快速上手
Quickstart
当前版本：Agent Pi DSH 3.5.3。DSH 3.x 与 Craft Agents OSS 2.x 经典版的内核、数据目录和更新通道完全分开。
Agent Pi DSH 的默认路径只有一条：打开工作区，直接说任务。工作台是加速器，不是闸门。
- 在官网下载区选择 Windows、macOS 或 Linux 安装包；Windows 用户安装前核对 SHA256：
579CAE929C685CB0BAD65F6476B4124593DD3649C68CCAE0E1CA8829E8EF7213。 - 打开 Agent Pi DSH，选择项目工作区（投标资料所在目录）。
- 配置 DeepSeek。新会话默认
deepseek-v4-flash-vision-exp。 - 可选：设置 → Codex 智能体 → 使用 ChatGPT 登录；无需 API Key。
- 用回形针上传资料后直接下任务：解析招标、对 BOQ、做策划、写正式成果。
- 成果写回
Agent Pi Outputs，出处以芯片标注，可点击定位到源文件。
提示：PDF 当文件读（能抽文本的不先转图）；图纸和照片交给官方视觉模型。
Current version: Agent Pi DSH 3.5.3. DSH 3.x and Craft Agents OSS 2.x Classic use separate engines, data directories and update channels.
The default path in Agent Pi DSH is a single one: pick a workspace and talk. The workbench accelerates work; it does not gate it.
- Choose the Windows, macOS or Linux package in the website download section. Windows users should verify SHA256 before installation:
579CAE929C685CB0BAD65F6476B4124593DD3649C68CCAE0E1CA8829E8EF7213. - Open Agent Pi DSH and pick the project workspace (the folder with the bid files).
- Connect DeepSeek. New sessions default to
deepseek-v4-flash-vision-exp. - Optional: Settings → Codex Agent → Sign in with ChatGPT. No API key is required.
- Attach files with the paperclip and give the job: parse the tender, reconcile the BOQ, plan the works, write Official Outputs.
- Deliverables are written back to
Agent Pi Outputs, with clickable citation chips locating every source.
Tip: PDFs are read as files (no image conversion when text is extractable); drawings and photos go to the official vision model.
安装与升级
Install & upgrade
安装包未签名：Windows SmartScreen 可能拦截，选「仍要运行」。覆盖安装前请完全退出——关到托盘不算退出。
| 事项 | 说明 |
|---|---|
| 不会自动升级 | 2.x 停在 2.6.5，避免把旧引擎静默换成新内核；3.x 请手动下载安装。 |
| 项目还在 | 工作目录和 Agent Pi Outputs 接着用。 |
| 会话不迁移 | 2.x 的旧聊天、旧模型连接不会进 3.x。 |
| 可并存 | 3.x 与 v2.6.5 经典版可同时安装。 |
| 数据目录 | %APPDATA%\agent-pi-dsh-desktop（首次启动写入可写 dsh-home 并组装 tender profile）。 |
若提示 app.asar 被占用：退出 Agent Pi DSH（含托盘），并暂时关掉占用安装目录的编辑器，再重新安装。覆盖安装不会改动已有知识库。
当前正式版同时提供 Windows x64、macOS arm64 和 Linux x64 构件；请从官网下载区进入对应的 GitHub Release 资产。
The installer is unsigned: if SmartScreen blocks it, choose “Run anyway”. Before an in-place upgrade, quit fully — closing to the tray is not enough.
| Topic | Detail |
|---|---|
| No auto-update | 2.x stays at 2.6.5 so the old engine is never silently swapped; install 3.x manually. |
| Projects stay | Your working directory and Agent Pi Outputs carry over. |
| Sessions stay | Old 2.x chats and model connections do not migrate into 3.x. |
| Side by side | 3.x and Classic 2.6.5 can be installed together. |
| Data directory | %APPDATA%\agent-pi-dsh-desktop (first launch writes a writable dsh-home and assembles the tender profile). |
If app.asar is locked: quit Agent Pi DSH (including the tray) and close any editor holding the install directory, then reinstall. In-place upgrades never touch your existing knowledge base.
The current release provides Windows x64, macOS arm64 and Linux x64 artifacts. Use the website download section to open the matching GitHub Release asset.
