dsh-ponytail

NanoAI independent derivative of @wenaixi/dsh-ponytail 5.6.2. Version 5.6.3-nanoai.1 retains the upstream runtime identity; its DSH 0.2.2-rc.1 support has not been runtime tested. The upstream npm package is not this derivative; obtain the exact NanoAI build before installing.

English | 简体中文

A DeepSeek Harness (DSH) port of DietrichGebert/ponytail, maintained independently (ADR-0010). Injects the 7-rung ladder directly into system prompts, provides 6 specialized skills, and registers zero tools.

What it is

A counter-weight to LLM boilerplate and over-engineering: deletion over addition, boring over clever, and fewest files possible. Challenge requirements before writing code (YAGNI), prioritize existing functions and standard libraries, and solve in one line when possible.

Features a Zero Cache Miss Architecture (ADR-0012): session baseline is locked permanently, and dynamic intensity switches take effect via turn-end reminders, leaving historical prompt / KV cache 100% untouched across long conversations. Ships as a pure dsh.bundle, writes zero garbage to your home root, uninstalls cleanly, and rebuilds itself through HMR.

Install

Needs the DSH runtime (npm i -g @deepseek-ai/dsh). Node and pnpm floors are in the badges above. Examples use the web profile; substitute your own profile name.

# Install (automatic HMR reload)
dsh plugin --profile web add ./wenaixi-dsh-ponytail-5.6.3-nanoai.1.tgz

# Remove
dsh plugin --profile web remove @wenaixi/dsh-ponytail

# Inspect composed patch tree
dsh --profile web --dump-config | grep -A2 ponytail

Usage

/ponytail            Switch to full mode (all 7 rungs active)
/ponytail lite       Do the bare minimum
/ponytail ultra      Challenge the requirement itself first
/ponytail off        Disable injection
/ponytail default lite   Persist default across sessions
stop ponytail        Equivalent to off (exact match)

Intensity levels reset on each new session according to env > Profile patch > fallback full. Therefore, /ponytail <level> is session-scoped. To persist across restarts and new sessions, use /ponytail default <level>.

Zero Cache Miss Architecture

When switching intensity levels within an ongoing session or updating global default configurations, the top-level SystemPrompt baseline remains strictly byte-identical and static.

Mode transitions take effect dynamically by appending an incremental system notification (<system-reminder>) at the end of the current turn's user message (aligned with the official dsh-tool-skill catalog update pattern, ADR-0012). This 100% protects historical KV Cache / Prompt Cache across all multi-turn conversation history, eliminating costly prefill recalculations and latency spikes.

Skills

Skill Bodies and Description Languages

Skill bodies are retrieved directly from upstream DietrichGebert/ponytail v4.10.3 in verbatim English. Two host-incompatible details are tailored for DSH: ponytail-gain data sources point to assets/*.svg (as upstream benchmarks/ is excluded), and ponytail-help configuration and update sections reference profile patches and dsh plugin.

Model catalog views and configuration forms share the exact same description string, governed by skillDescriptionLang (three-state: explicit Chinese zh / explicit English en / unconfigured auto). When unconfigured, it follows the host's locale.preference (English triggers English alignment; others default to Chinese). Explicit selection locks the language. Canonical descriptions reside in skills/descriptions.zh.json and skills/descriptions.en.json, served live via ctx.remote.ponytail.snapshot().

Configuration Panel

The card detail view provides an embedded settings panel: intensity level control, independent toggle switches for all 6 skills, a one-click reset button, and a 3-tier priority diagnostic chain displaying environment variables, profile patches, and built-in fallbacks with their active hit states.

The level segmented control reflects only the actual persisted configuration. When defaultMode is not configured in the patch, the control highlights the appended unset segment. The currently active runtime level is displayed distinctly in the diagnostic header text.

Configuration reads and writes use two official DSH channels:

Profiles lacking a full UI service bundle (headless, CLI) fall back gracefully to profiles/<name>/ponytail/config.json, fully aligned with the latest DSH pure profile architecture.

Upstream Mapping

The six hooks from upstream hooks/ are consolidated into a single DSH Cordis plugin:

Prompts are injected via systemPrompt.section('ponytail') at order: 50, positioned neatly between persona (0) and tools (100). When switched to off, prompt rendering overhead is zero. Flags and configurations are scoped per profile directory, ensuring full multi-instance isolation.

Development

pnpm typecheck              # tsc --noEmit
pnpm build                  # tsc into lib/, scripts/build-client.mjs generates lib/client.js
node scripts/verify.mjs     # Static contract gates: artifacts, skills, zero tools, slots, locales
node scripts/docs-verify.mjs # Documentation and version consistency gates
node scripts/behavior.test.mjs  # Behavioral test suite (113 tests pass)

dsh --profile web --patch ./cordis.patch.yml --dump-config  # Verify patch parsing
pnpm dsh web --patch ./cordis.patch.yml   # Hot reload dev runner

License

MIT. Full text in LICENSE.

Credits

• Upstream DietrichGebert/ponytail
• DeepSeek Harness for the plugin architecture and runtime
• dsh-plugin-dev for DSH plugin development patterns and architectural guidelines