Lsc-91-69/dsh-brain-compaction

人脑式上下文压缩逻辑,大幅减少长任务上下文占用以及token消耗

Project Overview项目介绍

This is a context compaction integration plugin for DeepSeek Harness 0.1.5. It integrates 7 community context compression tools, provides a unified configuration panel, compression verification and recall capability to save tokens, enable selective forgetting and cross-session memory. Use it when optimizing LLM context length. Note only dsh-argp engine works currently, some components need extra configuration.

这是适配DeepSeek Harness 0.1.5的上下文压缩整合插件,集成7个社区上下文压缩相关工具,提供统一配置面板、压缩验证和统一召回能力,可节省Token、实现选择性遗忘与跨会话记忆。需优化大模型上下文长度时使用,注意当前版本仅dsh-argp引擎可用,部分组件需额外配置。

Or use CLI install (for developers)或使用命令行安装(适合开发者)

CLI Install命令行安装

dsh plugin --profile web add "github:Lsc-91-69/dsh-brain-compaction"

Lsc-91-69/dsh-brain-compaction 加入你的 DSH 配置(web profile)即可启用。

READMEREADME

English  |  中文

dsh-brain-compaction — Human-brain-style context compaction, unified

Since 0.1.5 this targets DeepSeek Harness 0.1.5. See "DSH 0.1.5 compatibility" below.

A DeepSeek Harness plug-in (Cordis bundle) that composes seven community context-compression plug-ins into one coordinated, verifiable system — not a dependency collector:

Component Tools (namespaced) Default Row id
dsh-argp (atomic reference-graph pruning, 0-LLM, selective forgetting) recall_pruned / list_pruned / recall enabled dsh-argp
dsh-compaction-instant (VCC-style near-lossless) recall / search disabled (unusable on 0.1.5, see below) brain-compaction-instant (+brain-tool-recall / brain-command-recall)
@wanyantiande/dsh-headroom (tool-output compression + CCR store) headroom_retrieve disabled (needs local proxy) brain-headroom
dsh-memory-vault (cross-session memory) memory_remember / memory_recall / memory_forget enabled brain-memory-vault
dsh-sgme (multi-agent shared memory) memory_search wiki_* signal_* role_* (18 tools) disabled (needs SGME gateway) brain-sgme
dsh-mcp-lens (lazy MCP tool catalog) mcp_search / mcp_call enabled (zero servers) brain-mcp-lens
dsh-routing-suite (smart routing "We Need" chain, 0 extra LLM calls) no model tools; assemble-time guidance section + read-only status API enabled brain-routing-suite
dsh-context-doctor (injection token audit) context_audit optional (GitHub-only) context-doctor
Unified layer (this plug-in) brain_status / brain_verify / brain_recall enabled dsh-brain-compaction

DSH 0.1.5 compatibility

0.1.5 changed three runtime API shapes. Every one of them fails at module evaluation time (or fails silently), so a static contract test cannot see them — which is why this release also upgrades verification from "read the source" to "actually mount it".

0.1.5 change Old shape (1.1.0 — broken on 0.1.5) New shape (0.1.5)
@deepseek-ai/dsh-settings runtime exports narrowed to { SettingsProvider, SettingsConflictError, redactSecrets } import { installSettingsSection, settingsNamespace } from '@deepseek-ai/dsh-settings'throws on import (the whole plug-in row fails to load) the package-level function became the registry instance method ctx.settings.installSection(ctx, ns, schema, entry, hooks); a namespace is now a plain string literal; the package is imported type-only (zero runtime dependency)
settings became a service that must be resolved resolved as a soft dependency via ctx.inject(['settings'], …): when the provider is absent (headless, older, trimmed composition) the plug-in still mounts in composition-config-only mode with one warn; a provider lacking installSection degrades instead of throwing
Session no longer exposes .events session.events.filter(…) → always undefined, so brain_verify's compaction history silently reported 0 entries reads session.snapshotEvents() (with a fallback to the old shape); before/after token comparison works again
the browser module table (PLATFORM_MODULES) no longer has @deepseek-ai/dsh-client-runtime require("@deepseek-ai/dsh-client-runtime/client")missed the module table, the entire settings card (plus its CSS) disappears takes createSnapshotStore from the platform seed word @deepseek-ai/dsh-client-store (same zustand+immer engine; getSnapshot/subscribe/set/update contract unchanged)

Related adjustments:

  • Dependency lines: dsh-argp moves to ^1.1.0 (its peerDependencies name ^0.1.5-rc.1, and it too moved to the 0.1.5 settings source thunk); peerDependencies becomes >=0.1.5-rc.1.
  • dsh-compaction-instant is currently unusable: 0.1.4 still imports installSettingsSection / settingsNamespace, so enabling that row on 0.1.5 fails at startup. The patch keeps it disabled: true with a comment; switch engines once upstream ships a fix.
  • Engine detection hardened: argp 1.1.0 also exports ArgpT1Engine, ArgpRecallEngine, and ArgpProbeEngine; all four class names now count as argp, and an unknown class name is named in the report instead of a generic FAIL.
  • dsh.client.inject drops the removed dsh-client-runtime (and the unused connection).

