openclaw - 💡(How to fix) Fix Beta blocker: ollama - qwen3.5:9b produces garbage output under full agent system prompt [1 participants]

Bug type

Behavior bug (incorrect output/state without crash)

Beta release blocker

Yes

Summary

When OpenClaw sends a full agent system prompt (~42,000 chars) to ollama/qwen3.5:9b, the model produces repetition loops instead of a coherent response; confirmed across multiple sessions on v2026.4.12 and v2026.4.22.

Steps to reproduce

1. Configure OpenClaw with a full agent workspace (workspace files totaling ~27k chars: SOUL.md, AGENTS.md, USER.md, IDENTITY.md, TOOLS.md, HEARTBEAT.md, BOOTSTRAP.md)
2. Set agents.defaults.model.primary to ollama/qwen3.5:9b
3. Configure remote Ollama at http://host:11434 with qwen3.5:9b Q4_K_M
4. Start a new session and send any message (e.g. "hello")
5. Observe garbage output after 2-3 minute response time

Expected behavior

ollama/qwen3.5:27b produces coherent responses under identical workspace and tool configuration on the same hardware; qwen3.5:9b should behave similarly.

Actual behavior

Model produces repetition loops after 2-3 minutes: HeyHeyHeyHeyHeyHeyHeyHeyHeyHey... NONONONONONOhelloNONONONONONONOHELLONONO...

Ollama logs show first request returns 500 after 15s (race condition on load), second request returns 200 after 2m48s with garbage content.

[GIN] 2026/04/24 05:33:09 | 500 | 15.404728969s | POST "/api/chat" [GIN] 2026/04/24 05:36:00 | 200 | 2m48s | POST "/api/chat"

OpenClaw version

2026.4.22 (also reproduced on 2026.4.12)

Operating system

Ubuntu 24.04 (running as Unraid VM)

Install method

npm global

Model

ollama/qwen3.5:9b

Provider / routing chain

openclaw -> ollama (remote host 192.168.40.22:11434)

Additional provider/model setup details

Remote Ollama instance at http://192.168.40.22:11434 (no /v1) Model: qwen3.5:9b Q4_K_M, 9.7B parameters GPU stack: RTX 3090 24GB + RTX 3080 10GB + RTX 3060 12GB + RTX 3070 8GB = 54GB VRAM Model loads successfully: all 33 layers on GPU, 21.7GB VRAM, context_length 262144 agents.defaults.experimental.localModelLean: true was tried — did not resolve

Logs, screenshots, and evidence

Impact and severity

Affected: Any user running qwen3.5:9b with a full agent workspace Severity: High — model is completely unusable for agent sessions Frequency: 100% reproducible (5/5 attempts failed) Consequence: Agent session cannot start; falls back to garbage output or hangs

Additional information

System prompt breakdown at time of failure:

• Total: ~42,562 chars
• Project context (workspace files): ~27,152 chars
• Tool schemas: ~14,143 chars
• Skills prompt: ~3,747 chars

qwen3.5:27b handles identical prompt correctly on same hardware. localModelLean: true attempted — did not resolve. No errors in Ollama logs during inference — failure is silent garbage output only.