Lume

Most desktop AI assistants drift in long conversations: they lose sight of the original goal, start editing before investigating, skip verification after changes, and even get stuck in edit-revert oscillation loops. Lume injects a Codex-style task-execution protocol into every DSH Desktop session, covering context management, task decomposition, stage gating, change protection, verification loops, failure attribution, oscillation prevention, and result re-confirmation. The protocol is adaptive: a short version for casual chat and a full version for complex tasks. When two consecutive turns show clear failure signals it temporarily appends corrective clauses and lifts them on success. Cross-session reflection reminders are injected into the next conversation and fade out once performance recovers. No hidden chain-of-thought is ever exposed; the output matches the task complexity.

On top of the protocol sits a persona system where every character is a named, persistent individual with long-term memory keyed by persona, surviving across sessions and projects. Style evolves from your in-conversation requests and solidifies into learned style conventions that override the base contract when they conflict. Switching personas triggers a handover broadcast and per-turn lexical drift detection, so long-conversation switches stay reliable. Four built-in protected personas ship out of the box; custom characters are produced by a distillation pipeline that accepts novels, scripts, character sheets, or raw WeChat and QQ chat exports. The host-side pipeline runs in ten to ninety seconds in the background, keeps all source material local, treats it as untrusted text, and never injects it into the conversation context.

The management panel offers a force-directed memory star-map rendered on Canvas with date filtering and inline editing, a post-session reflection log that scores protocol execution across four dimensions on a 0-to-2 scale and writes it locally at zero extra token cost, and one-click JSON card export and import for cross-device migration. All data is stored in the DSH official storageDomain; nothing is uploaded anywhere.

Lume is built for DSH Desktop users who want their AI to work with genuine engineering discipline-read before writing, test before claiming done, attribute failures before retrying-while still speaking in a voice they enjoy living with day after day. The protocol and the persona are two independent, independently toggleable layers: keep the discipline and drop the character, or the other way around.