dsh-memory-plugin
为 DeepSeek-Harness (dsh) 提供三个宿主插件：L1 跨会话短期记忆（由钩子驱动的捕获与召回，零依赖）、memory_save 工具和 memory-timer 定时维护任务。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:YaoQC-Ai/dsh-memory-plugin#93b387cfa43ea2a23f839c16b4d04b8828c04b05Memory Plugin 以 dsh-memory-plugin 发布，当前版本为 0.2.2。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
v0.2.2：面向 DSH ≥0.2.0（在 0.2.0-rc.1、0.2.0-rc.2 与 0.2.1-rc.2、Node.js 24 验证）；不等于已验证所有 DSH 版本。0.1.x 旧宿主不再支持。
0.2.2: re-verified against DSH 0.2.1-rc.2 — runtime files byte-identical to 0.2.0/0.2.1. Diffed every host API the plugin touches (
agent/created,session/eventshapes,append(…,{surfaceOp}),pre-stepwaterfall,surface, V4 admission,form:'recall'): all unchanged;SESSION_FORMAT_VERSIONstays 4. The boot gate evaluatesengines.dshwithincludePrerelease, so>=0.2.0-rc.1 <0.3.0-0already admits 0.2.1-rc.2. Host harness default moves to0.2.1-rc.2. The experimentalappendPluginRecordAPI is gated topackages/experimental/callers only and is intentionally not adopted. 0.2.2：针对 DSH 0.2.1-rc.2 重新核验——运行件与 0.2.0/0.2.1 逐字节相同。逐文件比对了插件依赖的每个宿主 API（agent/created、session/event形状、append(…,{surfaceOp})、pre-stepwaterfall、surface、V4 admission、form:'recall'）：全部未变；SESSION_FORMAT_VERSION仍为 4（无 v4→v5 迁移）。boot 闸门用includePrerelease评估engines.dsh，故现有>=0.2.0-rc.1 <0.3.0-0已放行 0.2.1-rc.2。可选宿主集成默认期望版本改为0.2.1-rc.2。实验性appendPluginRecord/pluginRecordOf仅限packages/experimental/调用，本插件有意不采用。
0.2.1: declarations and verification only — no behavior change. The optional host harness now defaults to
0.2.0-rc.2, andpackage.jsongains a forward-lookingengines.dshrange (>=0.2.0-rc.1 <0.3.0-0) because DSH 0.2.1-alpha.4+ readsengines.dshfirst and falls back topeerDependencies. 0.2.1：只补声明与验证，行为不变。可选宿主集成默认期望版本改为0.2.0-rc.2；package.json新增前瞻性的engines.dsh（>=0.2.0-rc.1 <0.3.0-0），因为 DSH 0.2.1-alpha.4 起优先读engines.dsh、其次才回退peerDependencies。
0.2.0: recall injection moved off the removed
agent/session-startontoagent/created(injecting only onsource: startup | clear, and carryingsource.form: 'recall'for the UI's context-recall rendering); capture now accumulates per-turn text fromsession/eventinstead of the deprecatedsnapshotEvents()/eventAt()reads; and apeerDependenciesgate on@deepseek-ai/dsh ^0.2.0-rc.1makes an incompatible host skip this bundle at boot instead of failing every turn mid-conversation. The producer-ownedplugin:memory-shorttermsource kind and recognition of pre-0.2 rows are unchanged.
0.2.0：召回注入从已删除的
agent/session-start迁到agent/created（仅在source: startup | clear时注入，并带source.form: 'recall'以走 UI 的上下文召回渲染）；捕获改为从session/event增量累积本轮文本，不再用已废弃的snapshotEvents()/eventAt()回扫；新增对@deepseek-ai/dsh ^0.2.0-rc.1的peerDependencies闸门，使不兼容宿主在 boot 期跳过本组合包，而非每个回合在对话中途失败。生产者自有的plugin:memory-shorttermkind 与对 0.2 之前旧行的识别保持不变。
Cross-session short-term memory (L1) for DeepSeek-Harness (dsh). Zero dependencies, zero build step — plain Node built-ins and an inlined UserMessage.
