GLM 5.1 Together/FW, and Opus 4.7, Minimax 2.7 on together

[tawnymanticore]

What does this PR do?

GLM 5.1 Together/FW, and Opus 4.7, Minimax 2.7 on together

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: 8cac8474-219b-4cc7-af2c-b7116ef4340c

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch main

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[gemini-code-assist]

Code Review

This pull request introduces a multi-source feedback system for task runs, replacing the previous single-field user feedback with a dedicated data model and API. It also adds support for new models including Claude 4.7 and Grok 4.20, renames 'FewShot' components to 'TaskSample', and implements spec archiving and priority levels. Feedback was provided regarding missing multimodal flags for the new Claude model and potential race conditions in the UI's feedback loading state that could lead to a hanging loading indicator or data loss in the view.

[gemini-code-assist]

The anthropic provider for Claude Opus 4.7 is missing multimodal capability flags (supports_vision, multimodal_capable, multimodal_mime_types). These were present in the previous version (4.6) and are included in the openrouter provider for 4.7. Without these, multimodal features will be disabled when using the direct Anthropic provider for this model.

            KilnModelProvider(
                name=ModelProviderName.anthropic,
                model_id="claude-opus-4-7",
                structured_output_mode=StructuredOutputMode.json_schema,
                temp_top_p_exclusive=True,
                available_thinking_levels=CLAUDE_OPUS_4_7_ANTHROPIC_THINKING_LEVELS,
                default_thinking_level="high",
                suggested_for_evals=True,
                suggested_for_data_gen=True,
                supports_doc_extraction=True,
                supports_vision=True,
                multimodal_capable=True,
                multimodal_mime_types=[
                    KilnMimeType.PDF,
                    KilnMimeType.TXT,
                    KilnMimeType.MD,
                    KilnMimeType.JPG,
                    KilnMimeType.PNG,
                ],
            ),

[gemini-code-assist]

The feedback_loading state will hang if submit_feedback is called while a load_feedback request is pending. This happens because submit_feedback increments feedback_request_id, causing the finally block in load_feedback to skip clearing the loading state. You should ensure that feedback_loading is cleared when a request is superseded by a non-loading action like a submission.

[gemini-code-assist]

Incrementing feedback_request_id here cancels any pending load_feedback GET request. If a user adds feedback immediately after switching to a run (before the initial list loads), the GET request's result will be ignored, and the UI will only show the newly added feedback, losing all existing ones for that run. Consider allowing the GET request to complete and merging the results, or triggering a fresh load after the POST completes.

[github-actions]

📊 Coverage Report

Overall Coverage: 91%

Diff: origin/remote_config...HEAD

• app/desktop/studio_server/copilot_api.py (66.7%): Missing lines 434
• app/desktop/studio_server/utils/copilot_utils.py (76.9%): Missing lines 234-238,243
• libs/core/kiln_ai/datamodel/init.py (100%)
• libs/core/kiln_ai/datamodel/datamodel_enums.py (100%)
• libs/core/kiln_ai/datamodel/feedback.py (100%)
• libs/core/kiln_ai/datamodel/task_run.py (80.0%): Missing lines 155
• libs/server/kiln_server/feedback_api.py (100%)
• libs/server/kiln_server/server.py (100%)

Summary

• Total: 80 lines
• Missing: 8 lines
• Coverage: 90%

Line-by-line

View line-by-line diff coverage

app/desktop/studio_server/copilot_api.py

Lines 430-438

  430 
  431             for run in task_runs:
  432                 run.save_to_file()
  433                 saved_models.append(run)
! 434                 dataset_runs.save_pending_feedback(run)
  435 
  436             spec.save_to_file()
  437             saved_models.append(spec)
  438         except Exception:

---

app/desktop/studio_server/utils/copilot_utils.py

Lines 230-247

  230             self._pending_feedback[task_run.id] = feedback_text
  231 
  232     def save_pending_feedback(self, task_run: TaskRun) -> None:
  233         """Create Feedback children for a saved TaskRun if it has pending feedback."""
! 234         if not task_run.id:
! 235             return
! 236         feedback_text = self._pending_feedback.get(task_run.id)
! 237         if feedback_text:
! 238             fb = Feedback(
  239                 feedback=feedback_text,
  240                 source=FeedbackSource.spec_feedback,
  241                 parent=task_run,
  242             )
! 243             fb.save_to_file()
  244 
  245 
  246 def create_dataset_task_runs(
  247     all_examples: list[SampleApi],

---

libs/core/kiln_ai/datamodel/task_run.py

Lines 151-159

  151         """
  152         return self.thinking_training_data() is not None
  153 
  154     def feedback(self, readonly: bool = False) -> list[Feedback]:
! 155         return super().feedback(readonly=readonly)  # type: ignore
  156 
  157     # Workaround to return typed parent without importing Task
  158     def parent_task(self) -> Union["Task", None]:
  159         if self.parent is None or self.parent.__class__.__name__ != "Task":

---

---

• 📊 HTML Coverage Report - Interactive coverage report
• 📈 Diff Coverage Report - Detailed diff analysis
• Github Actions Run - View the full coverage report