stratagate-dsh 0.3.12 → 0.3.13
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +4 -0
- package/dist/client.js +2 -2
- package/dist/plugin.js +11 -39
- package/dist/plugin.js.map +1 -1
- package/docs/DSH.md +1 -1
- package/docs/DSH.zh-CN.md +1 -1
- package/docs/dreaming-agent.md +3 -3
- package/package.json +1 -1
package/docs/DSH.md
CHANGED
|
@@ -24,7 +24,7 @@ Click **🌙 Dreaming** in the long-term memory directory to let a dedicated Age
|
|
|
24
24
|
|
|
25
25
|
**打开 Dreaming 会话 ↗** opens the real DeepSeek Harness conversation, including tool calls and results, where users can refine the goal. This version writes only the directory. Event facts, raw messages, graph, and Profile remain protected. Dedicated read-only tools reuse Core retrieval without chat evidence batches, reinforcement, or changes to Block expansion and decay.
|
|
26
26
|
|
|
27
|
-
There is no fixed whole-run model-call allowance. Growing memory is split into batches and model-facing history is compacted while the full human transcript is retained. The Agent continues across batches and recovers occasional empty or text-only responses; three consecutive replies without tool action pause the run. Unproductive loops, request failures, changed sources
|
|
27
|
+
There is no fixed whole-run model-call allowance or cumulative token cap to configure. Growing memory is split into batches and model-facing history is compacted while the full human transcript is retained. The Agent continues across batches and recovers occasional empty or text-only responses; three consecutive replies without tool action pause the run. Unproductive loops, request failures, or changed sources pause the task without losing progress. The directory page shows task status, progress, and pause reasons; estimated consumption is retained internally for diagnostics, not provider billing. Resuming an older task clears its legacy cumulative token cap while preserving progress and the session. If another task changes the sources or directory, terminate and start again rather than overwrite newer content.
|
|
28
28
|
|
|
29
29
|
**AI 会看到的目录 ↗** opens a temporary window showing the entire actual default directory injection for the selected space, with no additional UI truncation. The existing directory summary budget still applies; factual answers require source retrieval.
|
|
30
30
|
|
package/docs/DSH.zh-CN.md
CHANGED
|
@@ -24,7 +24,7 @@
|
|
|
24
24
|
|
|
25
25
|
**打开 Dreaming 会话 ↗** 进入真实 DeepSeek Harness 会话,可查看工具调用和结果,也可补充整理方向。第一版只写目录;Event 事实、原始消息、图谱和 Profile 不参与改写。检索复用 Core,但不创建聊天证据批次、不强化记忆,也不改变 Block 的展开与衰减状态。
|
|
26
26
|
|
|
27
|
-
|
|
27
|
+