工作台与业务域
Workbench & domains
Agent Pi DSH 不是聊天壳。它把智能体放进真实项目作业：投标解析与组价、实施策划、投资研究。工作台挂在对话页页签上，默认路径仍然是打开工作区说话。
| 层 | 做什么 |
|---|---|
| 对话、模型、权限、技能目录 | dsh Web |
| 投标 / 实施 / 投资工作台 | 对话页页签 + 建项目 |
| 阶段准备、证据门禁、成果树 | tender-host 插件 |
| 并行拆活 | 内核原生 subagent / workflow，不再自研调度器 |
实施 / 投资与投标走同一套规则：落盘写成果，工作台不自建阶段运行器。阶段只准备 brief 和文件清单，并行扇出交给内核。
3.5.0 阶段记忆：阶段完成时先原子提交可校验的 StageMemory/Handoff；下一阶段只接收声明过的前序基线。主对话每轮只注入小型状态胶囊，精确事实按路径回读，因此重启、自动压缩或超长项目都不依赖聊天摘要充当业务底稿。项目事实只留在本项目，通用规范才进入知识库。
3.5.1 执行账本与双态控制：DSH 主智能体把目标、当前批次、计划、子任务、阻塞和下一动作回写到每会话执行账本；控制面板只维护磁盘事实、BOQ、证据、引用和人工门禁。面板同时展示执行态与事实态，并把差异定点送回主对话，不再靠机械扫描猜测智能体进度。
3.5.2 DSH alpha.3：长会话分页、渲染和图片投递能力升级；Agent Pi 同时订阅 Session 与 Chat projection，并在打包时拒绝已被上游删除的 SQLite 会话后端。内核指针今后每次变化都必须提升应用版本并通过 PR 门禁。
Agent Pi DSH is not a chat skin. It runs real jobs — bid parse and pricing, delivery planning, investment research. The workbench lives on chat-page tabs; the default path is still to open a workspace and talk.
| Layer | Role |
|---|---|
| Chat, models, permissions, skill catalog | dsh Web |
| Tender / delivery / investment workbench | Chat tabs + project creation |
| Stage prep, evidence gates, output tree | tender-host plugin |
| Parallel fan-out | Native subagent / workflow — no in-house scheduler |
Delivery and investment follow the same rules as tender: outputs land on disk, and the workbench never builds its own stage runner. Stages only prepare the brief and file list; fan-out belongs to the kernel.
3.5.0 stage memory: completing a stage first commits an atomic, validated StageMemory/Handoff; downstream stages consume only declared baselines. Each chat turn receives a small state capsule and reads exact facts back from disk, so restarts, automatic compaction and very large projects never depend on a chat summary as the business record. Project facts stay project-local; only reusable standards enter the knowledge base.
3.5.1 execution ledger and dual-state control: the DSH parent writes its objective, active batch, plan, assignments, blockers and next action into a per-session ledger. The workbench remains authoritative only for disk facts, BOQ, evidence, citations and human gates. It shows both states and sends exact mismatches back to the parent instead of guessing progress from files.
3.5.2 DSH alpha.3: long-session paging, rendering and image delivery are upgraded. Agent Pi subscribes to both Session and Chat projections and rejects the removed SQLite session backend during packaging. Every future kernel-pin change must raise the application version and pass the PR gate.
并行工人与长任务
Workers & long jobs
从 3.0 起，智能体循环交给 DeepSeek Harness：工具、并行子任务、会话、权限由内核直接跑。投标场景里大量持久 Bash 不再先卡三秒多才出字；推理强度可调到 low 省成本。
崩溃恢复：只救没递交的工人
工人正常写完并回推，父会话照常收口——这条路不加约束。父会话崩溃或应用重启后：
- 已经落到 Official Outputs 的任务不再重读 JSON、不再重派、不再重解析源文件；
- 只找回还没递交成果的工人：能续跑就续跑，续不上再下派这一条；
- 不会把整个阶段当新任务重开；
- 你在正式稿上改的字留在父会话里，不会把已完工工人叫醒。
原则：已完工任务的临时 JSON 不是权威——重启根据 Official Outputs 与已校验的 StageMemory/Handoff 判断是否还要救工人。
大历史不卡窗
上百个工人的父会话：点开某一个孩子的历史不应把整窗打死；冷子代理历史不再扫全部兄弟目录；大历史分页不再栈溢出；主对话上滚到头部自动「加载更早」。
Since 3.0 the loop belongs to DeepSeek Harness: tools, parallel sub-tasks, sessions and permissions run in the engine. Persistent Bash in tender jobs no longer stalls seconds per call; reasoning effort can be set to low to save cost.
