openclaw - 💡(How to fix) Fix [Bug]: Ollama local model ignores user prompt and summarizes OpenClaw runtime context instead [5 comments, 3 participants]

Bug type

Regression (worked before, now fails)

Beta release blocker

No

Summary

On OpenClaw 2026.4.26 in WSL2, local Ollama models work with raw ollama run, but in clean OpenClaw TUI sessions they ignore simple prompts like “What is your name?” and instead summarize runtime/context/tool metadata. /context detail shows workspace identity files are injected, and setting compat.supportsTools:false did not fix it.

Steps to reproduce

I asked Krill from discord what I should put here...

Use this for Steps to reproduce:

1. Install/run OpenClaw 2026.4.26 on Ubuntu under WSL2.
2. Configure Ollama as the model provider using native Ollama mode:
   • baseUrl: http://127.0.0.1:11434
   • api: ollama
   • apiKey: ollama-local
3. Pull a local Ollama model, e.g.:

bash ollama pull llama3.2

1. Allow/select the model in OpenClaw:

bash openclaw config set agents.defaults.models '{"ollama/llama3.2:latest":{}}' --strict-json --merge openclaw models set ollama/llama3.2:latest openclaw gateway restart 5. Start a fresh TUI session:

bash openclaw tui --session native-ollama-test

1. In the TUI, send:

text /model ollama/llama3.2:latest /think off /context detail What is your name? Answer from IDENTITY.md in one sentence.

1. Observe that /context detail shows AGENTS.md, SOUL.md, IDENTITY.md, USER.md, and MEMORY.md injected, but the model does not answer the question. Instead, it summarizes OpenClaw runtime/context/tool metadata.
2. Repeat with compat.supportsTools:false configured for the Ollama model; behavior still occurs.

Expected behavior

The assistant should answer the user prompt using the injected workspace identity, e.g. “My name is …”, instead of summarizing runtime/context metadata.

Actual behavior

The assistant ignores the user prompt and responds as if the runtime/context metadata is the task, e.g. by explaining JSON/tool/runtime context or saying no actionable request was provided.

OpenClaw version

OpenClaw 2026.4.26 (be8c246)

Operating system

Windows 11 host, Ubuntu 22.04 under WSL2 Linux kernel shown by OpenClaw: 6.6.87.2-microsoft-standard-WSL2 (x64) Node: 22.22.2

Install method

npm global install Gateway running as systemd user service: openclaw-gateway.service Command: /usr/bin/node /home/peter/.npm-global/lib/node_modules/openclaw/dist/index.js gateway --port 18789

Model

ollama/gemma4:latest ollama/llama3.2:latest

Provider / routing chain

OpenClaw TUI -> local OpenClaw Gateway ws://127.0.0.1:18789 -> native Ollama provider -> local Ollama daemon at http://127.0.0.1:11434 Ollama config: baseUrl: http://127.0.0.1:11434 api: ollama apiKey: ollama-local No /v1 endpoint, no proxy, no remote model router.

Additional provider/model setup details

No response

Logs, screenshots, and evidence

Apparently I can't upload the zip file for the exported diagnostics, but I have it.  Here's the summary at least.

# OpenClaw Diagnostics Export

Attach this zip to the bug report. It is designed for maintainers to inspect without asking for raw logs first.

## Generated

Generated: 2026-04-29T19:27:45.248Z
OpenClaw: 2026.4.26

## Contents

- included latest stability bundle (0 event(s))
- sanitized log tail (974 line(s), inspected 1061828 byte(s), raw messages omitted)
- config shape included (parsed)
- gateway status snapshot included (status/gateway-status.json)
- gateway health snapshot failed (gateway closed (1008): pairing required: device is asking for more scopes than currently approved (requestId: b741560b-514f-4470-98fb-e57d2cbf5
Gateway target: ws://127.0.0.1:18789
Source: local loopback
Config: $OPENCLAW_STATE_DIR/openclaw.json
Bind: loopback)

## Maintainer Quick Read

- `manifest.json`: file inventory and privacy notes
- `diagnostics.json`: top-level summary of config, logs, stability, status, and health
- `config/sanitized.json`: config values with credentials, private identifiers, and prompt text redacted
- `status/gateway-status.json`: sanitized service/connectivity snapshot
- `health/gateway-health.json`: sanitized Gateway health snapshot
- `logs/openclaw-sanitized.jsonl`: sanitized log summaries and metadata
- `stability/latest.json`: newest payload-free stability bundle, when available

Impact and severity

I was in the middle of a coding project and the when context got over 80% i asked my agent for a new session to get a fresh context. From that point on, my agent seems to be missing (even though it's loaded) and I can only talk directly to the model instead.

Additional information

I think the last good version was 4.22, but I'm not sure. I updated to see if it would help. It did not.