Update on my OpenClaw + Ollama + Gemma local setup issue.

I tested the same strict document-review prompt directly in Ollama using `ollama run gemma4:e4b`, and the model followed the required structure much better.

Direct Ollama output used the requested headings:

```text

DRAFT — HUMAN REVIEW REQUIRED

1. Executive Summary

2. Clarity Issues

3. Missing Information

4. Risky or Overstated Claims

5. Operational Gaps

6. Suggested Improvements

7. Recommended Next Action

8. Final Review Status

```

But when I run a similar workflow through OpenClaw, the agent keeps drifting into freeform summaries and adds its own headings, even after I tell it not to.

So now it looks like the model can follow the template directly, but OpenClaw may be adding session/profile/runtime behavior that changes the output style.

Has anyone seen this with OpenClaw local agents?

I’m trying to figure out:

1. How to fully reset/clear OpenClaw session context

2. Whether the default agent profile is adding hidden behavior

3. Whether there is a better no-tools/local-only config

4. Whether I should create a separate strict “document review” agent instead of using the main agent

5. Whether OpenClaw has a setting for system prompts or fixed output templates

Current stack:

* Windows desktop

* Ollama working

* Gemma direct test working

* OpenClaw gateway reachable