Crash recovery: rescue only undelivered workers
A live worker that finishes is handled normally. After a parent crash or app restart:
- Tasks already in Official Outputs are never re-read, re-dispatched, or re-parsed;
- only undelivered workers are located — resumed when possible, re-dispatched otherwise;
- the stage is never restarted as a brand-new task;
- your edits to Official Outputs stay put and never wake finished workers.
Rule: a finished task's transient JSON is not authoritative — after restart, Official Outputs plus the validated StageMemory/Handoff decide whether a worker still needs rescuing.
Large histories stay responsive
Parent sessions with a hundred workers: opening one child's history should not freeze the window; cold subagent histories no longer scan every sibling; large-history pagination no longer overflows; scrolling to the top of the main chat auto-loads earlier messages.
证据门禁、正式成果与出处
Evidence gates, Official Outputs & citations
证据门禁
项目特征缺口不能用模型记忆填。要么在资料里找到出处，要么由你尽调后调用 tender_evidence force_pass 授权放行。每一阶段开工前，先备 brief 和文件清单。
Official Outputs
统一成果树与阶段总报告写回 Agent Pi Outputs。BOQ 逐页组价阶段在 Markdown 总报告之外，还会出一份带公式的 BOQ 组价测算.xlsx（出厂预置 Univer，可在对话里预览、改公式）。
引用是出处，不是摘抄
Markdown 里的引用令牌显示成短芯片。点一下只看到：源文件、页或行、题目或段落，需要时再打开源文件。写作合同禁止把证据正文贴进稿里——芯片是标注，原文在库里。
右侧资源文件栏
「资源文件」挂在根 overlay 上：单击打开预览，右键或回形针「注入对话」，可一路收到窗口右缘只留下图标轨。对话框旁边的文件夹芯片只提供地址——模型按路径自己列目录、按需读文件。聊天里的文档链接优先用本机 pwsh.exe（PowerShell 7）打开。
Evidence gates
Project-fact gaps are never filled from model memory. Either a source is found in the materials, or you authorize a pass after diligence via tender_evidence force_pass. Every stage starts from a prepared brief and file list.
Official Outputs
A unified output tree and stage reports are written back to Agent Pi Outputs. The page-by-page BOQ pricing stage also produces a formula-carrying BOQ pricing workbook.xlsx (Univer is preinstalled — preview and edit formulas right in chat).
Citations locate, they do not dump
Citation tokens in Markdown render as short chips. One click shows only the source file, page or lines, heading or paragraph — open the source when needed. The writing contract forbids pasting evidence text into the draft: the chip is a locator, the original stays in the library.
The files rail
The files rail hangs on the root overlay: click to preview, right-click or paperclip to “inject into chat”, and it collapses to the window edge as an icon track. The folder chip in the composer only hands over a path — the model lists the directory and reads files on demand. Document links in chat prefer the local pwsh.exe (PowerShell 7).
