copilot]

Pull request overview

Adds support for additional iFlow model IDs and updates model-specific “thinking/reasoning” handling, with substantial additional refactoring in the iFlow provider around payload shaping and streaming error handling.

Changes:

• Register 5 new iFlow models and update model capability sets (thinking, GLM handling, reasoning_split, reasoning preservation prefixes).
• Expand iFlow provider request/stream handling (suffix-based thinking config parsing, image normalization + fallback logic, gzip handling, token usage estimation).
• Harden embedding cost calculation to handle missing usage/cost metadata.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 8 comments.

File	Description
src/rotator_library/providers/iflow_provider.py	Adds new model IDs and significantly refactors request building + streaming handling (thinking config, image handling, gzip, token estimation).
src/rotator_library/client/executor.py	Makes embedding cost calculation resilient to usage=None and input_cost_per_token being None or 0.0.

---

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