Verification (new in 0.1.5)

npm run build       # esbuild bundles the host half; the client half is copied verbatim
npm run test        # 42 static contract checks (incl. 0.1.5 regression guards)
npm run test:mount  # 16 real-runtime mount checks (built artifact onto real cordis + ToolRuntime + SettingsProvider)
npm run test:client # 16 browser-half smoke checks (stub require really executes the factory)
npm run verify      # all of the above

test:mount needs a DSH installation to resolve @deepseek-ai/*: it probes $DSH_RUNTIME_ROOT$DSH_HOME/profiles/node_modules → a few common paths and skips (exit 0) when none exists. It asserts what static analysis never can: the built artifact imports, the fiber really reaches ACTIVE, all three tools register, the settings namespace really registers and accepts a write, and brain_verify really reads 2 compaction summaries out of snapshotEvents().

Deep integration = composition + coordination, not dependencies

  • Assemble: cordis.patch.yml mounts every component row through the Cordis composition mechanism (row insertion, package-name references, disabled overrides). The profile uses nodeLinker: hoisted, so sub-plugins listed as regular dependencies resolve by row name — no per-plugin dsh plugin add.
  • Coordinate: src/host/integration.ts probes at runtime with zero imports — tool registry (ctx.tools.schemas()), services (ctx.get('compaction'), tokenMeter), engine class name. Any missing/disabled component degrades gracefully.
  • Arbitrate: ctx.compaction allows exactly one implementation per context. argp and instant both register the tool recall (duplicates throw), so the patch keeps exactly one engine row enabled (argp by default); brain_verify validates this.
  • One panel: Settings → Plugins → Brain Compaction binds multiple settings namespaces (this plug-in + compaction-instant + headroom) in one card; writes still go through each plug-in’s native scope (schema-validated, live-applied).
  • Routing chain (We-Need): dsh-routing-suite is a host row — zero extra LLM calls; it regex-classifies the first user task (fix/diagnose/review → inspect-first; build/create/implement → direct) and appends one ~90-char guidance line, but only for sessions whose preset id is exactly routing-suite. Deep integration: mounted under the unique id brain-routing-suite (status API and strategy surface stay available), and the same We-Need discipline is baked into the brain preset persona (always-on for any first task, no id check). The honest token story: direct tasks skip redundant exploration, maintenance tasks never skip root cause — fewer wasted turns, less context growth; it does NOT lower per-request injection cost. If you previously installed the standalone bundle, remove it before using this assembly.

Install

One-liner (recommended):

dsh plugin --profile web add "github:Lsc-91-69/dsh-brain-compaction"
# restart dsh web (stop it and start again, or use the GUI's restart control)

From a source checkout:

git clone https://github.com/Lsc-91-69/dsh-brain-compaction
dsh plugin --profile web add "file:<path-to-the-clone>"
# restart dsh web

Use the file: (or github:) spec, never link: — a link: spec does not install this package's sub-plug-in dependencies and the composed rows would fail to resolve. For iterative development you may link: and pnpm install the sub-deps yourself. Ship the built lib/ either way.

Verify the install (2 minutes)

  1. Restart dsh web — the profile composes its plugin rows at boot; a running process never hot-loads a new bundle.

  2. Check the composed tree (no server started, read-only):

    dsh --profile web --dump-config | Select-String 'brain-|compaction'
    

    Expected: dsh-brain-compaction ENABLED, dsh-argp ENABLED, compaction-basic disabled (this bundle's patch turns the stock summarizer off), brain-compaction-instant / brain-tool-recall / brain-command-recall disabled (engine exclusivity), and brain-memory-vault / brain-mcp-lens / brain-routing-suite ENABLED.

  3. Ask the model to run brain_verify in a fresh session. A healthy install reports engine: ArgpGraphEngine and PASS for argp / memory-vault / mcp-lens / routing-suite; instant / headroom / sgme / context-doctor are SKIP by design.

