[herjarsa]

Description

Fixes #2637

Kimi K2 models enter an infinite loop when using tool calls through OmniRoute because injectEmptyReasoningContentForToolCalls only handled DeepSeek V4 models, leaving Kimi K2 and other reasoning models without reasoning_content when tool_calls were present.

Root Cause

The injectEmptyReasoningContentForToolCalls function used isDeepSeekReasoningModel() to decide whether to inject empty reasoning_content for assistant messages with tool_calls. This only covered DeepSeek V4, but Kimi K2, Qwen-Thinking, GLM, and MiMo models also require reasoning_content to be passed back on subsequent turns.

When these models don't receive their reasoning_content back, they attempt to regenerate the response, generating the same tool_call again, causing an infinite loop.

Changes

• open-sse/translator/helpers/schemaCoercion.ts:

  • Replaced isDeepSeekReasoningModel() with requiresReasoningReplay() to cover all reasoning replay models
  • Now handles: DeepSeek V4, Kimi K2, Qwen-Thinking, GLM, MiMo, and any future models added to REASONING_REPLAY_MODEL_PATTERNS

• tests/unit/schema-coercion.test.ts:

  • Updated existing test to use a non-reasoning model (deepseek-v3) instead of deepseek-reasoner (which now correctly receives injection)
  • Added tests for Kimi K2 models via moonshot provider
  • Added test for Kimi K2 via Qoder provider (common routing scenario)
  • Added comprehensive test covering all reasoning replay providers
  • Added test verifying non-reasoning models are correctly skipped

Test Coverage

Tests verify that:

• DeepSeek V4 models via proxy providers (openrouter, fireworks, deepinfra) receive reasoning_content: ""
• Kimi K2, K2.5, K2.6, K2-thinking receive reasoning_content: ""
• Kimi K2 via Qoder provider receives reasoning_content: ""
• All reasoning replay providers (deepseek, opencode-go, siliconflow, nebius, deepinfra, sambanova, fireworks, together, xiaomi-mimo) receive reasoning_content: ""
• Non-reasoning models (gpt-4, claude, gemini, llama) do NOT receive injection

Related

• Issue [BUG] DeepSeek V4 thinking mode fails after tool calls because reasoning_content is not preserved/replayed #1628 (reasoning replay system)
• docs/routing/REASONING_REPLAY.md
• open-sse/services/reasoningCache.ts (defines requiresReasoningReplay())

Checklist

• Bug fix (non-breaking change which fixes an issue)
• Tests added/updated
• Code follows project style guidelines
• Changes are minimal and focused on the fix

[kilo-code-bot]

Kilo Code Review could not run — your account is out of credits.

Add credits or switch to a free model to enable reviews on this change.

[gemini-code-assist]

Code Review

This pull request refactors the reasoning content injection logic to use the requiresReasoningReplay helper, extending support to models like Kimi K2 and various third-party providers. The test suite has been significantly expanded to validate these additions. Review feedback identifies a critical syntax error in the test file due to orphaned code, an unused import in the schema coercion helper, and a recommendation to use unique test names for better maintainability.

[gemini-code-assist]

This change introduces a syntax error in the test file. While a new test block was added, the header for the previous test (formerly at line 134) was removed, but its body (lines 145-154) was left dangling. This results in orphaned code and an unmatched closing brace }); at line 154, which will prevent the test suite from running.

Please remove the redundant code block from lines 145 to 154 to restore the file's integrity.

[gemini-code-assist]

The import isDeepSeekReasoningModel is no longer used in this file. It was replaced by requiresReasoningReplay in the injectEmptyReasoningContentForToolCalls function. Removing unused imports helps maintain code cleanliness.

Suggested change

	import { isDeepSeekReasoningModel, requiresReasoningReplay } from "../../services/reasoningCache.ts";
	import { requiresReasoningReplay } from "../../services/reasoningCache.ts";

[gemini-code-assist]

This test shares the same name as the one defined on line 134. Duplicate test names can lead to confusing test reports and make it harder to identify which specific test failed. Consider using a more descriptive name for this comprehensive test case.

test("injectEmptyReasoningContentForToolCalls skips various non-reasoning models and providers", () => {

[diegosouzapw]

Muito obrigado pela contribuição, @herjarsa! A correção para o loop infinito de tool calls com o Kimi K2 foi integrada na release v3.8.3. Fizemos uma pequena limpeza em tests/unit/schema-coercion.test.ts para remover um trecho órfão remanescente. Seu trabalho é muito valioso para o OmniRoute! 🚀