Aevatar Vision Probe
Use this to confirm an attached image actually reaches a workflow vision step and the model can see it. Vision works on the WORKFLOW path (an Image input part is auto-attached to the workflow llm_call); it does NOT work on the plain chat reply path. So this probe drives a workflow.
You (the agent) interpret the run output honestly. The expected pass is a real description of the image the user sent. "No image received" is a real FAIL signal, not something to paper over — never invent image content.
Prerequisite (one-time)
The probe workflow probe_vision_describe must be registered in this chat's scope (it ships in the
repo's workflows/ directory; submit it to your scope if /workflow run reports it missing).
How to run it
Tell the user to send the image and the trigger in ONE message (desktop Lark: paste the image into the input box, then type the command, send together):
/workflow run probe_vision_describe(alias/run-workflow probe_vision_describe). A pure-text command with no image, or an image in a separate message, will not carry the attachment into the run.Read the run output. The workflow returns JSON:
{"saw_image": true|false, "description": "...", "dominant_colors": "...", "text_in_image": "..."}.Verdict:
saw_image: true+ a description that matches the image the user sent →VISION OK— the read-image path works and the session LLM route is multimodal. Image-dependent workflows can run.saw_image: false/ "no image received" →VISION BLOCKED. Most likely causes, report which: the session LLM route is not multimodal (fix:PUT /api/user-configpreferredLlmRoute to a vision route), or the attachment did not reach the run.未找到 workflow/暂未绑定可运行的 actor→ the probe workflow isn't registered/bound in this scope; that's a setup step, quote it.
Output
One block: did the run start, did the workflow return JSON, what was saw_image, and the verdict
above. Quote any error verbatim.
Guardrails
- Read-only: this probe only describes an image; it never sends, stores, or forwards anything.
- Never claim to have seen image content unless the run output actually returned a description;
saw_image: falseis an honest, valid result. - Never ask the user for tokens — your NyxID-brokered tools handle all credentials.