openclaw - 💡(How to fix) Fix [Bug]: Telegram image understanding intermittently fails for ollama/qwen3.5:cloud, including false image hallucinations [1 comments, 2 participants]

Bug type

Regression (worked before, now fails)

Beta release blocker

No

Summary

Telegram images sent to an Ollama cloud vision model are intermittently routed as text/fetch paths instead of multimodal vision content blocks, causing the agent to either deny image capability or hallucinate confident but false image descriptions.

Steps to reproduce

1. Configure an OpenClaw agent with ollama/qwen3.5:cloud as primary model (declared input: text, image)
2. Set agents.defaults.imageModel.primary to openai-codex/gpt-5.4
3. Start a fresh session (/new) with that agent via Telegram
4. Send one Telegram image (e.g. a photo)
5. Ask the agent: "Describe this image"
6. Observe: agent either claims it cannot view images, or produces a confident but false description unrelated to the actual image content

Expected behavior

When I send an image in Telegram to an agent using ollama/qwen3.5:cloud, the model should receive actual image input and either:

• describe the image correctly, or • clearly state that image input was unavailable

It should not fabricate image contents.

Actual behavior

In fresh sessions, sending a Telegram image and asking “describe this image” produces intermittent failures:

1. Sometimes the agent says it cannot view images at all
2. Sometimes it hallucinates a false description of the image

Example false result:

• Actual image: photo of a young woman in a winery • Returned description: monochrome historical architectural scene / ruins / archaeological site

OpenClaw version

OpenClaw: 2026.4.15

Operating system

macOS arm64, Mac Mini M4 16GB

Install method

No response

Model

ollama/qwen3.5:cloud

Provider / routing chain

Telegram Bot API → OpenClaw gateway → Ollama provider → ollama/qwen3.5:cloud (remote) Image media path (observed failure): Telegram inbound → OpenClaw media download/resize → file saved to local disk → incorrectly routed as text/fetch path (web_fetch with file:/// URI) instead of multimodal vision content block

Additional provider/model setup details

• Agent: ollama/qwen3.5:cloud primary, ollama/qwen3.5:9b local fallback • Global image model override: openai-codex/gpt-5.4 • OpenClaw version: 2026.4.15 (041266a) • Host: macOS arm64, Mac Mini M4 16GB • qwen3.5:cloud is declared as input: ["text", "image"] in both global and per-agent model configs • Same setup produced working vision results in prior sessions, then failed again in fresh sessions • No per-agent image model override set; only the global agents.defaults.imageModel.primary

Logs, screenshots, and evidence

Impact and severity

• Affected users: Any OpenClaw user routing Telegram images to an Ollama cloud model for vision • Severity: Data risk. The model does not simply fail safe — it fabricates confident but false visual descriptions, which is worse than a silent failure • Frequency: Intermittent. Vision worked correctly in some sessions, then failed in fresh sessions under the same configuration • Consequence: Unreliable Telegram-based image interpretation; agents can hallucinate image contents instead of admitting missing input, creating a trust and safety problem for any workflow relying on vision

Additional information

No response