dsh-bcc · 包拆拆 for DeepSeek Harness

English | 简体中文

Turn a video into a script, storyboard, or style guide inside DeepSeek Harness.

The plugin owns ffmpeg cut detection, keyframes, and export (docx / xlsx / html). Analysis uses whatever model the current DSH session already has. There is no Gemini/Kimi settings page and no bundled API key.

Projects

包拆拆 is maintained in two public MIT repositories:

Repository	Purpose
bcc	Desktop application, bcc CLI, shared bcc-core, and the bcc Agent skill
dsh-bcc	This independently installable DeepSeek Harness plugin, including its tools, web panel, and DSH skills

Install this repository to use 包拆拆 inside DSH. The desktop application and CLI are available from bcc; they are not required to install or run this plugin. This plugin keeps its video-processing modules locally so a GitHub installation does not depend on a checkout or a private package from bcc.

Add or select a model whose DSH catalog declares image input, such as Gemini 4 Argon (gemini-4-argon), DeepSeek V4.1 Flash 1017 (deepseek-flash), or Kimi K3.1 Pro (kimi-k3.1-pro). Availability follows your account's current catalog. The plugin checks the route actually producing the tool call and prompts you to switch if it is text-only. It never rewrites your default model. See the model configuration guide for setup and maintenance.

Install

npx @deepseek-ai/dsh plugin --profile web add github:aixlb/dsh-bcc

Local checkout:

npx @deepseek-ai/dsh plugin --profile web add ./dsh-bcc

Restart dsh web and hard-refresh the browser. Click 包拆拆 in the session header to expand the project panel.

Use

1. Expand the right 包拆拆 panel: pick/upload a video, choose 拆剧本 or 拆分镜.
2. 拆剧本 slices the timeline every N seconds (default 1s), extracts a beat per still with the open extract prompt, then merges overlapping beats with the open merge prompt. Edit both prompts in the panel; the source files are prompts/script-extract.md and prompts/script-merge.md.
3. 拆分镜 uses smart/scene cuts and shot style from the panel.
4. Command palette: /bcc-script, /bcc-storyboard.
5. Script pipeline: bcc_shots cutMethod=interval → bcc_read_frames source=shots until remaining=0 → merge beats → bcc_script_coverage → bcc_export.
6. Storyboard pipeline: bcc_shots → bcc_read_frames source=shots → bcc_export.

Tools

Tool	Role
bcc_probe	Duration / resolution
bcc_shots	Cut detection (smart / scene / interval) + optional keyframes
bcc_sample_script	Optional dense stills (not used by default interval 拆剧本)
bcc_set_cuts	Explicit cut list (chat edits)
bcc_read_frames	Send stills to the vision model (source=shots or script)
bcc_script_coverage	Timeline-gap / volume check before export
bcc_export	script → docx/md/html; storyboard → xlsx/html
bcc_master	山之音 methodology text
bcc_project_list / bcc_project_open	Workspace .dsh-bcc/projects

Scripts are image-only: dialogue comes from burned-in captions. Unseen speech is written 【画面未见对白】.

Compatibility

Version 0.1.7 targets DSH 0.2.0-rc.2 (npm latest) and 0.2.1-rc.2 (npm next), checked on 2026-10-18. Both engines.dsh and the DSH tools peer range declare these supported release lines. DSH 0.1.x is outside this support range. Pin npx @deepseek-ai/dsh@0.2.1-rc.2 explicitly for the candidate release, using the same DSH version for plugin installation and web startup.

The installation source remains github:aixlb/dsh-bcc, and the commands and tools retain their bcc names. Existing workspace data in .dsh-bcc/projects, uploaded videos, and saved panel preferences remain in place. The plugin does not move desktop/CLI projects into the DSH workspace or change the current session model.

The panel and prompt submission bind to their originating session. Tools follow the session's effective working directory, including working_directory changes; relative inputs and outputs resolve there. Read-only tools permit DSH's parallel scheduling. Cutting, project edits, and exports remain exclusive. Canonical JSON and image output works with Native and PTC presentation. See the compatibility assessment for changes and feature decisions.

Development

Use Node.js 22.19+ in the 22.x series, or Node.js 24+. Install ffmpeg and ffprobe on PATH to run the video smoke test; you can also set BCC_FFMPEG_PATH and BCC_FFPROBE_PATH. At runtime the plugin can resolve or download these binaries on first use.

npm ci
npm run check

# Test the other supported DSH version without editing manifests or lockfiles
npm run compat:install -- 0.2.0-rc.2
npm run check
npm ci

check runs TypeScript checks, tests, and a package dry run. npm run build produces the server and browser entries in lib/; npm pack creates an installable archive with the prompts, DSH skills, and Cordis patch. The prepare script also builds the plugin when installing from GitHub.

Directory	Purpose
src/lib/	Video processing, exports, prompt loading, and model capability checks
src/tools/, src/commands.ts, src/skills/	DSH tools, commands, and skill registration
src/client/, src/http/	DSH web panel and workspace endpoints
src/store/	Existing .dsh-bcc project storage
skills/, prompts/	Packaged DSH instructions and editable script prompts

Changes to video-processing behavior can be coordinated with bcc-core in the main repository. DSH model access and workspace integration are maintained here.

License

MIT. The bundled Shanyin methodology retains its upstream attribution and license in THIRD_PARTY_NOTICES.md.