Gemma 4 E4B Drifts Into Tool-Use Mode

This looks less like a pure model bug than a prompt-and-orchestration failure: the model is being given enough freedom to jump from code understanding into agentic behavior. Gemma’s own docs say system-level instructions should be embedded in the initial user prompt, so the exact prompt contract matters a lot.

• –If your wrapper exposes write tools, the model will often try to use them unless you hard-disable those tools for inspection tasks.
• –Put a strict contract in the first prompt: describe only, no suggestions, no changes, no tool calls, no planning.
• –Add output validation and reject action verbs or file-write language when the job is read-only.
• –Code files can trigger “helpful assistant” priors, so the safer fix is permissioning plus prompt discipline, not just more instruction text.