A visible screenshot doesn't mean the model got it: a 3-step agent image-path diagnostic
derspenti · reddit · 2026-09-10
The author flags a common diagnostic blind spot: seeing a screenshot in the chat window doesn't prove the model actually received it. Using the dsh-univer-office plugin in DeepSeek Harness as an example, Office screenshot tools render documents and return images via an attachment store — a path requiring an available store and an image-capable model route. A browser preview and an image attached to the model's next request are different observations.
A minimal check with one slide and one screenshot:
- Did the screenshot operation produce an actual image attachment, not just a filename or text?
- Did the configured route carry the attachment into the model request?
- Can the model identify a visual detail not already in the surrounding text?
These steps separate a missing attachment from a model that received the image and misread it. The author stresses this is a proposed diagnostic, not a benchmark, and that local multimodal failures should be attributed to the pipeline — not the model — only after recording what actually reached it.
More from coding & agent
- User Flags opencode with Free Muse 1.3 Reading Entire Code Directories Despite Deny Permission — ConcernedIndInvestor · 2026-09-10
- MCP Server Design Question: Should Tool Metadata Be Internal Config or an Exposed Resource? — Sea-Letterhead-9529 · 2026-09-10
- SWE-Bench Pro Verified: leakage and reward hacking inflated agent scores, some models drop sharply — Shanghai-AI-Laboratory · 2026-09-10
- Qoder launches Sonus, a built-in model specializing in Computer Use and autonomous coding — Scobleizer · 2026-09-10
- Millie's ternary 35B MoE hits 56% on SWE-bench Verified, runs local coding agent on iPhone 17 Pro — MannyKayy · 2026-09-10
- Distilling Linus Torvalds' code review style into a Skill from 32k kernel emails — Mte90 · 2026-09-10