feat: add thinking mode support for GLM-4.6 and GLM-4.6V models #11114

roomote · 2026-01-30T16:23:26Z

This PR attempts to address Issue #11071.

Related GitHub Issue

Addresses: #11071 (comment by @mark-ucalgary)

Description

According to the Z.ai documentation for GLM-4.6 and GLM-4.6V, both models support thinking mode. However, the current codebase only defines thinking support for GLM-4.7.

This PR adds thinking mode support for GLM-4.6 and GLM-4.6V models across:

Model type definitions
GLM detection logic
Z.ai provider implementation
Test coverage

Changes Made

1. Model Definitions (`packages/types/src/providers/zai.ts`)

Added supportsReasoningEffort: ["disable", "medium"] to GLM-4.6 (international and mainland)
Added supportsReasoningEffort: ["disable", "medium"] to GLM-4.6V (international and mainland)
Added reasoningEffort: "medium" and preserveReasoning: true to both models

2. GLM Detection (`src/api/providers/utils/glm-model-detection.ts`)

Updated supportsThinking detection to include GLM-4.6 base model
Updated to include GLM-4.6V and its variants (v-flash, v-flashx)
Now detects: GLM-4.6, GLM-4.6V, GLM-4.6V-Flash, GLM-4.6V-FlashX, and GLM-4.7 variants

3. Z.ai Provider (`src/api/providers/zai.ts`)

Updated comments to reflect support for GLM-4.6, GLM-4.6V, and GLM-4.7
Changed detection from checking specific model ID to checking for supportsReasoningEffort array
This makes the implementation generic and future-proof for any GLM model with thinking support

4. Tests (`src/api/providers/utils/tests/glm-model-detection.spec.ts`)

Added test cases for GLM-4.6 base model thinking support
Added test cases for GLM-4.6V vision variants thinking support
Reorganized tests for clarity between models that support thinking and those that do not

Test Procedure

Run the GLM detection tests:

cd src && npx vitest run api/providers/utils/__tests__/glm-model-detection.spec.ts

The updated tests verify:

✅ GLM-4.6 base model has supportsThinking: true
✅ GLM-4.6V and its variants have supportsThinking: true
✅ GLM-4.7 variants continue to have supportsThinking: true
✅ GLM-4.5 variants correctly have supportsThinking: false

Impact

For Z.ai provider users:

GLM-4.6 and GLM-4.6V models will now properly send the thinking parameter
Users can control thinking mode via the reasoning effort settings
Thinking content will be preserved in conversation history

For LM Studio and OpenAI-compatible endpoint users:

The detection logic in PR fix: combine tool call sync fix with GLM model detection for issue #11071 #11096 will now correctly identify GLM-4.6 and GLM-4.6V as thinking-capable models
Console logs will show supportsThinking: true for these models

Pre-Submission Checklist

Issue Linked: This PR is linked to Issue [BUG] GLM4.5 via LMStudio as well as via an OpenAI-compatible endpoint stuck repeating file reads #11071
Scope: Changes are focused on adding thinking support to GLM-4.6/4.6V
Self-Review: Performed thorough self-review of code changes
Testing: Updated tests to cover GLM-4.6/4.6V thinking support
Documentation Impact: No documentation updates required (internal API changes)
Contribution Guidelines: Read and agree to guidelines

Screenshots / Videos

N/A - Backend feature, no UI changes.

Documentation Updates

No documentation updates are required.

Additional Notes

This PR builds on the GLM detection work in PR #11096 and addresses the specific feedback from @mark-ucalgary regarding GLM-4.6 and GLM-4.6V thinking support as documented by Z.ai.

Feedback and guidance are welcome!

Important

Adds thinking mode support for GLM-4.6 and GLM-4.6V models, updating model definitions, detection logic, provider implementations, and tests.