  4. Open Settings → Plugins — the "Brain Compaction" card is contributed by this plug-in's browser half. (That tab renders the intersection of the namespaces the Host serves and the cards the browser registered, so a missing card means the Host half did not mount — check the host log for an import error.)

Troubleshooting a fresh install

  • The requested module '@deepseek-ai/dsh-settings' does not provide an export named 'installSettingsSection' → you are on a pre-0.1.5-dsH release. This version (0.1.5) targets DSH 0.1.5-rc.1+; upgrade both sides.
  • tool "recall" is already registered → argp and instant are both enabled; restore engine exclusivity (see "Engine switching").
  • Duplicate row id → a sub-plug-in was also added as its own bundle; run dsh plugin --profile web remove <name> and keep this bundle's assembly.
  • The profile cannot resolve dsh-brain-compaction → re-run dsh plugin --profile web install; the dependency must be present in the profile's own node_modules.

Optional component (Context Doctor — GitHub-only, not published to npm):

dsh plugin --profile web add "github:Zhenyu98/dsh-context-doctor#main"

Install from GitHub & distribution

# any machine (append #v0.1.5 to pin a release)
dsh plugin --profile web add "github:<your-name>/dsh-brain-compaction"
# restart dsh web

This works because the repository root IS the package: package.json, lib/, src/, preset/, scripts/, test/ all sit at the top level beside the README. pnpm resolves a github: spec by packing the repo root, so a layout that nests the package one level down (e.g. repo/<name>/package.json) fails with "no package.json" — verified against real GitHub tarballs: a subdirectory layout yields a root containing only README.md and the subdirectory, and pnpm has nothing to pack.

  • The repo must contain lib/ (build artifacts) — a GitHub install uses the repo as-is and does not build; contract test 5d guards this.
  • Installing pulls the sub-plug-in dependencies automatically (npm versions resolve by row name under nodeLinker: hoisted); do NOT dsh plugin add any sub-plug-in separately (duplicate row ids).
  • For local development use file: (or link: + pnpm install the sub-deps).
  • Run npm run verify before publishing: build + 42 contract assertions + 16 real-runtime mount assertions + 16 browser-half assertions. The mount suite needs a DSH install to resolve @deepseek-ai/* (it probes $DSH_RUNTIME_ROOT$DSH_HOME/profiles/node_modules → each parent directory's node_modules) and skips cleanly when none exists.

The agent preset ships along: the validated 人脑式上下文压缩 preset lives in preset/ (compression-discipline persona + session-level argp engine row). Copy it to $DSH_HOME/.agent-presets/brain-compaction/; see preset/README.md.

Release checklist: git init, commit everything (.gitignore excludes node_modules/; commit lib/, preset/, src/, scripts/, test/, cordis.patch.yml, READMEs, LICENSE), tag v0.1.5, then dsh plugin --profile web add "github:<you>/dsh-brain-compaction#v0.1.5".

In a fresh session ask the model to run brain_verify — component matrix (PASS/SKIP/FAIL), session token snapshot, recent compaction history, and the context_audit hook, all in one report.

Engine switching (argp ↔ instant ↔ headroom)

Edit the plug-in’s cordis.patch.yml (or override its rows in your own profile patch): set disabled: true on the currently enabled engine row, remove disabled from the target row (instant also needs brain-tool-recall and brain-command-recall enabled), restart.

Session-level engine: the 人脑式上下文压缩 preset places its engine row inside the per-session compaction realm in agent.cordis.yml (default dsh-argp) — that row is the engine brain-preset sessions actually use; the host-level engine row serves presets without their own compaction group. Keep the two planes consistent when switching (the preset already replaced compaction-basic with dsh-argp for you).

Verification (before/after token comparison)

  1. Call brain_verify (summary): engine state + component matrix + current tokens + recent compaction/summary leaf fields (shadowed/checkpoint). (This used to report nothing on 0.1.5 — session.events was removed; it now reads session.snapshotEvents().)
  2. If Context Doctor is installed, call context_audit: per-item token cost of AGENTS.md chain / skill catalog / tool schemas / MCP surface, duplicates and conflicts (per-request injection; compaction shrinks the conversation, not this baseline).
  3. Run two manual compactions (/compact), then brain_verify again and compare sessionTokens and the compaction history — that is the before/after Token delta.
  4. brain_status for arbitrary snapshots; brain_recall(seq|query) for unified recovery of pruned content.
  5. Three shipped suites (npm run verify): test/contract.test.mjs (42 static contract assertions), test/mount.test.mjs (16 real-runtime mount assertions), test/client.test.mjs (16 browser-half smoke assertions).