One package, three host plugins (each is a cordis.patch.yml row; this README focuses on the L1 plugin and summarizes its siblings):
| Plugin row | File | What it does |
|---|---|---|
dsh-memory-plugin | index.js | L1 short-term cross-session memory — capture / recall / evict / recover (this document) |
dsh-memory-plugin/save | save.js | memory_save model tool — writes a concept page into <dir>/memory/knowledge/ |
dsh-memory-plugin/timer | timer.js | memory-timer — scheduled maintenance task (default 02:00 wall-clock, dry-run safe) |
为 DeepSeek-Harness（dsh）提供跨会话短时记忆（L1）。零依赖、零构建 —— 纯 Node 内置模块 + 内联构造 UserMessage。
一个包、三个宿主插件（每个都是 cordis.patch.yml 的一行；本 README 详述 L1 插件，并概述两个兄弟插件）：
| 插件行 | 文件 | 作用 |
|---|---|---|
dsh-memory-plugin | index.js | L1 跨会话短时记忆 —— 捕获 / 召回 / 淘汰 / 恢复（本文档） |
dsh-memory-plugin/save | save.js | memory_save 模型工具 —— 把概念页写入 <dir>/memory/knowledge/ |
dsh-memory-plugin/timer | timer.js | memory-timer —— 定时维护任务（默认本地时钟 02:00，dry-run 安全） |
Makes an agent remember what recent sessions talked about when a new session starts.
session/event → turn/end only when reason.kind === 'completed', once per live session turn.agent/created (source startup/clear), checks claimed and pending memory before injecting recent digests; resume/compact skip re-injection.Capture is rule-based text extraction: no extra LLM calls. Local file I/O is synchronous; failures are logged. It is not a semantic summarization pipeline.
Four business listeners; no services injected (the fake dispose-event guard was dropped — Cordis auto-detaches listeners when the fiber unloads):
| Event | Action |
|---|---|
agent/created | On startup/clear: deduplicate claimed/pending memory, then inject recall (resume/compact skip) |
session/event | Accumulate this turn's user/assistant text; capture on turn/end when reason.kind === 'completed' |
agent/pre-step | After next(), update decision.messages, preserve extra fields, settle own pending messages |
agent/request-error | Outer prepend listener: await compaction recovery, append recall only for an uncancelled overflow retry with a newer surface generation |
Empty first steps and rejected steps stay unchanged. Retry recovery does not authorize retries, change the compactor's checkpoint/counts, or add an extra step. Capture deduplication is in-process only; aborted/error/empty turns and historical seed events are not captured. A committed event is not a disk-flush guarantee.
Session-event shapes are asymmetric (verified against packages/core/session/src/types.ts):
user/message → data IS the UserMessage (use data.content, ONLY when data.source.kind === 'user')
assistant/message → data is {turn, step, message} (use data.message.content)
turn/start → data is {turn} (opens a fresh per-turn capture buffer)
DSH also injects scaffolding — the workspace AGENTS.md, runtime-context snapshots, the skills catalog — as user/message events, each stamped with its own source.kind. Only a genuine user prompt carries source.kind === 'user' (the headless bundle and session-controller both stamp it so, and DSH's own tests discriminate injected context by source.kind !== 'user'). Capture (turnTextOf) therefore whitelists kind === 'user' rather than blacklisting known plugin kinds — the source map is merge-extensible, so a blacklist would leak every newly added injection kind. Without the filter the digest fills with boilerplate that is then re-injected as "background" and amplified each session.
session.append only checks JSON round-trip safety, not a strict per-field schema — so an inlined frozen literal with a fresh randomUUID() injects exactly like llm.createUserMessage(), with no import of DSH internals.