知识库
Knowledge base
入口在左侧栏「专业化工作台」下面的知识库按钮（工作台顶栏入口是同一页）。索引按文档自己的编排切条款：章 / 节 / 条、Clause / Article / Section、A1.2.3。检索是 MiniSearch BM25，不是向量库。
3.4.2 影子导航：长篇叙事 PDF / Markdown 会额外建立 PageIndex 兼容结构树，用于先定位应读章节；它不替代原文、MiniSearch、精确条款或 BOQ 表格。无需第二 API Key；真实项目审计门禁通过前，默认仍走 3.4.1 路径。
两条入库路，同一套索引
- 本页导入：选择文件或拖入，先落入
knowledge-base/originals/原始文档区，点「解析入库」才抽文本或走 MinerU（有进度条）。也可以贴路径：原文件、知识包文件夹、MinerU 产物文件夹。 - 对话导入：贴上文件后发送 「把这个 PDF 准确整理完整内容，做成知识包再入库。」对话只生成
…-知识包/；写好后在右侧对该文件夹或pack.json右键「一键导入知识包」。普通文件可右键「一键导入知识库」。
注意：只贴文件不说话不会进库——对话导入必须带上那句提示词（或提到知识库 / 入库 / 知识包 / 准确整理 / 完整内容 / 全文转录）。
对话里由模型自己出图再建知识包
宿主不先转 PDF。模型必须自己调 kb_prepare_document，再用官方 read_image 按印刷页复刻。不要用 vision_*，也不要自己截页。
用户模板
把你编好的较好文档入库为用户模板并勾选「本次任务选用」后，本轮业务稿复刻它的格式、大纲、章节顺序、标题语域和内容深度；数字、地名、合同号仍走规范、合同和本项目资料。文件名以「模板」结尾或带 template 时，右侧一键入库自动归入此类。
分类子目录与 .apkb 传递包
分类下可建子目录（如 规范 → COTO 2020）；已入库且文件名能认出的章节自动挂上。已入库条目、有文件的子目录、本机技能（不含出厂捆绑技能）都可导出 .apkb——这不是 zip，对方在同一页导入后，条目回到原分类和子目录，技能写入本机技能目录。
MinerU 表格能看
官方 MinerU 把表写成 HTML；入库落稿会把 <table> 收成标准 Markdown 表。旧稿点「全部重建」即可转表；预览对未重建的稿做兜底渲染。
The entry is the Knowledge Base button under “专业化工作台” in the left rail (the workbench header opens the same page). Indexing follows each document's own structure: chapters / sections / clauses, Clause / Article / Section, A1.2.3. Retrieval is MiniSearch BM25, not a vector store.
3.4.2 shadow navigation: long narrative PDF / Markdown sources also receive a PageIndex-compatible structure tree to locate chapters before reading. It never replaces originals, MiniSearch, exact clauses or BOQ tables. No second API key; the 3.4.1 path remains default until audited real-project gates pass.
Two import paths, one index
- On-page import: pick or drop files — they land in
knowledge-base/originals/first, and only “Parse into KB” extracts text or runs MinerU (with a progress bar). You can also paste a path: a file, a knowledge-pack folder, or a MinerU output folder. - Chat import: attach the file and send “把这个 PDF 准确整理完整内容，做成知识包再入库。” Chat produces only a
…-知识包/folder; then right-click it (or itspack.json) on the files rail → “Import knowledge pack”. Plain files get “Import into KB”.
Note: attaching a file without the prompt does not import anything — chat import requires the instruction (or a mention of 知识库 / 入库 / 知识包 / 准确整理 / 完整内容 / 全文转录).
The model prepares its own pages
The host never pre-converts the PDF. The model must call kb_prepare_document itself, then use the official read_image to transcribe printed pages. Do not use vision_*, and do not screenshot pages yourself.
User templates
Import a well-crafted document as a user template and tick “use for this task”: the next deliverable mirrors its format, outline, section order, heading register and depth — while figures, place names and contract numbers still come from the specs, the contract and this project's files. Files ending in “模板” or containing template auto-classify on one-click import.
Sub-folders and .apkb transfer packs
Categories can hold sub-folders (e.g. Specs → COTO 2020); imported chapters with recognizable filenames attach automatically. Entries, populated sub-folders and local skills (not factory-bundled ones) export as .apkb — not a zip; importing on the same page restores the original category and sub-folder, and skills land in the local skill directory.
Readable MinerU tables
Stock MinerU emits HTML tables; on import they are folded into standard Markdown tables. Rebuild older entries with “Rebuild all”; the preview falls back to rendering raw <table> for entries not yet rebuilt.
插件、模块与市场
Plugins, modules & the market
Harness 的企业能力：业务要什么，就按企业所需做成插件——技能、工具、工作台页、验收门禁都可以加。投标 / 实施 / 投资是第一批；合同、分包、物资、财务尽调用同一套装配方式往上叠，不必为了新工序换一套产品。
一句话蒸馏领域模块
模块管理的主按钮是「去对话里创造」：把本单活的打法一句话蒸馏成领域模块（workbench-domain-builder），沉淀为可复用的企业经验。JSON 粘贴收在高级里。
插件市场
设置 → 插件市场：发现社区为 Agent Pi 打造的能力，一键安装。从外部仓库装插件时，点确认会预写正确的 pnpm 构建放行键。对话里后来装的 bundle 插件，重启后保留。
内置能力
web_fetch：模型可取页面和图片地址，写成 Markdown 图；- AnySearch 网络搜索；
- 出厂预置 Univer：表格在对话里可预览、可改公式；
dsh-genui3.3.2 起不再出厂预装（已装的请在设置里卸载）。
The enterprise property of Harness: whatever the business needs becomes a plugin — skills, tools, workbench views, acceptance gates. Tender / delivery / investment are the first set; contracts, subcontracting, materials and financial diligence stack on the same assembly. You do not replace the product for every new procedure.