Config matrix

Parameter Plane How
Engine choice composition cordis.patch.yml disabled flags + restart (only argp is usable on 0.1.5)
argp maxPasses / recencyGuard composition row config (default 256 / 10)
instant thresholdRatio / retainTurns / retainTokens / auto / checkpointCap runtime unified panel / settings compaction-instant (that plug-in is unusable on 0.1.5)
headroom toggle / threshold / proxy runtime unified panel / settings headroom
selfTestAfterBoot / verifyDetail runtime unified panel / settings brain-compaction (0.1.5: ctx.settings.installSection)
sgme baseUrl / agentKey / … composition row config + env vars
vault injectLimit / recallLimit composition row config (default 8 / 10)
mcp-lens servers / allowTools / denyTools composition row config (empty, fail-closed)
context-doctor defaultCwd / cacheTtlMs composition its own patch

Layout

src/host/index.ts       entry (settings soft-dependency registration, tools, events, boot self-check)
src/host/config.ts      unified config (brain-compaction namespace + schema; settings package is type-only)
src/host/integration.ts component matrix / engine arbitration / verify / unified recall
src/client.js           unified settings panel (verbatim ModuleLoader factory; platform word dsh-client-store)
cordis.patch.yml        assembly: component rows + engine exclusivity + basic disabled
scripts/build.mjs       esbuild packaging (src/host/*.ts → lib/*; falls back to the official CLI on EPERM)
test/contract.test.mjs  static contract tests (42, incl. 0.1.5 regression guards)
test/mount.test.mjs     real-runtime mount tests (16; needs a DSH install, else skipped)
test/client.test.mjs    browser-half smoke tests (16; no browser needed)

Development

npm run build       # esbuild host + verbatim client copy
npm run test        # static contract tests (42)
npm run test:mount  # real-runtime mount tests (16; skipped without a DSH install)
npm run test:client # browser-half smoke tests (16)
npm run verify      # all of the above

Troubleshooting

  • Boot error tool "recall" is already registered: argp and instant are both enabled — restore engine exclusivity.
  • Boot error duplicate row id: a sub-plug-in was previously added as its own bundle (dsh plugin --profile web remove <name>, keep this bundle’s assembly).
  • Boot error The requested module '@deepseek-ai/dsh-settings' does not provide an export named 'installSettingsSection': you are on a pre-0.1.5 version (that one only targets 0.1.0-rc.x). Upgrade to 0.1.5+.
  • Startup fails after enabling the instant engine: dsh-compaction-instant@0.1.4 is not 0.1.5-compatible (same removed exports). Keep that row disabled: true, or wait for an upstream release.
  • The "Brain Compaction" card is missing from Settings → Plugins: either the host half did not mount (the host log carries the import error), or the host brain-compaction namespace failed to register — on 0.1.5 the plug-in configuration tab only renders the intersection of namespaces the Host serves and card keys registered, so a failed registration means no card (this is not a rendering bug). npm run test:mount answers this directly.
  • brain_verify engine FAIL naming a class not in the engine table: the ctx.compaction implementation is new. If it is another argp assembly, add the class name to ENGINE_CLASS_MAP in src/host/integration.ts.
  • brain_verify engine FAIL with no class name: something re-enabled compaction-basic or preempted the engine.
  • A section is missing from the panel: that plug-in’s row is disabled by default (see table); enable and restart.

License

MIT. Sub-plug-ins keep their own licenses (dsh-argp MIT, dsh-compaction-instant MIT, dsh-memory-vault MIT, @wanyantiande/dsh-headroom MIT, dsh-mcp-lens MIT/EULA, dsh-sgme MIT, dsh-context-doctor BSD-3-Clause).


English  |  中文

人脑式上下文压缩逻辑 · dsh-brain-compaction

0.1.5 起适配 DeepSeek Harness 0.1.5。 详见下方「DSH 0.1.5 适配」。

一个 DeepSeek Harness 插件(Cordis bundle)。它不是"Hell World 依赖收集器",而是把 社区里七个人脑式上下文压缩方向的插件组装成一个统一、协同、可验证的体系