From a git host (pnpm links the checkout; nothing is built):
dsh plugin --profile <name> add github:YaoQC-Ai/dsh-memory-plugin
From a local checkout:
dsh plugin --profile <name> add ./dsh-memory-plugin
Verify the layer, then start:
dsh --profile <name> --dump-config # shows a "# == dsh-memory-plugin" layer
dsh --profile <name>
The package declares peerDependencies: { "@deepseek-ai/dsh": "^0.2.0-rc.1" }. On an incompatible host, dsh skips this bundle at boot with a stderr warning (its three plugin rows vanish from --dump-config) instead of failing every turn mid-conversation. The package also declares engines.dsh: ">=0.2.0-rc.1 <0.3.0-0"; hosts from 0.2.1-alpha.4 check that field before the peer range, so the same gate keeps working there.
Every knob is optional; add a config: block to the patch row to override.
| Key | Default | Meaning |
|---|---|---|
dir | $DSH_HOME or ~/.dsh | Base dir; digests live under <dir>/memory/shortterm/ |
maxChars | 2048 | Cap on the injected digest |
maxEntries | 50 | Turns kept per session file |
recentFilesCap | 20 | Session files aggregated on recall (most recent first) |
assistantMaxChars | 280 | Truncation length for a single assistant line in the digest |
maxCandidates | 500 | Evicted-turn candidates kept in compile_candidates.md (the L1 aging archive — excluded from recall so it cannot pollute the injected digest) |
<dir>/memory/shortterm/
├── <session-id-1>.md # newest turn first, capped at maxEntries
├── <session-id-2>.md
└── ...
Evicted turns are not dropped: they roll into <dir>/memory/shortterm/compile_candidates.md (a shared aging archive, capped at maxCandidates). That file lives beside the per-session digests but is excluded from recall — buildDigest never reads it, so the archive cannot crowd out real recency or be re-injected as "background". Entry boundaries in both files are ## <ISO timestamp> lines; Markdown headings like ## foo inside an assistant reply stay part of their turn and are not split into fake entries.
Per-session files reduce write collisions, but the shared candidate archive has no cross-process lock. Digests are not the source of truth; retain the host session logs for recovery.
memory_save (save.js) — a model tool that writes a self-contained concept page into <dir>/memory/knowledge/ (maintaining <dir>/memory/knowledge/_index.md). It takes title (concept name), content (self-sufficient body) and domain, one of persona | execution | knowledge; it validates that the value is one of those three before writing. This is the write half of L2 — the read/retrieval half (memory_search) is not implemented yet.memory-timer (timer.js) — a scheduled maintenance task: each wall-clock day at 02:00 (configurable via hour / minute on the patch row) it fires once and — by default — dry-runs: it appends a trigger receipt to <dir>/memory/_log.md and records state in <dir>/memory/timer-state.json, so you can observe the cadence before wiring it to real work. Consecutive failures are tracked against failAlertThreshold. It injects timer (ctx.interval) from the base bundle's cordis-plugin-timer.L1 gives recency; it does not do semantic search. For that, layer the official Memorix MCP memory example MCP memory server in your own profile's cordis.patch.yml (do not fork it into this package — it needs an external memorix binary, and forcing it would break installs that don't have one):
- insert:
- id: memory-memorix
name: '@deepseek-ai/dsh-mcp-client'
config:
serverName: memorix
transport: stdio
command: memorix
args: [serve]
cwd: !!js process.cwd()
This exposes mcp__memorix__* tools the model can call natively.
memory_save) already ships; the scheduled compiler and memory_search retrieval do not.dsh; a session-query patch override will accompany the memory_search iteration.npm test
# Optional: DSH_SOURCE must point to a built 0.2.1-rc.2 source tree (override the expected version with DSH_EXPECTED_VERSION).
npm run test:host
The 41 unit tests require no host dependencies. The optional integration harness uses real host drivers with a simulated provider, temporary storage and no network. Tests are included in Git, not in the npm runtime package.
