dsh-infinite-context
DeepSeek Harness 的上下文管家插件。它让 AI 助手在超长对话中不"失忆"、不"爆窗"：旧对话自动整理成摘要并存档，重要的事情每轮主动提醒，需要细节时随时找回原文。
English
跳到英文说明
中文
它解决什么问题
和 AI 助手长期协作时，你会遇到两个问题：
- 窗口装不下 —— 对话太长，超出模型一次能读取的上限，请求直接报错。
- 压缩即失忆 —— 普通的压缩会把旧对话直接丢掉：做过什么决定、试过什么方案、有什么约束，全都没了。
本插件的思路是压缩不等于删除：旧对话被折叠成一份结构化摘要，原文完整存档、随时可取回；同时把"当前任务进行到哪了"主动放在模型眼前，不靠它自己想起来去查。
核心功能
1. 智能折叠：像整理会议纪要
对话接近上限时，较早的内容自动折叠成一份结构化摘要——总目标、关键决定、改过的文件、遇到的问题、待办事项，一目了然。
折叠的时机和深度不是死规则：
- 看阶段：模型正在连续调用工具时推迟，一段话说完的间隙才动手；
- 看内容：大量工具输出适合折深一些，重要指令则尽量保留原文；
- 会自我修正：每次折叠记录"实际结果与计划的偏差"，下次自动补偿；
- 有安全底线：模型正在思考或回答时绝不动手，两次折叠之间有最小间隔，避免反复折腾。
2. 一切可恢复
折叠前，原文完整存档，摘要开头给出一个取回编号（如 fl_ab12cd34ef56）。模型或你随时可以用 memory_expand 工具读回原文，支持分页和关键词定位；读不完会明确告知总长度和下一次怎么取。
超长的工具输出（一次大文件读取、一份长日志）同样先存档，再用一个带编号的短摘要占位——需要细节时一键展开。
3. 记忆金字塔
每次折叠的摘要存入本地数据库（重启不丢）。积攒多了，插件会自动把主题相近的几条合并成更高层的"长期记忆"；新对话或新问题时，最相关的历史记忆自动检索并注入给模型。
4. 任务状态提醒（工作记忆）
每次折叠后，插件从摘要中提取"总目标 / 待办 / 下一步 / 关键约束"，之后每一轮都主动放在模型眼前。
这解决一个真实痛点：记忆检索靠"相似度"，而五十轮之前的待办和当前问题往往并不相似——靠模型自己想起来去查，生产环境中一次也没发生过。主动提醒才是可靠的。
5. 提前准备，而不是让你等
快到折叠点时，插件会在后台提前把摘要写好；真正需要折叠时直接取用，现场零等待。如果后台准备会和你的对话抢同一个模型服务器的位置（本地小窗口部署时可能发生），它会自动让路——对话永远优先。
6. 会自己学习的参数
- 每个模型路线"摘要实际需要多久"，由插件持续测量并自动调整等待时间——本地部署的大模型慢但健康，等待会自动放宽；云端卡死的流不会浪费第二次机会。
- 每次折叠"实际落点与计划的偏差"持续记录，下一次自动补偿。
- 提供商报"装不下"时，记下服务器的真实容量，之后按真实容量规划，而不是按声明值硬撑。
- 你在配置里手工指定的值优先级最高，学习只在未指定时生效。
7. 其他能力
- 工具结果即时提炼：大段输出先提取关键行（错误、计数、文件路径），原文存档备用。
- 本地部署优化：自动识别 127.0.0.1 / 局域网上的模型，按探测到的真实窗口规划、注入精简指令、放宽等待。
- 中英文双语场景：摘要格式与关键信息提取都覆盖中文的错误、警告与超时行。
模型可用的工具
| 工具 | 作用 |
|---|
memory_search | 搜索全部历史记忆 |
memory_expand | 展开被折叠/存档的原文（支持分页、关键词） |
memory_status | 查看记忆系统状态与用量统计 |
memory_index | 浏览记忆索引 |
memory_maintain | 记忆库体检：重复、冲突、低价值条目 |
配置
所有配置在 cordis.yml，每一项都有注释。最常用的：
| 配置 | 作用 | 默认 |
|---|
compaction_governance | 统一折叠调度（建议保持开启） | true |
summarization_timeout_ms | 单次摘要的等待上限（毫秒）；0 = 不限制 | 自动：本地 240s / 云端 90s，随输入伸缩，并随实测学习 |
task_state_enabled | 每轮注入任务状态提醒 | true |
speculative_fold | 后台提前准备折叠摘要 | true |
tool_archive | 超长工具结果存档 | true |
rag_top_k 等检索项 | 每轮注入多少条相关记忆 | 见 cordis.yml |
完整清单与每项的详细说明见 cordis.yml 内注释。
安装
需要 Node.js ≥ 22.19.0 和 DSH 0.2.x。
npm pack
dsh plugin --profile <profile> add ./dsh-infinite-context-<version>.tgz
安装后重启 DSH 生效。数据（记忆、存档、统计）存放在 DSH 数据目录的 storages 下，升级插件不会丢失。
兼容性说明
- 折叠通过 DSH 自带的
compaction-basic 完成，工具调用与结果的配对永远完整。
- 本地模型（127.0.0.1 / 内网地址）自动识别：按探测到的真实窗口规划、放宽摘要等待、注入精简指令。
- 默认的语义向量化是内置实现，零第三方依赖；需要更强的语义检索可自行安装
@huggingface/transformers。
- 本地模型只支持特定推理档位时，摘要调用沿用会话当前的档位设置。
许可
MIT
English
What problem does it solve
Long-running sessions with an AI assistant hit two walls:
- The window fills up — the conversation outgrows what the model can read at once and requests start failing.
