@dsh-cc/output-styles

@dsh-cc/output-styles

English | 中文

Claude Code-compatible output style selection for the DeepSeek Harness. The selected style contributes a system-prompt section through the systemPrompt registry — empty for the default style — so the model's communication contract mirrors the familiar Claude Code Explanatory / Learning styles plus project- and user-authored custom styles.

Output styles

Three built-in styles ship with the package:

Style	Effect
default	Contributes no section.
Explanatory	Directs the model to explain its implementation choices and codebase patterns as it works.
Learning	Directs the model to teach collaboratively, requesting short hands-on practice contributions gated by a single TODO(human) marker.

Custom styles are loaded from /.claude/output-styles/*.md (project) and the harness home ~/.dsh/output-styles/*.md, with the later project directory overriding a same-named earlier style. Each file name becomes the style name; frontmatter supplies a description (required) and an optional keep-coding-instructions boolean (default true):

---
description: Keep answers concise and skip the usual preamble.
keep-coding-instructions: false
---
Your concise coding and communication instructions here.

When keep-coding-instructions is false, the contributed section states that it replaces the default coding-instruction section; when true (the default), the default coding instructions are retained alongside the style's prose. Malformed or missing frontmatter fails the plugin load loud.

Selection

The active style is the settings key outputStyle in the cc-output-styles namespace, layered over the plugin composition outputStyle config. Without a settings provider the composition value stands; then /output-style still switches in-session.

A switch re-emits system-prompt/change, so the next assembled prompt picks up the new style's section.

/output-style command

Registered through ctx.commands; every composed command adapter discovers it without a model turn.

Input	Result
/output-style	Show the current style and the selectable styles.
/output-style	Set the active style; unknown names return an error listing the selectable styles.

Composition

The plugin injects systemPrompt and commands. Custom apps mount their owners plus this row:

- id: system-prompt
  name: '@deepseek-ai/dsh-system-prompt'
- id: commands
  name: '@deepseek-ai/dsh-commands'
- id: cc-output-styles
  name: '@dsh-cc/output-styles'
  config:
    outputStyle: Explanatory

Model Experience

Output style section

What the model sees

While a non-default style is active, its prose appears as the cc:output-style prompt section, rendered before the deployment persona. The default style contributes nothing.

Token effect

The default style adds no tokens. Non-default styles add their section prose to each assembled prompt.

KV Cache effect

The contribution is a fixed prompt-section provider; a style switch changes the section text and so the assembled prompt for the following turn, like any system-prompt section change.

Known Limitations and Deferred Work

• Prompt-section replacement only — keep-coding-instructions: false is expressed in the contributed section's lead-in rather than by unregistering the harness's own coding-instruction provider, which owns that slot.
• Composition fallback is static — without a settings provider the selection reverts to the composition outputStyle after a reload; /output-style persistence needs the settings service.
• No plugin-forced styles — Claude Code's plugin-authored force-for-plugin output styles are out of scope for this row.