Troubleshooting

Agent Says It Saw Your Image (But It Didn’t)?

Fix false image-analysis replies fast: confirm channel capability, run a known-good test, and stop hallucinated screenshot summaries.

A recurring pain point in the community is this pattern: the agent confidently describes an image, but the description is wrong or generic. In plain terms, the model likely never received the image bytes in that session context.

Reality check: confident wording is not proof the image was visible. Always verify with a concrete image detail test.

Fast symptom checklist

Why this happens

Known-good 2-minute verification flow

  1. Start a fresh session (new thread/chat) in the same surface you plan to use.
  2. Upload one clean image with an obvious detail (e.g., a red circle in top-left).
  3. Ask: "What exact text is visible in the top line? What color is the top-left shape?"
  4. If answer is vague/wrong, stop task execution and fix attachment path first.

Do not continue production work until this test passes. Otherwise you'll get polished nonsense and lose trust in outputs.

Copy/paste prompt to force honesty

Before you analyze this image, do a visibility check: 1) Confirm whether you can directly access the image bytes in this message. 2) Quote one exact text snippet from the image. 3) Name one specific visual detail (color + location). If you cannot verify those, stop and say "IMAGE_NOT_VISIBLE".

Recovery steps if test fails

  1. Retry in a fresh session in the same platform surface.
  2. Use a smaller PNG/JPG and a simple filename.
  3. Post image + question in the same message turn.
  4. If still failing, test the same image in dashboard chat to isolate Discord/channel path issues.
  5. If dashboard works but Discord fails, treat as channel/integration config issue (not model intelligence).

Support-ready escalation packet

Image visibility failure report Surface used (dashboard / Discord DM / Discord channel): Channel/thread ID (if Discord): Image type + size: Prompt used: Expected result: Actual result: Did 2-minute verification pass? (yes/no): Does same image work in dashboard? (yes/no):

Best practice: require a "visibility proof" step before every screenshot-dependent workflow. It prevents silent failures and fake confidence.