dsh-tavern-fengyue

dsh-tavern-fengyue

English | 中文

Tavern Fengyue (DSH) — an RPG character-card engine running on the dsh host, with direct import of SillyTavern character cards (.png with embedded data, or .json) and Fengyue cards.

This project aims to guide the development paradigm for tavern-style agents in the agent era: tavern-like applications should stand on mainstream agent infrastructure, not on a bespoke DSL stack.

Table of Contents

• Design Principles
• Features
• Quick Start
• Workspace Layout
• SillyTavern and Fengyue Card Import
• Docs
• Contributing
• License
• Disclaimers

Design Principles

LLM handles reasoning; everything else runs as a script.

The division of labor, layer by layer:

SillyTavern raises the learning curve for both players and card authors — players face regex injection, world info, and injection depth; authors need to master multiple bespoke DSLs just to build a card with real functionality. This project folds everything back into traditional scripts:

• ST configuration languages retired — regex, STScript, world info, and macros all become local .mjs scripts; ST's multi-source depth injection (each source re-injecting its own prompt into the same context every turn) becomes a single live postPrompt per turn, paired with dsh's built-in automatic context compression. In the agent era, reuse the agent infrastructure that already works — don't reinvent wheels.
• Context structure reuses the mainstream agent paradigm — systemPrompt (fixed) + compressed history + pure-append dialogue + latest user message + per-turn dynamic postPrompt.
• Workspace as memory — runtime/ IS an llm-wiki: a standing root index + detail archives fetched on demand; no vector DB, no RAG, no plugin black boxes — retrieval, long-term memory, and skills all live as scripts and local files.
• Built-in card authoring agent — player interaction stays at "pick a card + chat"; the author's learning cost shifts to the AI, which writes the scripts and config; this two-way convergence forces the concept system to stay lean.
• Tools and local files first — modern agent architecture's external memory + tool-calling pattern; not bespoke frameworks.
• Embrace tool calls; push mechanical logic into tools — fixed rules (dice rolls, pathfinding, table lookups, state machines) are encapsulated in preset/tools/ and run by script instead of being reproduced by the model in prose. Two wins: it strips "computing by the rules" — reasoning the LLM should never carry — out of the model, saving tokens while removing the instability of mechanical logic reproduced by generation; and the model reads the live objective results a tool returns every turn, so with script injection the narrative always has a dynamic objective reference instead of being improvised.

Breaking the traditional impossible triangle. "Rich functionality / low comprehension cost / token cost" are at odds in the traditional ST framework. This project collapses every concept into the system framework itself, and all three improve at once:

• Near-unlimited functionality — flexible scripts can carry almost any logic, unconstrained by pre-defined concept sets
• Extremely low comprehension cost — players see only cards and chat; authors let the AI write the configuration, no need to learn proprietary DSLs
• Significantly lower token spend — a fixed context structure boosts prefix-cache hit rates; state lives on disk maintained by the tail agent, not re-emitted every turn by the model

Features

• Minimalist UI — pick a card + chat, that's it; advanced customization relies on the built-in agent
• Direct import of SillyTavern character cards (.png with embedded data, or .json) and Fengyue cards — world info auto-converts to lorebook.mjs, regex to scripts, complex logic translated by the agent with a full report — nothing silently dropped
• Tail agent — after each narrative turn, state changes are written to runtime/ so the frontend renders panels directly, without the model burning output tokens repeating static state
• Auto/manual save — keeps last 10 world snapshots (chat history + model context + local files) on every send, roll back anytime
• Cross-platform — all scripts run on Node.js

Quick Start

Prerequisites: Node.js ≥ 22.19 or ≥ 24, pnpm (npm i -g pnpm or corepack enable).

Compatibility: verified against @deepseek-ai/dsh 0.1.7-rc.2 (npm next) — the list lives in config/dsh-compatibility.json, and pnpm bootstrap installs the newest verified version into the profile. The 0.1.7 migration (2026-09-27; six breaking host surfaces adapted, see the decision note) retired 0.1.5 support — the profile no longer assembles on a 0.1.5 host, and the dev tooling asserts the verified list fail-loud. Upstream dsh is in developer preview with breaking changes expected; if a newer dsh breaks the tavern profile, re-run pnpm bootstrap to restore the verified profile, or follow the external-install fallback in docs/release/independence-and-release.zh.md.

How this is distributed

This repository is a DSH profile project: it declares the tavern-fengyue profile and its bundle layer in package.json, and every package under packages/ is built from source. Install from source today — the steps directly below.

The same bundle is being prepared for the standard dsh plugin channel (published to npm as dsh-tavern-fengyue plus its three sibling packages, then installed per profile with dsh plugin add). That path has not been rehearsed end to end yet, so this README deliberately does not document it until it is.