组件 工具(命名空间) 默认 行 id
dsh-argp(原子引用图剪枝 · 0-LLM · 选择性遗忘) recall_pruned / list_pruned / recall ✅ 启用 dsh-argp
dsh-compaction-instant(VCC 式近无损即时压缩) recall / search ⬜ 禁用(0.1.5 下不可用,见下) brain-compaction-instant (+brain-tool-recall/brain-command-recall)
@wanyantiande/dsh-headroom(工具输出压缩 + CCR 可逆存储) headroom_retrieve ⬜ 禁用(需本地代理) brain-headroom
dsh-memory-vault(跨会话记忆库) memory_remember / memory_recall / memory_forget ✅ 启用 brain-memory-vault
dsh-sgme(拾光记忆引擎·多智能体共享) memory_search wiki_* signal_* role_* 等 18 个 ⬜ 禁用(需原厂网关) brain-sgme
dsh-mcp-lens(MCP 工具懒加载) mcp_search / mcp_call ✅ 启用(零服务器) brain-mcp-lens
dsh-routing-suite(智能路由 "We Need" 思维链, 0 额外 LLM 调用) 无模型工具;assemble 注入引导段 + 只读状态 API ✅ 启用 brain-routing-suite
dsh-context-doctor(注入物 token 审计) context_audit ⬜ 可选(GitHub-only) context-doctor
统一层(本插件) brain_status / brain_verify / brain_recall dsh-brain-compaction

DSH 0.1.5 适配

0.1.5 把三处运行时 API 形状改了。它们全都是"模块求值期直接抛错/静默失效"级别, 静态契约测试看不见——所以本版本除了改代码,还把验证方式从"读源码"升级成"真挂载"。

0.1.5 的变化 旧写法(1.1.0,在 0.1.5 上必坏) 现写法(0.1.5)
@deepseek-ai/dsh-settings 运行时导出收窄为 { SettingsProvider, SettingsConflictError, redactSecrets } import { installSettingsSection, settingsNamespace } from '@deepseek-ai/dsh-settings'import 即抛(整个插件行加载失败) 包级函数改为注册表实例方法 ctx.settings.installSection(ctx, ns, schema, entry, hooks);namespace 变成普通字符串字面量;该包只做 type-only 导入(运行时零依赖)
settings 变成需要解析的服务 软依赖 ctx.inject(['settings'], …):provider 缺席(headless、旧版、裁剪过的组合)时插件照常挂载,退化为"组合层配置",只打一条 warn;installSection 形态不符时同样降级而不是崩
Session 不再有 .events 属性 session.events.filter(…) → 恒为 undefinedbrain_verify 的压缩历史静默报 0 条 改读 session.snapshotEvents()(保留旧形状兜底),压缩前后对比恢复可用
浏览器模块表(PLATFORM_MODULES)不再有 @deepseek-ai/dsh-client-runtime require("@deepseek-ai/dsh-client-runtime/client")missed the module table整张设置卡片(含 CSS)消失 直接从平台 seed 词 @deepseek-ai/dsh-client-storecreateSnapshotStore(同一 zustand+immer 引擎,getSnapshot/subscribe/set/update 契约不变)

连带调整:

  • 依赖线dsh-argp 提到 ^1.1.0(其 peerDependencies 明确 ^0.1.5-rc.1,且自身也 改用 0.1.5 的 settings 源 thunk);peerDependencies 改为 >=0.1.5-rc.1
  • dsh-compaction-instant 暂不可用0.1.4import { installSettingsSection, settingsNamespace } ——在 0.1.5 上该行一旦启用即启动失败。patch 里保持 disabled: true 并留注释;等上游适配后再切引擎。
  • 引擎识别加固:argp 1.1.0 除 ArgpGraphEngine 外还导出 ArgpT1Engine/ArgpRecallEngine/ ArgpProbeEngine,四个类名都认作 argp;未知类名会在报告里点名,而不是笼统 FAIL。
  • dsh.client.inject 清掉已删除的 dsh-client-runtime(顺带移除未使用的 connection)。

验证方式(0.1.5 新增)

npm run build      # esbuild 打包 host + 逐字复制 client
npm run test       # 42 项静态契约(含 0.1.5 回归守卫:禁止改回已删除的 API 形状)
npm run test:mount # 16 项真实运行时挂载(把构建产物挂到真的 cordis + ToolRuntime + SettingsProvider)
npm run test:client# 16 项浏览器半区烟雾(桩 require 真的把 factory 跑一遍)
npm run verify     # 以上全部

test:mount 需要一个 DSH 安装来解析 @deepseek-ai/*:按 $DSH_RUNTIME_ROOT$DSH_HOME/profiles/node_modules → 若干常见路径 依次探测,都没有则跳过(exit 0)。 它断言的是静态测试永远看不见的东西:构建产物可 import、fiber 真的 ACTIVE、三个工具 真的注册、settings namespace 真的注册且写入生效、brain_verify 真的从 snapshotEvents() 里读出了 2 条压缩历史。