One-sentence module distillation
The main button in module management is “create in chat”: distill the playbook of the current job into a domain module in one sentence (workbench-domain-builder) and keep it as reusable company experience. Raw JSON paste lives under Advanced.
Plugin market
Settings → Plugin market: discover community capabilities for Agent Pi and install in one click. When installing from an external repository, confirming pre-writes the correct pnpm build-approval keys. Bundle plugins installed later in chat survive restarts.
Built-ins
web_fetch: the model fetches pages and image URLs into Markdown images;- AnySearch for web search;
- Univer preinstalled: preview and edit spreadsheets with formulas in chat;
dsh-genuiis no longer factory-preinstalled since 3.3.2 (uninstall in Settings if present).
模型、视觉与 Codex
Models, vision & Codex
内核钉在 DeepSeek Harness dsh-v0.1.2-alpha.3（dd6322d604），桌面壳 Electron 43.4.1；Agent Pi 兼容层不修改官方 DSH 内核。
- 新会话默认 DeepSeek-V4-Flash-Vision-Exp；Flash / Pro 仍是纯文本——只选它们看不见图片。已有会话沿用当时记在日志里的模型。
- 贴图：规范化 → Files API，失败再 inline。工作区图片用官方
read_image。系统提示不再点名已卸掉的vision_*。 - PDF 当文件读，能抽文本的不先转图；图纸和照片交给视觉模型。
- Codex 是子智能体，不是普通模型 Provider：设置页用 ChatGPT 登录后，DSH 主智能体通过
subagent_codex委派独立任务；凭据只在 Agent Pi 专属本机目录。
选模型口诀：要识图，选 Vision-Exp；只跑文本长任务，Flash / Pro 更省。推理强度可调到 low 进一步降本。
The kernel is pinned to DeepSeek Harness dsh-v0.1.2-alpha.3 (dd6322d604) on Electron 43.4.1; the Agent Pi compatibility layer does not modify the official DSH kernel.
- New sessions default to DeepSeek-V4-Flash-Vision-Exp; Flash / Pro are text-only — with those, images are invisible. Existing sessions keep the model recorded in their logs.
- Pasted images: normalize → Files API, inline as fallback. Workspace images use the official
read_image. The system prompt no longer names the removedvision_*tools. - PDFs are read as files — no image conversion when text is extractable; drawings and photos go to the vision model.
- Codex is a subagent, not a regular model provider: sign in with ChatGPT in Settings, then the DSH parent delegates self-contained tasks through
subagent_codex. Credentials stay in Agent Pi’s private local directory.
Rule of thumb: pick Vision-Exp when images matter; Flash / Pro is cheaper for long text jobs. Set reasoning effort to low to cut cost further.
