Articles tagged "DeepSeek V4"

dsh-image-describer: Equipping DeepSeek Harness's Text-Only Models with Visual Understanding Capabilities

2026-09-19 DeepSeek Harness plugin DeepSeek Harness plugin dsh-image-describer DeepSeek V4 Visual Understanding Image Recognition

The `dsh-image-describer` plugin in DeepSeek Harness (DSH) intercepts the message stream and converts image content into structured text, enabling text-only LLMs to understand images. It supports dual-mode interaction, including pasting images into the chat box (attachment mode) and referencing local file paths in the conversation (Tool mode), and automatically performs OCR and content recognition. Before use, enable the image input channel for the text model in the configuration file and specify a multimodal vision model, such as MiniMax-M3, along with its API key.

Starfie1d1272/dsh-v4-anchor: DeepSeek V4 Standard First Request Anchoring and Promotion Recovery

2026-09-03 DeepSeek Harness plugin DeepSeek V4 dsh plugin dsh-v4-anchor Standard anchoring RL-Shaped

`Starfie1d1272/dsh-v4-anchor` is an RL anchoring patch for the DeepSeek Harness Standard plugin, specifically designed for the DeepSeek V4 model. Its core functionality is to constrain the session to a minimized RL-shaped surface (containing only `bash/pwsh` and `str_replace_editor`) before the first persistent `tool/call`, temporarily removing runtime contexts, the Skill Catalog, and `AGENTS.md` injection. Once the model completes its first tool call, the plugin automatically restores full Standard capabilities, including the original prompt, the full tool directory, and skills.

Miraculous Standard: DeepSeek V4 Pro/Flash Unified Anchored Presets

2026-08-27 DeepSeek Harness plugin DeepSeek DSH model inference Agent preset DeepSeek V4

Miraculous Standard is an agent preset provided by rinDBeans for DeepSeek Harness, unifying coverage across four combinations of V4 Pro/Flash with the official API and opencode-go. The initial request only exposes a Minimal toolset for trajectory anchoring. After upgrade, it unlocks the Pro resident set or Flash full directory based on the model, coupled with context-gate and epoch management to address compaction and resume anchoring issues. This article compiles installation commands, model configurations, and key items from agent.cordis.yml, while explaining applicable scenarios and known limitations.