Install and run from source

In the project directory:

pnpm install       # required after cloning — a checkout ships no build output
pnpm build
pnpm bootstrap     # registers this checkout as the `tavern-fengyue` profile
pnpm tavern        # starts the host; Ctrl-C stops it

The host serves the Web UI at http://127.0.0.1:3081. The tavern profile claims its own port so it can run alongside dsh web, which defaults to 3080; a longer-running instance prints its full URL (with a login token) on boot. --port overrides it: pnpm tavern --port 3099.

pnpm tavern runs the host this repo pins as its own devDependency (node_modules/@deepseek-ai/dsh), so a globally installed dsh is not required — and no DSH_HOME in your shell either: the dev orchestrator injects it. The profile lives in its own harness home (~/.dsh-tavern-fengyue), leaving an existing ~/.dsh untouched; an exported DSH_HOME wins and relocates the home (re-run pnpm bootstrap after changing it).

Daily Use

Day to day, pnpm tavern from this checkout is all it takes — no env vars needed. Both the profile's packages and the host come from here (they are link: dependencies), so after a source edit pnpm build is all it takes — the profile re-applies the bundle layer live.

The separate harness home exists so the host's real ~/.dsh — and any profile you already keep there — is never touched; the tavern profile can sit alongside the rest of your dsh setups.

The one case where you set DSH_HOME by hand is driving the same profile with a globally installed dsh: a different host binary, which would otherwise look in its default ~/.dsh and miss the profile — and whose version then has to match the verified line above (the pinned devDependency is what pnpm tavern guarantees):

DSH_HOME=~/.dsh-tavern-fengyue dsh --profile tavern-fengyue               # bash / zsh
$env:DSH_HOME="$HOME\.dsh-tavern-fengyue"; dsh --profile tavern-fengyue   # PowerShell

(Or put the profile in the host's default home from the start: DSH_HOME=~/.dsh pnpm bootstrap.)

First Play

1. Set your API Key in the sidebar "API Key" (or set DEEPSEEK_BASE_URL / DEEPSEEK_API_KEY for a self-hosted / proxied endpoint)
2. Click "＋ New Tavern Session" → pick a card / import an ST character card / Fengyue card / start from a blank skeleton with the AI writing agent
3. Each turn: model narrates → tail agent bookkeeps → panel refreshes itself
4. Header "Save / Load / Reset" freezes or restores this world at any point

Stop: Ctrl-C in the host terminal; pnpm stop under background mode.

Workspace Layout

Each session creates tavern_workspace/<timestamp>/:

<workspace>/
├── preset/              ← card source (persists across sessions)
│   ├── meta.json        card metadata
│   ├── prompt/          systemPrompt / postPrompt / maintenancePrompt
│   ├── setup/           initial-state template
│   ├── scripts/         {{script}} template scripts
│   ├── tools/           main-agent CLI tools
│   └── assets/          media files
├── runtime/             ← live state (tail agent maintains)
└── savings/             ← save snapshots (runtime backups)

Everything that enters the model's context is a local file you can open and audit — this IS the memory system of the whole project.

SillyTavern and Fengyue Card Import

On import, world info folds into lorebook.mjs; regex, STScript, and community-derived extensions (such as the MVU variable-update convention) land in preset/st-import/ with a translation report — capabilities preserved, concepts consolidated. Full mapping: sillytavern-mechanics-and-import.zh.md.

On MVU: the framework fully supports the MVU paradigm, but we recommend replacing it with a named tool from preset/tools/ — a tool call can narrate, mutate the state, and return the freshly updated value in one turn, weaving state reads/writes naturally into the narrative flow. If you want to keep MVU's "inline variable-update directive written into narrative" paradigm specifically to avoid tool-call interruption, the framework's hook scripts + frontend JS presentation layer fully support that path too.

Docs

Deep-dive docs live in docs/ — organized by area, currently Chinese-first; an English pass follows once the set is reorganized. See the docs index for the full map:

• Design · file tree · scripts & tools
• Card hooks · SillyTavern mechanics & import · field map
• Dynamic post injection · autosave & retry · tail-session archive
• Keyless testing · mock-LLM testing
• Independence & release

docs/notes/ holds dated decision/bug-fix notes plus the chronological devlog — a historical record; on conflict, the code and READMEs win.

Contributing

PRs welcome — see CONTRIBUTING.md and CODE_OF_CONDUCT.md. pnpm lint, pnpm build, and pnpm test must pass before review (CI runs the same on Ubuntu / macOS / Windows).

License

MIT for this project's own code — see LICENSE. Upstream DSH packages keep their own licenses; vendored snapshots and the SRD 5.1 corpus (CC-BY 4.0, not MIT) are listed in THIRD_PARTY_NOTICES.md.

Disclaimers

• Not affiliated — DeepSeek / DeepSeek Harness, SillyTavern, Dungeons & Dragons / SRD (Wizards of the Coast), and Genshin Impact (HoYoverse) are products of their respective owners. This independent project is not affiliated with, endorsed by, or sponsored by any of them; names are used for identification and interoperability only.
• Content is the card's, not ours — the engine runs whatever card you author or import. Imported community cards speak for their authors; this project does not vet, endorse, or curate card content, and responsibility for restricted content lies with the importer under their local rules.
• As is — provided without warranty under the MIT license. Example-preset imagery is AI-generated, as labelled in THIRD_PARTY_NOTICES.md.