DSH Image Transcription Plugin

中文 · Upstream: DeepSeek Harness · Topic

A source release for the DeepSeek Harness image-transcription feature. It is implemented as two Cordis plugins:

• @deepseek-ai/dsh-image-transcription converts image-bearing messages into durable visual transcriptions when the primary model is text-only.
• @deepseek-ai/dsh-tool-image-recall registers recall_image, allowing the primary model to ask the configured vision model to re-read a cited historical image.

Features

• Stable transcriptionId references bind text transcriptions to their original attachments.
• Compaction adds a compact image index; original attachments remain stored in the session log.
• recall_image sends only the cited image batch and the current question to the visual route.
• The Web UI shows elapsed transcription status and retains clear failures.

Configure the image-transcription model

This plugin requires an image-transcription model configuration before it can process images.

1. Open Settings → Models.
2. In Image transcription model, select the visual provider and model to use when the primary model is text-only.
3. Expand Configuration and enter the provider API key, endpoint, protocol, and available model ID.
4. Save the configuration.

License and attribution

This release contains modifications to DeepSeek Harness, licensed under MIT. Keep the included LICENSE and retain upstream attribution in derivative distributions.