深度融合 = 组合 + 协调,而不是依赖

  • 组装cordis.patch.yml 用 Cordis 组合机制(insert 行、包名引用、disabled 覆盖) 把全部组件行一次装入 profile 组合树;profile 用 nodeLinker: hoisted,所以子插件 作为本 bundle 的 dependencies 可被 loader 直接按行名解析(无需逐个 dsh plugin add)。
  • 协调src/host/integration.ts 运行时零 import 探测——工具注册表 (ctx.tools.schemas())、服务 (ctx.get('compaction') / tokenMeter)、引擎类名 判定当前激活引擎;任何组件缺失/禁用都优雅降级,不会崩。
  • 互斥仲裁ctx.compaction 每上下文只允许一个实现。argp 与 instant 都会注册 工具 recall(重名即抛错),因此 patch 里恰好一个引擎行启用(默认 argp); brain_verify 会校验这一点。
  • 统一面板:Settings → Plugins → 人脑式上下文压缩。一张卡片绑定多个 settings namespace(本插件 + compaction-instant + headroom),哪些在场就渲染哪些小节; 写入仍走各插件原生 scope(schema 校验 + live 生效)。
  • 路由思维链(We-Need)dsh-routing-suite 是 host 行——0 次额外 LLM 调用, 用正则对首条用户任务分类(修复/排查/审查→检查优先;新建/实现→直接执行并验证), 仅在预设 id 恰为 routing-suite 的会话注入一句约 90 字符的引导。深度融合 = 本插件用自研 id brain-routing-suite 挂载它(状态 API/策略面可用), 且把同一套 We-Need 纪律固化进 brain 预设的 persona(任何首任务常驻生效, 不依赖 id 判断)。 token 节省的真实机理:直接任务免去无谓探索, 维护任务避免跳过根因——减少返工轮次 与上下文生长; 它不改变注入物的常驻成本, 请勿夸大。 若你在此前单独安装过 dsh-routing-suite, 请先移除独立 bundle 再使用本插件的组装。

安装

一条命令(推荐):

dsh plugin --profile web add "github:Lsc-91-69/dsh-brain-compaction"
# 重启 dsh web(dsh web 停掉再起, 或使用 GUI 的重启控件)

从源码目录安装:

git clone https://github.com/Lsc-91-69/dsh-brain-compaction
dsh plugin --profile web add "file:<克隆下来的目录路径>"
# 重启 dsh web

file:(或 github:)规格,不要link:——link: 不安装本包的子插件 依赖,组装行会解析失败。开发迭代时可以 link: 自行 pnpm install 子依赖。 两种方式都依赖仓库里已提交的 lib/ 构建产物。

安装后验证(约 2 分钟)

  1. 重启 dsh web:profile 在启动时组合插件行,运行中的进程不会热加载新 bundle。

  2. 核对组合树(不启服务,只读):

    dsh --profile web --dump-config | Select-String 'brain-|compaction'
    

    预期:dsh-brain-compaction ENABLEDdsh-argp ENABLEDcompaction-basic disabled(本 bundle 的 patch 关掉了内置摘要引擎)、 brain-compaction-instant / brain-tool-recall / brain-command-recall disabled(引擎互斥),以及 brain-memory-vault / brain-mcp-lens / brain-routing-suite ENABLED

  3. 在新会话里让模型调用 brain_verify。健康状态应报告 engine: ArgpGraphEngine,且 argp / memory-vault / mcp-lens / routing-suite 四项 PASS;instant / headroom / sgme / context-doctor 按设计 SKIP。

  4. 打开 设置 → 插件:应能看到「人脑式上下文压缩」卡片(由本插件的浏览器半区 提供)。该页只渲染"Host 已服务的 namespace ∩ 浏览器已注册的卡片 key"的交集, 所以卡片不出现意味着 Host 半区没挂上——去宿主日志里找 import 错误。

新装常见问题

  • The requested module '@deepseek-ai/dsh-settings' does not provide an export named 'installSettingsSection':装的是 0.1.5 之前的版本。本版本(0.1.5)要求 DSH 0.1.5-rc.1 及以上,两边都要升。
  • tool "recall" is already registered:argp 与 instant 同时启用,恢复引擎互斥 (见「引擎切换」)。
  • 报行 id 重复:某个子插件也被单独装成了 bundle,先 dsh plugin --profile web remove <name>,只保留本 bundle 的组装。
  • profile 解析不到 dsh-brain-compaction:重跑 dsh plugin --profile web install,该依赖必须落在 profile 自己的 node_modules 里。

