← Back to list

Lume

DSH Desktop enhancement plugin — it puts verifiable facts on the table; constraints are injected only when relevant, so idle chat costs no extra tokens. 1) Task-execution discipline: each turn is classified from the sentence plus recent-turn trajectory (question / research / discussion / diagnosis / execute) with explicit boundaries, so questions do not touch files and diagnosis does not silently start fixing; it flags citations to code lines never opened this session, negative claims about symbols never seen, requirement items the deliverable does not cover, and unverified assumptions dressed up as open questions; at 75%/90% context pressure it warns and exports structured session memory (goal, settled decisions, open items, key locations) that a new session receives on its first turn; past sessions, including ones too full to continue, are re-distilled into cross-session knowledge at startup, idempotently. 2) Checkable artefacts: task contract (goal / scope / counts estimated then backfilled / completion criteria / non-goals); change ledger that auto-records every edit together with how it was verified, and surfaces still-unverified entries at delivery; hypothesis ledger including ruled-out items, where any conclusion requires evidence, the decision method and a counter-example check; design decisions (point / choice / rejected options / impact); project knowledge accumulated per working directory across sessions — numbered, deletable by name, sentences only, secrets rejected. 3) Behaviour triggers keyed on trajectory rather than wording: drifting exploration, editing without verifying, retrying a dead path, missing contract, missing design, stale hypotheses, criteria drift, knowledge not captured, editing a file never read this session; each notice carries concrete counts and repeats at most once every two turns. 4) Runtime metrics: route decisions, trigger hits, block assembly and external outcome signals are written to a local lume-metrics.jsonl and readable through lume_metrics, which reports trigger efficacy with an explicit observational caveat and excludes mechanisms that have no mechanical criterion. 5) Persona system: distil named characters from chat logs, novels, scripts and character-setting documents, with tone, catchphrases and reply length anchored to real material statistics; per-persona long-term memory plus temporary facts that expire after 30 days; corrections become style rules, approved replies are pinned as corpus, lexical leakage detection after a persona switch, memory star map, persona card export/import. Layered injection keeps session-stable text in the system prompt and sends volatile content as a tail snapshot, so the prompt cache stays valid.

DSH Desktop 增强插件：在会话里补上「事实到位」这一步——约束按需注入，闲聊轮不额外加 token。一、任务执行纪律：按这一句话加最近几轮轨迹判定请求类型（问答/查找/讨论/诊断/执行）并划清边界，问答轮不改文件、诊断轮不越权修复；引用本次没打开过的代码行、对没见过的符号下否定断言、需求条目与交付物对不上、把「我没查」包装成「待你定」，都会被当场点出；上下文占用到 75%/90% 预警并导出结构化会话记忆（目标/已拍板/未决/关键定位），新会话开局注入，续接一句「继续」即可；历史会话（含已撑满聊不动的）启动时自动补蒸馏出跨会话知识，幂等不重复。二、可检查的产出：任务契约（目标/范围/数量先估后回填/完成判据/非目标）、改动台账（每处改动自动入账并记下验证方式，未验证的条目在交付时点出）、假设台账（含已排除项；下结论必须带证据、裁决方式与反例检查）、设计决策（决策点/选择/放弃理由/影响面）、按工作目录跨会话累积的项目知识（带编号、可点名删除，只收句子并拦截密钥）。三、行为触发器按轨迹纠偏、不看措辞：撒网不收敛、连写不验、死路重撞、契约缺失、设计缺失、假设过期、判据漂移、知识未记、未读就改——每条带具体数字，同类两轮内不重复。四、运行时度量：路由判定、触发器命中、块装配与外部结果信号写本机 lume-metrics.jsonl，lume_metrics 可查触发器效能（口径为命中后 3 轮内出现真验证命令或载具从无到有，明确标注观察性、无机械口径的标未判定且不计入分母）。五、人设系统：从聊天记录、小说、剧本、人物设定文档蒸馏具名角色（语气、口头禅、回复篇幅锚定真实素材统计）；以人设为键的长期记忆，以及 30 天后自动过期的临时记忆；纠正自动转风格约定、认可的回复摘录为语料；切换人设后按签名词做词法泄漏检测；记忆星图可视化与角色卡导出导入。注入分层：系统段只放会话恒定内容，易变内容随对话尾部快照下发，前缀缓存每步可用。