Behavior:
- Adds thinking mode support for GLM-4.6 and GLM-4.6V models in zai.ts, glm-model-detection.ts, and zai.ts.
- Updates GLM detection logic in glm-model-detection.ts to include GLM-4.6 and GLM-4.6V.
- Modifies createStream in base-openai-compatible-provider.ts to handle thinking mode for GLM models.
Tests:
- Adds test cases for GLM-4.6 and GLM-4.6V thinking support in glm-model-detection.spec.ts.
- Reorganizes tests to distinguish models with and without thinking support.
Misc:
- Updates comments in zai.ts to reflect new model support.
- Adjusts NativeToolCallParser to handle tool call synchronization more robustly.

^{This description was created by}^{for 23256cd. You can customize this summary. It will automatically update as commits are pushed.}

…1071 This PR combines: 1. PR #11093 fix: NativeToolCallParser processFinishReason hasStarted check 2. GLM model detection utility for LM Studio and OpenAI-compatible providers 3. mergeToolResultText optimization for GLM models 4. Disable parallel_tool_calls for GLM models 5. GLM-4.7 thinking parameter support 6. Diagnostic logging for GLM detection Closes #11071

This commit adds GLM model detection and related optimizations to the OpenAI Compatible provider (OpenAiHandler), which handles the "OpenAI Compatible" option in the UI when users set a custom base URL. Changes: - Import GLM detection utilities and Z.ai format converter - Add glmConfig property to track GLM model configuration - Detect GLM model on construction when model ID is available - Re-detect GLM model in createMessage if model ID changes - For GLM models: - Use convertToZAiFormat with mergeToolResultText to prevent conversation flow disruption - Disable parallel_tool_calls as GLM models may not support it - Add thinking parameter for GLM-4.7 models - Add console logging for detection results and applied optimizations This addresses the feedback in issue #11071 where the user clarified they are using the "OpenAI Compatible" provider, not the "OpenAI" provider. The previous changes in this PR only affected the base-openai-compatible-provider.ts (used by Z.ai, Groq, etc.) and lm-studio.ts, but not openai.ts which handles OpenAI Compatible.

- Add supportsReasoningEffort to GLM-4.6 and GLM-4.6V model definitions - Update GLM detection logic to recognize GLM-4.6/4.6V thinking support - Update Z.ai provider to handle thinking mode for all GLM models with reasoning - Update tests to reflect GLM-4.6/4.6V thinking support Addresses issue #11071 where Z.ai documentation shows GLM-4.6 and GLM-4.6V support thinking mode, but this was not reflected in the codebase.

roomote · 2026-01-30T16:24:10Z

Rooviewer See task on Roo Cloud

Review complete. Found 2 issues to address:

Inconsistency between GLM detection and model definitions: glm-4.6v-flash and glm-4.6v-flashx are marked as supporting thinking in glm-model-detection.ts but don't have supportsReasoningEffort in zai.ts
Log messages say "GLM-4.7 thinking mode" but now apply to GLM-4.6 models as well (in base-openai-compatible-provider.ts, lm-studio.ts, and openai.ts)

_{Mention @roomote in a comment to request specific changes to this pull request or fix all unresolved issues.}

roomote · 2026-01-30T16:29:34Z

src/api/providers/utils/glm-model-detection.ts

+	// GLM-4.6, GLM-4.6V, and GLM-4.7 have built-in thinking support
+	// For GLM-4.6, only the base model and vision variants support thinking
+	const supportsThinking =
+		version === "4.7" ||
+		(version === "4.6" &&
+			(variant === "base" || variant === "v" || variant === "v-flash" || variant === "v-flashx"))


This detection includes v-flash and v-flashx variants as supporting thinking, but the model definitions in zai.ts don't include supportsReasoningEffort for glm-4.6v-flash or glm-4.6v-flashx. This inconsistency means the Z.ai provider (which checks Array.isArray(info.supportsReasoningEffort)) will correctly exclude these models from thinking mode, but LM Studio and OpenAI Compatible providers will incorrectly try to send the thinking parameter to them. Either add supportsReasoningEffort to the flash variants in zai.ts if they truly support thinking, or remove v-flash and v-flashx from this detection logic.

