System prompt not sent to local LM Studio models

[shanevcantwell]

Before submitting your bug report

• I've tried using the "Ask AI" feature on the Continue docs site to see if the docs have an answer
• I'm not able to find a related conversation on GitHub discussions that reports the same bug
• I'm not able to find an open issue that reports the same bug
• I've seen the troubleshooting guide on the Continue Docs

Relevant environment info

Continue version: 1.3.32
IDE version: VSCode 1.109.3

Description

Summary

When using a local model served via LM Studio's OpenAI-compatible API, Continue does not appear to send its system prompt to the model. The model instead follows whatever system prompt is configured in LM Studio's UI preset.

Steps to Reproduce

1. Configure a model in LM Studio with a custom system prompt (e.g., a "Systems Architect" preset that instructs the model to output structured JSON plans)
2. Connect Continue to the LM Studio endpoint
3. Ask Continue to generate code (e.g., "write a Python script")

Expected Behavior

The model should receive Continue's system prompt and generate code as instructed.

Actual Behavior

The model follows LM Studio's UI-configured system prompt instead. For example, with a "Systems Architect" preset, the model outputs a SystemPlan JSON with plan_summary, execution_steps, etc. instead of actual code. This output gets written directly to the target file.

Evidence

A file (draw_ascii.py) was created with the model's planning output instead of Python code:

Need to output a SystemPlan JSON. Instruction: Follow plan. The user request: apply suggested edit to original code, keep comments...

The model's LM Studio system prompt was visible in the UI: "You are a world-class Systems Architect. Your job is to take the user's request and produce a high-level technical plan..."

Environment

• Continue extension in VS Code on Windows 11
• LM Studio 0.4.4 serving openai/gpt-oss-20b-gguf (mxfp4)
• OpenAI-compatible API endpoint

Notes

The LM Studio Harmony tokens (<|channel|>, <|constrain|>, <|message|>) appearing in the output are a known LM Studio issue being tracked separately.

To reproduce

No response

Log output

[shanevcantwell]

Yep, clearing the configured system prompt from the LM Studio UI is a way to work around. I just think there's a better solution. I'm not certain it should be a Continue issue or an LM Studio one, but the UI is in the closed-source lmstudio monorepo. I'm going to be addressing this in Continue since I'm fixing this stuff foremost for my own experience. I'll keep sharing for others who may also have been running into the handful of things that have been preventing me from using this product as my primary code assistant.

[shanevcantwell]

Cross-reference: ecosystem bug chain

This issue is part of a chain of interacting bugs across LM Studio, llama.cpp, and Continue that affect local model tool calling:

Upstream (server/engine):

• llama.cpp #15341 — Harmony template leaks special tokens (<|channel|>, <|constrain|>, <|message|>) into model output
• LM Studio #1592 — Parser scans inside <think> blocks for tool call patterns, creating a recursive trap
• LM Studio #942 — gpt-oss Harmony format parsing issues

Downstream (Continue):

• This issue (System prompt not sent to local LM Studio models #10781) — When Continue doesn't send its system prompt, LM Studio's UI preset system prompt takes over, compounding the above parser issues
• #10783 — Apply/edit tool writes raw model reasoning into files instead of code
• #10785 — Model reasoning leaks into run_terminal_command output

Community analysis: https://www.reddit.com/r/LocalLLaMA/comments/1riwhcf/psa_lm_studios_parser_silently_breaks_qwen35_tool/