可选组件(Context Doctor,GitHub-only 未发布 npm):

dsh plugin --profile web add "github:Zhenyu98/dsh-context-doctor#main"

从 GitHub 安装与分发(给别人用时)

# 别人/另一台机器(按需加 #v0.1.5 固定版本)
dsh plugin --profile web add "github:<你的用户名>/dsh-brain-compaction"
# 重启 dsh web

这条命令能成立的前提是:仓库根目录就是包本体——package.jsonlib/src/preset/scripts/test/ 与 README 并列在顶层。pnpm 解析 github: 规格时会 对仓库根那一层打包,所以"包被多套一层目录"的结构(如 仓库/<包名>/package.json)会因找不到根 package.json 而安装失败——这一点已用真实 GitHub tarball 验证:子目录结构的打包结果是根目录下只有 README.md 和那个子目录, pnpm 无处可打包。

  • 仓库必须提交 lib/(构建产物)——GitHub 安装直接按仓库内容安装,不跑构建; test/contract.test.mjs 的 5d 项会检查它。
  • 安装即自动拉取子插件依赖(npm 版本,nodeLinker: hoisted 下按行名可解析); 不要额外 dsh plugin add 任一子插件,避免行 id 重复。
  • 请用 file: 而非 link: 规格安装本机开发副本(link: 不安装子插件依赖)。
  • 发布前跑一遍 npm run verify:构建 + 42 项契约 + 16 项真实运行时挂载
    • 16 项浏览器半区。挂载套件需要一份 DSH 安装来解析 @deepseek-ai/*(按 $DSH_RUNTIME_ROOT$DSH_HOME/profiles/node_modules → 各级父目录的 node_modules 依次探测),都没有时会干净跳过而不是误报。

Agent 预设一起分发:本仓库 preset/ 内是已验证的「人脑式上下文压缩」预设 (persona 压缩纪律 + 会话级 argp 引擎行),复制到 $DSH_HOME/.agent-presets/brain-compaction/ 即可在预设选择器中见到;详见 preset/README.md

发布清单git init 后提交全部(node_modules/ 已被 .gitignore 排除, lib/preset/src/scripts/test/cordis.patch.yml、README、 LICENSE 都提交),打 tag v0.1.5,然后 dsh plugin add github:<你>/dsh-brain-compaction#v0.1.5

重启后新会话里让模型调用 brain_verify——组件矩阵 PASS / SKIP / FAIL + token 快照

  • 最近压缩历史 + context_audit 接入点,即验证完成。

引擎切换(argp ↔ instant ↔ headroom)

编辑插件的 cordis.patch.yml(或你自己的 profile patch 覆盖它):把当前启用引擎行 加 disabled: true,把目标引擎行去掉 disabled(instant 还要同时启用 brain-tool-recall / brain-command-recall),重启。参考文件内注释。

⚠️ DSH 0.1.5 下只有 dsh-argp 可切dsh-compaction-instant@0.1.4 仍引用已删除的 包级 settings 导出,启用即启动失败;headroom 0.3.0 未做启用后链路的实测。见 「DSH 0.1.5 适配」。

会话级引擎人脑式上下文压缩 预设(agent.cordis.yml 的 compaction 组)把引擎行 直接放在会话 realm 内(默认 dsh-argp)——该行才是 brain 预设会话真正使用的 压缩引擎;host 层的引擎行服务于不带压缩组的预设。切换时两处要保持一致(同理把 预设组内的 compaction-basic 换成 dsh-argp,该预设已替你做)。

验证机制(压缩前后 Token 对比)

  1. 会话中调用 brain_verify(summary):得到引擎状态 + 组件矩阵 + 当前 token + 最近 compaction/summary 事件叶子字段(shadowed/checkpoint 前后数值)。 (0.1.0 时代这里恒为空——session.events 在 0.1.5 已被移除;现在走 session.snapshotEvents()。)
  2. 若已装 Context Doctor:调用 context_audit,得到 AGENTS.md 指令链 / 技能目录 / 工具 schema / MCP 工具面的逐项 token 成本、重复与冲突(常驻注入物,压缩不降低 这部分;真正被压缩的是会话历史)。
  3. 连续两次手动压缩(/compact)后再次 brain_verify:比较两次 sessionTokenscompaction/computation 差值——即压缩前后 Token 对比。
  4. brain_status 任意时刻快照;brain_recall(seq|query) 统一召回被剪内容。
  5. 三层自带测试(npm run verify):test/contract.test.mjs 42 项静态契约、 test/mount.test.mjs 16 项真实运行时挂载、test/client.test.mjs 16 项浏览器半区烟雾。

