dsh-prompt-profile

English | 中文

Reusable Markdown prompt profiles for DeepSeek Harness. A profile can select a provider, model, and reasoning effort for one turn, render command arguments into its body, and restore the previous selection afterward.

Why this exists 💡

Model choice and review instructions are often repeated across sessions. dsh-prompt-profile turns those combinations into small, reusable Markdown profiles that can be invoked for one turn and then leave the Agent's previous model selection intact.

Features ✨

• Discover profiles from user and project directories, with project profiles taking precedence.
• Substitute positional arguments into profile bodies with shell-style placeholders.
• Select a provider, model, and reasoning effort for one turn.
• Restore the previous Agent selection automatically or opt out per profile.
• List available profiles and invoke nested profiles by slash-separated names.

Install 📦

dsh plugin --profile demo add ./dsh-prompt-profile

Create a user profile under ~/.dsh/prompt-profiles/ or a project profile under <repo>/.dsh/prompt-profiles/:

---
description: Deep code review
provider: deepseek
model: deepseek-reasoner
reasoningEffort: high
restore: true
---
Review $1 for correctness, regressions, and missing tests. Extra context: ${@:2}

Project profiles override user profiles with the same relative name.

Use 🚀

/prompt-profile list
/prompt-profile review src/index.ts "focus on cancellation"

Nested files use slash-separated names. For example, .dsh/prompt-profiles/review/security.md is review/security.

Supported placeholders:

Frontmatter 🧾

Omit both provider and model to use the current agent selection.

Configure ⚙️

- id: dsh-prompt-profile
  name: dsh-prompt-profile
  config:
    userDirectory: ~/.dsh/prompt-profiles
    projectDirectory: .dsh/prompt-profiles

Develop 🧑‍💻

pnpm install
pnpm run check

Scope 🎯

Version 0.1 intentionally omits prompt chains, loops, skill injection, subagents, worktrees, and best-of-N orchestration. Those features should be added only when Harness extension points can preserve their lifecycle and durable-session semantics.

License 📄

MIT

Credits 🙏

The core idea comes from pi-prompt-template-model. This first Harness-native release focuses on the smallest useful unit: deterministic discovery, argument substitution, scoped model routing, and restoration.