Description
After updating to LM Studio 0.4.19 Build 2, the Anthropic-compatible endpoint (/v1/messages) stopped working correctly with Claude Code when using Qwen models that have custom Jinja templates with thinking support.
Symptoms:
The model correctly processes very large prompts (16k+ tokens containing full Claude Code system reminders + dozens of tool schemas).
Prompt evaluation completes successfully.
The model then generates only a single sentence (~50-60 tokens) and stops.
Tool calling / agent loop in Claude Code fails to continue (it treats the short response as the end of the turn).
This regression appeared immediately after updating from 0.4.18 to 0.4.19 Build 2. Downgrading back to 0.4.18 fully restores normal behavior.
Steps to Reproduce
Load a Qwen3.5 model (in my case unsloth/Qwen3.5-122B-A10B-Q4_K_M) that uses a custom Jinja template with / reasoning support.
Point Claude Code CLI at LM Studio’s Anthropic endpoint (ANTHROPIC_BASE_URL=http://localhost:1234).
Send a complex agentic prompt that triggers heavy tool use (e.g. “analyze this project”).
Observe that the model only outputs one sentence and then stops, instead of making tool calls.
Environment
LM Studio version: 0.4.19 Build 2 (broken) → 0.4.18 (working)
Model: unsloth/Qwen3.5-122B-A10B-GGUF (Q4_K_M)
Chat Template: Custom Jinja template with explicit reasoning_content → handling and custom <tool_call> format.
Client: Claude Code CLI (Anthropic)
Endpoint used: Anthropic-compatible /v1/messages
OS: Windows
Relevant Logs
Prompt processing completes normally (~16k tokens processed at ~2900+ t/s).
Generation then stops after only 57 tokens with a clean Finished streaming Anthropic response.
Even after explicitly disabling thinking/reasoning mode in LM Studio, the issue persisted.
Suspected Cause
Changes introduced in 0.4.19 (particularly making “LM Studio Engine Protocol” default to ON and updates to reasoning content handling) appear to have broken how the Anthropic emulation layer assembles prompts for models using custom Jinja templates that include thinking support.
Workaround
Downgrade to LM Studio 0.4.18 — issue is immediately resolved.
Additional Context
This setup was working reliably before the 0.4.19 update. The model itself is capable — the problem is isolated to the Anthropic compatibility layer after the update.
Description
After updating to LM Studio 0.4.19 Build 2, the Anthropic-compatible endpoint (/v1/messages) stopped working correctly with Claude Code when using Qwen models that have custom Jinja templates with thinking support.
Symptoms:
The model correctly processes very large prompts (16k+ tokens containing full Claude Code system reminders + dozens of tool schemas).
Prompt evaluation completes successfully.
The model then generates only a single sentence (~50-60 tokens) and stops.
Tool calling / agent loop in Claude Code fails to continue (it treats the short response as the end of the turn).
This regression appeared immediately after updating from 0.4.18 to 0.4.19 Build 2. Downgrading back to 0.4.18 fully restores normal behavior.
Steps to Reproduce
Load a Qwen3.5 model (in my case unsloth/Qwen3.5-122B-A10B-Q4_K_M) that uses a custom Jinja template with / reasoning support.
Point Claude Code CLI at LM Studio’s Anthropic endpoint (ANTHROPIC_BASE_URL=http://localhost:1234).
Send a complex agentic prompt that triggers heavy tool use (e.g. “analyze this project”).
Observe that the model only outputs one sentence and then stops, instead of making tool calls.
Environment
LM Studio version: 0.4.19 Build 2 (broken) → 0.4.18 (working)
Model: unsloth/Qwen3.5-122B-A10B-GGUF (Q4_K_M)
Chat Template: Custom Jinja template with explicit reasoning_content → handling and custom <tool_call> format.
Client: Claude Code CLI (Anthropic)
Endpoint used: Anthropic-compatible /v1/messages
OS: Windows
Relevant Logs
Prompt processing completes normally (~16k tokens processed at ~2900+ t/s).
Generation then stops after only 57 tokens with a clean Finished streaming Anthropic response.
Even after explicitly disabling thinking/reasoning mode in LM Studio, the issue persisted.
Suspected Cause
Changes introduced in 0.4.19 (particularly making “LM Studio Engine Protocol” default to ON and updates to reasoning content handling) appear to have broken how the Anthropic emulation layer assembles prompts for models using custom Jinja templates that include thinking support.
Workaround
Downgrade to LM Studio 0.4.18 — issue is immediately resolved.
Additional Context
This setup was working reliably before the 0.4.19 update. The model itself is capable — the problem is isolated to the Anthropic compatibility layer after the update.