Introduction

DSH’s plugin mechanism allows agents to connect to local tools on demand. For scenarios where agents need to produce images, speech, music, or SFX, a common issue is the inconsistency of the same character or voice line across multiple calls, making it difficult to immediately listen to and verify the generated results.

Deepseek-Continuity is a DSH plugin maintained by linxuhao with an MIT license. It provides local generation for images, speech, music, and SFX, WAV transcription for listening back, and cross-call identity consistency; failed generations are not treated as successful returns.

Below, its positioning is introduced first, followed by installation and typical usage.

Visual Identity and Voice Lines

Fixed Visual Identity:

  • create_character, create_animal, create_object, or import_subject fixes a character, animal, or prop.
  • Subsequently, call subject_image to generate an image that maintains that identity.

Fixed Voice Lines:

  • create_actor or import_actor fixes a voice line.
  • Subsequently, call actor_tts to generate dialogue.

The focus of these calls is cross-call consistency: fix the identity first, then generate specific content.

One-time Generation and Playback

One-time Generation:

  • generate_image is used for single image generation.
  • generate_speech is used for single speech generation.

Playback and Post-processing:

  • transcribe transcribes the generated WAV back to text.
  • remove_bg and slice_sheet are used for post-processing.

transcribe is suitable for reading the generated WAV back to text and comparing it with expected dialogue or a script.

Music, SFX, Backend, and State

Music Generation:

  • Uses Stable Audio.
  • Max 120 seconds.
  • No loop points.

SFX Generation:

  • Procedural sfxr.
  • No model.
  • No VRAM.

Backend and State:

  • Can interface with OpenAI-compatible /v1/images/* and /v1/audio/transcriptions backends.
  • continuity_status is used to query the plugin status.
  • Local lazy-loading models, so the GPU is unaffected when idle.
  • Failed generations are not treated as successful returns.

Installation and Activation

dsh plugin calls pnpm. If pnpm is not yet installed, execute first:

corepack enable pnpm

Configure the backend:

uvx --from dsh-continuity continuity-setup

You can also run from source code:

uvx --from git+https://github.com/linxuhao/Deepseek-Continuity continuity-setup

Add the plugin to an existing app bundle profile:

dsh plugin --profile <your-profile> add dsh-plugin-continuity

The PyPI package is dsh-continuity; do not use the unrelated continuity-mcp.

Typical Usage

  1. Generate image after fixing a character:

First call create_character, then call subject_image to generate an image maintaining the same character.

  1. Generate dialogue after fixing a voice line:

First call create_actor, then call actor_tts line-by-line to generate WAV for the same voice line.

  1. Playback and check:

Call transcribe to read the generated WAV back and compare it with the dialogue.

Suitable Scenarios and Notes

Suitable for scenarios in DSH where agents need to generate local images, speech, music, and SFX, and require fixed characters, animals, props, or voice lines.

Notes to be aware of:

  • Style requests may only partially take effect; reference images dominate the output style.
  • Clothing/equipment may stack rather than replace.
  • The plugin runs with the current dsh process permissions; check the source code and license before installing.
  • License is MIT.

Repository: https://github.com/linxuhao/Deepseek-Continuity