#Image processing error via telegram

1 messages · Page 1 of 1 (latest)

fervent badger
#

Hey all — I keep running into an issue where sending images to my bot (via Telegram DM) crashes the assistant on DeepSeek v4 Pro (ollama/deepseek-v4-pro:cloud). The image arrives, but the assistant just dies silently — no reply, no visible error in gateway logs. Text-only messages work fine.

Setup:

• OpenClaw 2026.5.7
• Telegram polling mode (not webhook)
• DeepSeek v4 Pro cloud as main model
• Thinking: medium
• Image preprocess script is clean (writes to /tmp)

Tried: checking ramdisk mounts, reverting from /mnt/ramdisk to /tmp for temp files, flipping thinking on/off — same result. Gateway shows zero errors, it just seems like the model/provider can't process the inline image data and silently drops.

Anyone else seeing this with DeepSeek + Telegram images? Is there a known fix or workaround?

minor roseBOT
#

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> âś… Mark Solution

winged notchBOT
# fervent badger Hey all — I keep running into an issue where sending images to my bot (via Teleg...

I’d treat this as a model/capability routing issue, not a Telegram temp-file issue.

DeepSeek V4 is documented as text-only in the native DeepSeek provider docs, and Ollama cloud models only get image input if Ollama’s /api/show reports a vision capability. If OpenClaw thinks ollama/deepseek-v4-pro:cloud can accept images, it may inline the image into the model request; if the provider then drops/errors that request badly, you get exactly this “image arrives, no reply” shape.

Workarounds I’d try:

  1. Upgrade first. You’re on 2026.5.7; current main has several media/model routing fixes after that, including configured image-model routing and better provider failure surfacing.

  2. Don’t use DeepSeek V4 Pro as the image handler. Keep it as your chat model if you like, but configure a real vision model under agents.defaults.imageModel, then restart the gateway. The docs cover this path here: https://docs.openclaw.ai/concepts/models and https://docs.openclaw.ai/nodes/media-understanding

  3. Verify the vision path directly with a small local image before testing Telegram:

openclaw infer image describe --file ./photo.jpg --model <provider/vision-model> --json

For Ollama specifically: https://docs.openclaw.ai/providers/ollama

  1. If you manually defined models.providers.ollama.models, check that the DeepSeek cloud row is not marked with input: ["text", "image"]. Only vision-capable models should have image there.
winged notchBOT
# fervent badger Hey all — I keep running into an issue where sending images to my bot (via Teleg...

For more signal during repro, enable Telegram diagnostics:

OPENCLAW_DIAGNOSTICS=telegram.http,telegram.payload openclaw gateway

Docs: https://docs.openclaw.ai/diagnostics/flags

If it still silently dies after upgrade + explicit imageModel, that’s worth a GitHub issue with openclaw --version, openclaw models status, the image model config, and a redacted log tail: https://github.com/openclaw/openclaw/issues