MIT
让 agent 在新会话开始时「记得」最近若干会话聊过什么。
session/event → turn/end 且 reason.kind === 'completed' 时，按轮次提取 user + assistant 文本，写入会话摘要。agent/created（source 为 startup/clear）检查已认领与 pending 中是否已有记忆，再注入近期摘要；resume/compact 不重复注入。捕获是规则化文本提取，不增加模型调用；本地文件 I/O 是同步的，失败记告警。它不是语义摘要流水线。
四个业务监听器，不注入服务（已去掉那个挂在不存在事件上的 dispose 守卫 —— Cordis 会在 fiber 卸载时自动摘除监听器）：
| 事件 | 动作 |
|---|---|
agent/created | startup/clear 时检查已有记忆，必要时注入召回（resume/compact 跳过） |
session/event | 增量累积本轮 user/assistant 文本，turn/end 且 reason.kind === 'completed' 时落盘 |
agent/pre-step | await next() 后修正 decision.messages，保留扩展字段，清理本插件 pending |
agent/request-error | prepend 外层等待压缩器恢复；仅在溢出 retry、未取消且 surface generation 增加时追加记忆日志 |
空首步、reject 不恢复；后续工具续步不受空首步规则影响。retry 恢复不自行授权重试，不改压缩器计数或 checkpoint，不产生额外 step。去重仅限进程内；取消、错误、空轮次和历史 seed 不捕获。事件提交不代表已经 flush 到磁盘，进程崩溃后的漏捕获不自动补偿。
会话事件形状不对称（已对照 packages/core/session/src/types.ts 核对）：
user/message → data 本身就是 UserMessage （取 data.content，且仅当 data.source.kind === 'user'）
assistant/message → data 是 {turn, step, message} （取 data.message.content）
turn/start → data 是 {turn} （开启本轮捕获缓冲）
DSH 还会把工作区 AGENTS.md、runtime-context 快照、skills 目录这些脚手架也作为 user/message 事件注入，各自带着不同的 source.kind。唯有真实用户 prompt 才是 source.kind === 'user'（headless bundle 与 session-controller 都如此打标，DSH 自身测试也用 source.kind !== 'user' 区分注入上下文）。因此捕获过滤（turnTextOf）采用白名单 kind === 'user'，而非黑名单已知插件 kind——source map 可被插件扩展，黑名单会漏掉每一种新增的注入 kind。不加这道过滤，digest 会被样板塞满，再作为「背景」注入下一会话，逐次放大。
session.append 只校验 JSON 无损可序列化，不做逐字段严格 schema 校验 —— 所以内联一个带新 randomUUID() 的冻结字面量，注入效果与 llm.createUserMessage() 完全一致，且无需 import 任何 DSH 内部模块。
从 git 托管安装（pnpm 链接该 checkout，不跑任何构建）：
dsh plugin --profile <name> add github:YaoQC-Ai/dsh-memory-plugin
从本地 checkout 安装：
dsh plugin --profile <name> add ./dsh-memory-plugin
先验证层、再启动：
dsh --profile <name> --dump-config # 会显示一层 "# == dsh-memory-plugin"
dsh --profile <name>
本包声明 peerDependencies: { "@deepseek-ai/dsh": "^0.2.0-rc.1" }。宿主不兼容时，dsh 会在 boot 期跳过本组合包并往 stderr 打警告（--dump-config 里三行插件消失），而不是让每个回合在对话中途失败。本包同时声明 engines.dsh: ">=0.2.0-rc.1 <0.3.0-0"；0.2.1-alpha.4 起的宿主先查这个字段、再回退 peer 范围，因此同一道闸门在新宿主上依然生效。
所有旋钮都可选；在 patch 行补 config: 块即可覆盖。
| 键 | 缺省 | 含义 |
|---|---|---|
dir | $DSH_HOME 或 ~/.dsh | 基准目录；摘要存放在 <dir>/memory/shortterm/ |
maxChars | 2048 | 注入摘要的字符上限 |
maxEntries | 50 | 每个会话文件保留的轮次条数 |
recentFilesCap | 20 | 召回时聚合的会话文件数（最近优先） |
assistantMaxChars | 280 | 摘要里单条助手文本的截断长度 |
maxCandidates | 500 | compile_candidates.md 候选池保留的淘汰轮次块数（L1 老化留档 —— 召回时排除，不会污染注入摘要） |
<dir>/memory/shortterm/
├── <session-id-1>.md # 最新轮次在最前，上限 maxEntries
├── <session-id-2>.md
└── ...