配置矩阵

参数 调整方式
引擎选择(argp/instant/headroom) 组合 cordis.patch.yml(改 disabled,重启);0.1.5 下仅 argp 可用
argp maxPasses / recencyGuard 组合 config(本插件的 patch 默认 256/10)
instant thresholdRatio/retainTurns/retainTokens/auto/checkpointCap 运行时 统一面板 / settings compaction-instant(0.1.5 下该插件暂不可用)
headroom 压缩开关/阈值/代理参数 运行时 统一面板(headroom 启用后)/ settings headroom
本插件 selfTestAfterBoot / verifyDetail 运行时 统一面板 / settings brain-compaction(0.1.5:ctx.settings.installSection
sgme baseUrl/agentKey/…(网关) 组合 config + 环境变量(patch 注释)
vault injectLimit/recallLimit 组合 config(默认 8/10)
mcp-lens servers/allowTools/denyTools 组合 config(默认空, fail-closed)
context-doctor defaultCwd/cacheTtlMs 组合 其自身 patch

目录结构

src/host/index.ts       插件入口(apply:settings 软依赖注册、工具注册、事件监听、boot 自检)
src/host/config.ts      unified config(brain-compaction namespace + schema, type-only 依赖 settings 包)
src/host/integration.ts 组件矩阵/引擎仲裁/验证报告/统一召回(运行时探测, 零 import 耦合)
src/client.js           统一配置面板(逐字发布, ModuleLoader factory, 平台词 dsh-client-store)
cordis.patch.yml        组装:组件行 + 引擎互斥 + compaction-basic 禁用
scripts/build.mjs       esbuild 打包(src/host/*.ts → lib/*;EPERM 时回落官方 CLI 二进制)
test/contract.test.mjs  静态契约测试(42 项,含 0.1.5 回归守卫)
test/mount.test.mjs     真实运行时挂载测试(16 项,需 DSH 安装;缺则跳过)
test/client.test.mjs    浏览器半区烟雾测试(16 项,无需浏览器)

开发

npm run build       # esbuild 打包 host + 逐字复制 client(需要 esbuild 可解析)
npm run test        # 静态契约测试(42 项)
npm run test:mount  # 真实运行时挂载测试(16 项;无 DSH 安装则跳过)
npm run test:client # 浏览器半区烟雾测试(16 项)
npm run verify      # 以上全部

常见问题

  • 启动报 tool "recall" is already registered:argp 与 instant 同时启用了——把 引擎互斥恢复到一个启用。
  • 启动报行 id 重复:之前单独 dsh plugin add 过子插件——先 dsh plugin --profile web remove dsh-argp 等,再保留本 bundle 的组装。
  • 启动报 SyntaxError: The requested module '@deepseek-ai/dsh-settings' does not provide an export named 'installSettingsSection':装的是 0.1.5 之前的版本(那版只 适配 0.1.0-rc.x)。升级到 0.1.5+。
  • 启用 instant 引擎后启动失败dsh-compaction-instant@0.1.4 未适配 0.1.5(同样的 已删除导出)。保持该行 disabled: true,或等上游发新版。
  • 设置页看不到「人脑式上下文压缩」卡片:两种可能——(a) Host 半区没挂上(宿主日志里 会有 import 错误);(b) 0.1.5 的插件配置页只渲染 「Host 已服务的 namespace ∩ 已注册的卡片 key」的交集,所以 Host 端 brain-compaction namespace 注册失败时卡片不会出现(这不是渲染 bug)。 npm run test:mount 能直接判出 namespace 是否注册成功。
  • brain_verify 引擎 FAIL 且提示"不在已知引擎表内"ctx.compaction 是某个新引擎 实现。若是 argp 的新装配面,把类名加进 src/host/integration.tsENGINE_CLASS_MAP
  • brain_verify 引擎 FAIL(无类名):profile 的 cordis.patch.yml 里 compaction-basic 行被其它层重新启用,或被 argp 之外的东西抢占。
  • 面板里某小节不见:对应子插件行处于 disabled(默认禁用项见上表),先启用再重启。

许可证

MIT;子插件各自遵循其原始许可证(dsh-argp MIT、dsh-compaction-instant MIT、 dsh-memory-vault MIT、@wanyantiande/dsh-headroom MIT、dsh-mcp-lens MIT(EULA 见包内)、 dsh-sgme MIT、dsh-context-doctor BSD-3-Clause)。

上一个 Prev dsh-universal-worldbook 下一个 Next DeepSeek-Harness-Desktop