One model writes the story, another does the bookkeeping

What it does DSH Tavern is a roleplay chat frontend — a “tavern” in the SillyTavern sense — rebuilt on DSH (DeepSeek Harness), an agent runtime. It imports the SillyTavern ecosystem directly: character cards, world books, presets, regex beautification, MVU variable cards, Tavern Assistant scripts. The difference is the turn itself — a foreground model writes only prose, while background models settle variables, generate candidate replies, retrieve memories, and design NPC profiles. Runs on Windows, macOS, Linux, and Android, and a desktop-hosted session can be played remotely from a phone.

The interesting bit The split doubles as a cache strategy. The README pins SillyTavern’s slowness on world-book edits invalidating the KV prefix cache; by keeping the foreground prompt stable and offloading volatile work, DSH Tavern claims 95%+ cache hits and roughly 10-second turns. Memory follows the same logic: every turn is saved verbatim and retrieved on demand, instead of pre-compressing summaries and hoping you guessed what you’d need later.

Key highlights

• Deterministic status bars — MVU cards render from a fixed template, variables are settled by a background model, and a failed update retries alone instead of torching the whole turn.
• Script mode — bind a novel or outline to a card and play along its mainline, read in segments; a full-length novel never enters context whole.
• Conversational card surgery — tell the agent what you dislike, it proposes changes, you confirm, it writes. A separate debug agent reads your actual play logs to fix broken regex or stale status bars.
• Guide and Skill — per-session steering (“more inner monologue, dialogue under three lines”) plus scene-specific writing skills invoked on demand, instead of always-on prompt filler.
• Consistent NPCs, durable data — recurring characters get a background-designed profile so they stop drifting between appearances; updates and reinstalls preserve your saves.

Caveats

• SillyTavern third-party plugins can’t be installed (DSH plugins can); the README says built-in substitutes cover “most” cards, which is doing some work in that sentence.
• The headline numbers — ~10-second turns, 95%+ cache hits, “more alive than the original” — are the project’s own claims and user testimonials, not independent benchmarks.
• README, demo video, and community channels are Chinese-first; expect some translating.

Verdict If you play or author SillyTavern character cards and have watched a status bar disintegrate mid-scene, this is a real re-architecture, not a reskin. If your setup leans on niche tavern plugins or you need an English-first community, watch from the bleachers for now.