- Compaction means amnesia — naive compaction throws old context away: decisions made, approaches tried, constraints stated — all gone.
This plugin's principle: compaction is not deletion. Old turns are folded into a structured summary while the original text is archived and retrievable on demand; and the current task state is actively placed in front of the model every turn instead of relying on it to remember.
Core features
1. Smart folding, like meeting minutes
As the conversation nears its limit, older turns fold into a structured summary — goals, decisions, files changed, problems hit, pending work.
When and how deep to fold is not a fixed rule:
- It watches the phase: deferred while tools run in a loop, applied at turn boundaries;
- It watches the content: tool-heavy bulk folds deep, important instructions stay verbatim;
- It self-corrects: each fold records how far the result landed from the plan and compensates next time;
- It has hard safety rails: the model is never interrupted mid-generation, and a minimum interval prevents fold churn.
2. Everything is recoverable
Before folding, the original text is archived and the summary carries a reference (e.g. fl_ab12cd34ef56). The model — or you — can read the exact original back via the memory_expand tool, with paging and keyword search; truncated reads state the full length and the exact next call.
Oversized tool outputs (a big file read, a long log) are archived the same way and replaced by a short stub that carries the reference.
3. The memory pyramid
Every fold summary is persisted locally (survives restarts). As they accumulate, topically-similar summaries are consolidated into higher-level long-term memories; the most relevant ones are retrieved and injected for new questions automatically.
4. Task-state reminders (working memory)
After each fold the plugin extracts "goal / pending / next step / key constraints" from the summary and places it in front of the model every single turn.
This fixes a real gap: memory retrieval works by similarity, and a fifty-turn-old pending job is rarely similar to the current question — in production the model never once searched on its own. Pushing beats hoping.
5. Prepared in advance, instead of making you wait
As the surface approaches the fold trigger, the summary is precomputed in the BACKGROUND; the actual fold consumes it with zero waiting. If the background call would contend with your conversation for the same local server (possible on small local deployments), it stands down — the conversation always wins.
6. Parameters that learn
- How long a summary takes per model route is measured continuously and the wait adjusts itself — a big locally-deployed model is slow but alive, so waits widen; a dead cloud stream does not get a second patient chance.
- How far each fold actually landed from its plan is tracked and compensated.
- When a provider rejects a request for size, the server's real capacity is recorded and future folds plan against reality, not the declared number.
- Values you set explicitly always win; learning only applies where you did not.
7. Also included
- Immediate distillation of tool results: signal lines (errors, counts, paths) extracted, full text archived.
- Local-deployment optimizations: automatic discovery of 127.0.0.1 / LAN models, planning against probed real windows, concise-mode directives.
- Bilingual (Chinese/English) summaries and signal extraction, including Chinese error and warning lines.
Tools available to the model
| Tool | Purpose |
|---|
memory_search | Search all stored memories |
memory_expand | Expand folded/archived original text (paging, keyword) |
memory_status | Memory system status and usage statistics |
memory_index | Browse the memory index |
memory_maintain | Health check: duplicates, conflicts, stale entries |
Configuration
Everything lives in cordis.yml with a comment on every key. The common ones:
| Key | Purpose | Default |
|---|
compaction_governance | Unified fold scheduling (keep on) | true |
summarization_timeout_ms | Per-attempt summary deadline (ms); 0 = unlimited | auto: local 240s / cloud 90s, scaled by input, refined by learning |
task_state_enabled | Per-turn task-state reminder | true |
speculative_fold | Background precomputation of fold summaries | true |
tool_archive | Archive oversized tool results | true |
rag_top_k etc. | How many memories to inject per turn | see cordis.yml |
The full list with per-key explanations is in the comments of cordis.yml.
Install
Requires Node.js ≥ 22.19.0 and DSH 0.2.x.
npm pack
dsh plugin --profile <profile> add ./dsh-infinite-context-<version>.tgz
Restart DSH after installing. Data (memories, archives, statistics) lives under the DSH data directory's storages folder and survives upgrades.
Compatibility
- Folding runs through DSH's built-in
compaction-basic; tool call/result pairing is always preserved.
- Local models (127.0.0.1 / private LAN) are detected automatically: planning against probed real windows, wider summary deadlines, concise-mode directives.
- The default embedder is built in with zero third-party dependencies; install
@huggingface/transformers yourself for stronger semantic retrieval.
- When a local model only accepts specific reasoning levels, summary calls follow the session's current setting.
License
MIT
