視覺

這是什麼

暴露給 agent 的圖片理解能力。

白話解釋

Hermes 無法自己看一眼圖片，所以當它需要知道圖片裡面有什麼時，它會尋求外部幫助。它會下載圖片，交給能夠實際看圖的路徑，然後得到特定問題的答案——這個螢幕截圖寫了什麼、這個圖表中有什麼——而不是只憑文字描述來猜測。

為什麼重要

能夠實際查看圖像，而不僅僅是推理關於圖像的文字描述，這就是視覺工具集所增加的功能——註冊為一個單一的 vision_analyze 工具。它在研究型工作流程中補充了網路搜尋和檔案工具：擷取螢幕截圖或圖表，然後直接分析，而不是依賴周圍的文字來描述其內容。並非每個設定的模型都支援圖像輸入，這使得這項能力既受工具集配置限制，也受模型選擇限制。

How it works

vision_analyze takes an image URL or a local file path, downloads and normalizes it — including converting SVGs to PNG — before deciding how to answer (tools/vision_tools.py). When the active main model’s provider natively accepts images inside tool results, Hermes takes a fast path: the raw image bytes get handed straight back so the main model sees the picture itself on its next turn, no extra model call involved. When it doesn’t, Hermes falls back to a separate auxiliary vision model — Gemini 3 Flash Preview via OpenRouter by default — which describes the image in text, and that description is what the main model actually reads. Downloaded images are cleaned up from a temporary cache afterward; local file paths are left untouched since they weren’t downloaded in the first place.

A concrete example

Someone pastes a link to a screenshot of a cryptic error dialog and asks what’s wrong.

1. Hermes calls vision_analyze with the image URL and the question.
2. If the main model can see images directly, it reads the screenshot itself on the next turn; otherwise the auxiliary vision model describes the error dialog’s contents in text.
3. Either way, the main model gets the information it needs without the user having to manually describe what the screenshot shows.