_{Fix it with Roo Code or mention @roomote and request a fix.}

roomote · 2026-01-30T16:29:55Z

src/api/providers/base-openai-compatible-provider.ts

+		if (this.glmConfig.isGlmModel && this.glmConfig.supportsThinking) {
+			const useReasoning = this.options.enableReasoningEffort !== false // Default to enabled for GLM-4.7
+			;(params as any).thinking = useReasoning ? { type: "enabled" } : { type: "disabled" }
+			console.log(`[${this.providerName}] GLM-4.7 thinking mode: ${useReasoning ? "enabled" : "disabled"}`)


The log message says "GLM-4.7 thinking mode" but this code now also executes for GLM-4.6 and GLM-4.6V models. Consider updating to something like "GLM thinking mode" or using this.glmConfig.displayName for accuracy. The same issue exists in lm-studio.ts and openai.ts.

Suggested change

console.log(`[${this.providerName}] GLM-4.7 thinking mode: ${useReasoning ? "enabled" : "disabled"}`)

console.log(`[${this.providerName}] ${this.glmConfig.displayName} thinking mode: ${useReasoning ? "enabled" : "disabled"}`)

_{Fix it with Roo Code or mention @roomote and request a fix.}

roomote added 3 commits January 30, 2026 01:09

github-project-automation bot added this to Roo Code Roadmap and Roo Code Roadmap Jan 30, 2026

github-project-automation bot moved this to Triage in Roo Code Roadmap Jan 30, 2026

github-project-automation bot moved this to New in Roo Code Roadmap Jan 30, 2026

roomote bot mentioned this pull request Jan 30, 2026

[BUG] GLM4.5 via LMStudio as well as via an OpenAI-compatible endpoint stuck repeating file reads #11071

Open

roomote bot commented Jan 30, 2026

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat: add thinking mode support for GLM-4.6 and GLM-4.6V models #11114

feat: add thinking mode support for GLM-4.6 and GLM-4.6V models #11114

Uh oh!

roomote bot commented Jan 30, 2026 •

edited by ellipsis-dev bot

Loading

Uh oh!

roomote bot commented Jan 30, 2026 •

edited

Loading

Uh oh!

roomote bot Jan 30, 2026

Uh oh!

roomote bot Jan 30, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

	console.log(`[${this.providerName}] GLM-4.7 thinking mode: ${useReasoning ? "enabled" : "disabled"}`)
	console.log(`[${this.providerName}] ${this.glmConfig.displayName} thinking mode: ${useReasoning ? "enabled" : "disabled"}`)

feat: add thinking mode support for GLM-4.6 and GLM-4.6V models #11114

Are you sure you want to change the base?

feat: add thinking mode support for GLM-4.6 and GLM-4.6V models #11114

Uh oh!

Conversation

roomote bot commented Jan 30, 2026 • edited by ellipsis-dev bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Related GitHub Issue

Description

Changes Made

1. Model Definitions (packages/types/src/providers/zai.ts)

2. GLM Detection (src/api/providers/utils/glm-model-detection.ts)

3. Z.ai Provider (src/api/providers/zai.ts)

4. Tests (src/api/providers/utils/__tests__/glm-model-detection.spec.ts)

Test Procedure

Impact

Pre-Submission Checklist

Screenshots / Videos

Documentation Updates

Additional Notes

Uh oh!

roomote bot commented Jan 30, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

roomote bot Jan 30, 2026

Choose a reason for hiding this comment

Uh oh!

roomote bot Jan 30, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

roomote bot commented Jan 30, 2026 •

edited by ellipsis-dev bot

Loading

1. Model Definitions (`packages/types/src/providers/zai.ts`)

2. GLM Detection (`src/api/providers/utils/glm-model-detection.ts`)

3. Z.ai Provider (`src/api/providers/zai.ts`)

4. Tests (`src/api/providers/utils/tests/glm-model-detection.spec.ts`)

roomote bot commented Jan 30, 2026 •

edited

Loading