没有固定的整轮模型调用次数或累计 token 上限,无需配置消耗额度。记忆增长时自动拆批,并压缩模型使用的历史上下文;用户可查看的完整会话记录保留。完成一批后自动进入下一批;偶发空响应或普通说明会提示继续,连续三次没有工具行动才暂停。重复无效调用、请求失败或来源变化会暂停,保留进度。目录页只展示整理状态、进度和暂停原因;内部保留估算消耗用于排查问题,不代表供应商账单。旧任务继续时清除历史累计 token 上限,保留进度与会话。来源或目录被其他任务修改时应终止并重新开始,避免覆盖新内容。
|
|
28
28
|
|
|
29
29
|
**AI 会看到的目录 ↗** 打开临时窗口,展示当前空间实际默认注入的全部目录文本;窗口不再额外截断。默认注入仍遵守原有目录摘要预算,具体事实需要检索来源。
|
|
30
30
|
|
package/docs/dreaming-agent.md
CHANGED
|
@@ -30,11 +30,11 @@ Event 分页按递归稳定字段顺序序列化,页结果和持久断点绑
|
|
|
30
30
|
|
|
31
31
|
批次最多 24 个 Event,并按预计上下文容量进一步拆小;这是页面大小,不是整轮调用限额。目录页的进度按这些批次计算(例如 `1/14` 批约为 7%),不是简单按章节数计算;一个较大的章节可能拆成多个批次。每个工作步骤和人工继续时使用 DSH 的 surface replacement 重建模型工作上下文,只携带一个当前批次、必要的现有总览、用户方向、最近工具结果和错误反馈;原始会话日志完整保留。目录 sections/overview 页绑定内容摘要并持久记录覆盖,任务明确给出已读范围与下一未读位置,压缩后不重新开始浏览。失败写入提案单独保留,不受最近 12 条工具日志的窗口影响;过大提案只提供错误与分批修正提示,完整内容仍在会话记录。用户方向本身过长时暂停并提示精简,避免静默丢失指令。
|
|
32
32
|
|
|
33
|
-
`dreaming_finish` 表示批次已全部检查,允许无需修改;所有冻结批次完成后进入 completed
|
|
33
|
+
`dreaming_finish` 表示批次已全部检查,允许无需修改;所有冻结批次完成后进入 completed。重复无效工具调用、连续无工具进展的模型步骤或失败时暂停。没有累计 token 限额,界面和控制接口不提供上限设置。内部计量保留实际模型输入结构的 token 估算和输出估算/供应商输出计数,仅用于诊断;目录页和模型任务上下文不展示消耗数字,不声称价格或账单精度。旧任务继续时在原子事务中清除历史 tokenLimit,保留任务、批次进度、证据断点与会话身份。
|
|
34
34
|
|
|
35
|
-
无进展计数在新增完整 Event 证据、推进连续有效来源页、读取经校验的目录新项或成功修改目录时重置。搜索关键词变化、空结果与重读已覆盖页均不重置;每批允许有限探索。空响应(含 `(empty)`)或普通文字回复结束但任务未完成时,在同一 Agent Loop 内提示继续;连续三次没有工具行动才暂停,不按整轮累计。复用持久化的 idleModelSteps
|
|
35
|
+
无进展计数在新增完整 Event 证据、推进连续有效来源页、读取经校验的目录新项或成功修改目录时重置。搜索关键词变化、空结果与重读已覆盖页均不重置;每批允许有限探索。空响应(含 `(empty)`)或普通文字回复结束但任务未完成时,在同一 Agent Loop 内提示继续;连续三次没有工具行动才暂停,不按整轮累计。复用持久化的 idleModelSteps,工具检查点或完成批次后重置;不使用后台定时重试。用户暂停、终止和真实请求错误优先于自动继续。切批消息明确上一批已完成、当前批待处理,完成全部批次后只生成总结,不再催促工具调用。
|
|
36
36
|
|
|
37
|
-
原生 Agent 在请求最终组装后的 `llm/stream` 边界计量实际消息(含临时任务和系统提示)及 JSON
|
|
37
|
+
原生 Agent 在请求最终组装后的 `llm/stream` 边界计量实际消息(含临时任务和系统提示)及 JSON 工具定义,自动继续也走相同入口。单次响应最多 32768 Token;若一轮回复尚未产生任何原生工具调用,普通文本超过 8192 Token 会中断并暂停,避免模型重复描述调用计划而持续消耗。reasoning 不计入这条普通文本边界,但 reasoning、普通说明和真正的工具参数均计入单次响应消耗及 32768 Token 总上限。模型失败时拦截宿主自动重试并暂停。没有整轮时长或累计调用次数硬限额;完成后的总结请求也计量,但不重新激活任务或租约。
|
|
38
38
|
|
|
39
39
|
目录写入使用带必需 kind 的明确参数结构,操作数组必须扁平;错误不视为成功提交,Agent 可以纠正并重新提交。暂停时不为说明原因额外调用模型,而是在原生会话写入插件通知,报告真实原因、检查批次与已保存修改。start 每次创建新 run/session;resume 保留本轮会话身份,重建工作上下文,避免累计发送旧轮次历史。
|
|
40
40
|
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "stratagate-dsh",
|
|
3
|
-
"version": "0.3.
|
|
3
|
+
"version": "0.3.13",
|
|
4
4
|
"description": "Recent conversations stay vivid. Older ones fade into summaries, not oblivion. StrataGate gives DeepSeek Harness six-layer, time-decaying memory, while lasting events and relationships settle into a knowledge graph. Bring your memories from other AIs with you—no need to start over.",
|
|
5
5
|
"type": "module",
|
|
6
6
|
"main": "./dist/index.js",
|