被挤出的轮次不丢弃：滚动进入 <dir>/memory/shortterm/compile_candidates.md（全会话共享的老化留档，上限 maxCandidates）。它与各会话摘要同目录，但召回时被排除 —— buildDigest 从不读它，所以留档不会挤占「近期性」、也不会被当作「背景」重新注入。两类文件的条目边界都是 ## <ISO 时间戳> 行；assistant 正文里的 Markdown 标题（如 ## foo）仍属于本轮，不会被误切成伪条目。
按会话分文件可减少写入冲突，但共享候选池没有跨进程锁。摘要不是事实源，恢复依据仍是宿主保留的完整会话日志。
memory_save（save.js） —— 一个模型工具：把自足的概念页写入 <dir>/memory/knowledge/（并维护 <dir>/memory/knowledge/_index.md）。入参 title（概念名）、content（自足正文）、domain（三选一：persona 用户/人格 · execution 工作方式 · knowledge 其它），写入前校验取值必须属于这三者。这是 L2 的写半边；读/检索半边（memory_search）尚未实现。memory-timer（timer.js） —— 定时维护任务：每个自然日本地时钟 02:00（patch 行 hour/minute 可改）触发一次，默认只 dry-run：向 <dir>/memory/_log.md 追加触发凭证、状态记入 <dir>/memory/timer-state.json，便于先观察节奏再接真实任务；连续失败次数按 failAlertThreshold 跟踪告警。它从 base bundle 的 cordis-plugin-timer 注入 timer（ctx.interval）。memory-save.config.wikiDir：可选绝对路径，优先于 dir；未设时仍写 <dir>/memory/knowledge/。升级有本机定制路径的安装前，必须显式配置旧落点。memory-timer.config.enabled: false：完全禁用启动补跑和轮询。只允许一个进程启用 timer，建议 Web 启用，headless 禁用。_log.md 与状态文件的事务。profile patch 示例（这是两个 profile 各自的片段，不要合并）：
# headless/cordis.patch.yml
- id: memory-timer
config:
enabled: false
# web/cordis.patch.yml
- id: memory-timer
config:
enabled: true
wikiDir 应在两个 profile 中均指向同一个实际绝对路径。patch 会替换整个 config，不是深层合并；保留现有其他配置项。包默认不指定任何用户路径。
L1 提供的是「近期性」，不做语义检索。要语义召回，请把官方的 Memorix MCP memory 示例 MCP memory 服务叠进你自己 profile 的 cordis.patch.yml（不要 fork 进本包 —— 它需要外部 memorix 二进制，强加进来会让没装它的用户安装即坏）：
- insert:
- id: memory-memorix
name: '@deepseek-ai/dsh-mcp-client'
config:
serverName: memorix
transport: stdio
command: memorix
args: [serve]
cwd: !!js process.cwd()
它会暴露 mcp__memorix__* 工具，模型可原生调用。
memory_save）已随包落地；定时编译器与 memory_search 检索尚未实现。dsh 的会话全文搜索默认关闭；session-query 的 patch 覆盖将随 memory_search 迭代一起提供。npm test
41 项单测无需宿主依赖。可选真实宿主集成测试使用已构建的 0.2.1-rc.2 源码（期望版本可用 DSH_EXPECTED_VERSION 覆盖）、真实驱动与压缩器、模拟提供方和临时目录，不访问网络：
$env:DSH_SOURCE = '<已构建的宿主源码绝对路径>'
npm run test:host
测试代码随 Git 发布，不包含在 npm 运行包中。
MIT