版本沿革
Version lineage
| 版本 | 一句话 |
|---|---|
| 3.6.0 | DSH rc.1；DWG 只读预览；GPL 对应源码与十五项资产验证后发布 |
| 3.5.2 | DSH alpha.3 正式升级；长会话与图片投递适配；内核升级强制提升应用版本 |
| 3.5.1 | 每会话执行账本；主智能体执行态与系统事实态双态对照；差异驱动续派 |
| 3.5.0 | 不可变阶段记忆与 handoff；精准下游失效；重启与压缩后从磁盘基线恢复 |
| 3.4.2 | PageIndex 长文档影子导航；知识面路由；不可变证据包；五域覆盖账本 |
| 3.4.1 | DSH alpha.1 兼容；投标七阶段收敛；右侧文件栏与 Agent 模式恢复 |
| 3.3.5 | 投标执行链收紧；BOQ 全量覆盖；ChatGPT/Codex 登录；父会话自动接续 |
| 3.3.4 | 解析必须抽出实际 BOQ；企业工效优先；预览改价全局重算 |
| 3.3.3 | 南非道路投标五份深度分析稿；右侧多格式预览；Univer 软件内编辑 |
| 3.3.2 | 知识库完善（独立入口、知识包、用户模板、子目录、.apkb、MinerU 转表）；内核 0.1.1-rc.2；不再预装 GenUI |
| 3.3.1 | 内核 0.1.1-rc.1；官方视觉模型；卸 vision-router |
| 3.3.0 | 内核 0.1.0-rc.8；关 stealth，识图只留工具 |
| 3.2.3 | 对话框文件夹只给路径；右键恢复注入对话 |
| 3.2.2 | 卸掉 J-Space 出厂技能；不再劫持 DSH 循环 |
| 3.2.1 | 预装 dsh-genui、web_fetch、AnySearch；市场放行写对 pnpm 包名 |
| 3.2.0 | 内核 0.1.0-rc.7；崩溃只救未递交工人；引用变出处芯片；资源文件栏不再丢 |
| 3.1.0–3.1.3 | 蒸馏模块、知识库、成果树、质检；子代理历史加载热修 |
| 3.0.0 | 发动机换成 DeepSeek Harness |
| 2.6.5 | 经典版（Craft Agents OSS / Goal Loop），可并存 |
3.0 vs 2.x
| 2.x 经典版 | 3.x | |
|---|---|---|
| 内核 | Claude Agent SDK + Pi + 自研控制面 | DeepSeek Harness |
| 企业扩展 | 改控制面才能加流程 | 按企业所需做插件，流程随业务增长可高度定制 |
| 长任务 | Goal Loop 审查回合 | 内核原生 subagent / workflow，少绕一层 |
| 费用 | 控制面和循环更容易空转 | 同样活，回合更短，token 更省 |
| 图片 | 先识图再塞进对话 | 视觉模型走原生多模态；PDF 当文件读 |
| 数据目录 | ~/.agent-pi | %APPDATA%\agent-pi-dsh-desktop |
| Version | In one line |
|---|---|
| 3.6.0 | DSH rc.1, read-only DWG preview, and GPL source plus fifteen-asset verification before publication |
| 3.5.2 | DSH alpha.3 release, long-session and image-delivery adaptation, mandatory application version bump for kernel upgrades |
| 3.5.1 | Per-session execution ledger, dual main-agent/fact state, and mismatch-driven continuation |
| 3.5.0 | Immutable stage memory and handoffs, precise downstream invalidation, disk-baseline recovery after restart and compaction |
| 3.4.2 | PageIndex shadow navigation for long documents, typed knowledge-surface routing, immutable evidence packages, five-domain coverage ledger |
| 3.4.1 | DSH alpha.1 compatibility, converged seven-stage tender flow, restored files rail and agent modes |
| 3.3.5 | Hardened tender execution, full BOQ coverage, ChatGPT/Codex sign-in, and automatic parent wake |
| 3.3.4 | Analysis must extract the real BOQ; enterprise productivity first; reviewed rates recalculate |
| 3.3.3 | Five-manuscript SA road-tender analysis depth; multi-format right rail; in-app Univer editing |
| 3.3.2 | Knowledge base polish (own entry, knowledge packs, user templates, sub-folders, .apkb, MinerU tables); kernel 0.1.1-rc.2; GenUI no longer preinstalled |
| 3.3.1 | Kernel 0.1.1-rc.1; official vision model; vision-router removed |
| 3.3.0 | Kernel 0.1.0-rc.8; stealth off, vision via tools only |
| 3.2.3 | Composer folder hands over a path; right-click inject restored |
| 3.2.2 | J-Space removed; no more hijacking the DSH loop |
| 3.2.1 | Preinstalled dsh-genui, web_fetch, AnySearch; correct pnpm allow keys |
| 3.2.0 | Kernel 0.1.0-rc.7; crash rescues only undelivered workers; citations become locator chips; files rail fixed |
| 3.1.0–3.1.3 | Module distillation, knowledge base, output tree, QC; subagent history hotfixes |
| 3.0.0 | Engine swapped to DeepSeek Harness |
| 2.6.5 | Classic (Craft Agents OSS / Goal Loop), side-by-side install |
3.0 vs 2.x
| 2.x Classic | 3.x | |
|---|---|---|
| Engine | Claude Agent SDK + Pi + in-house control plane | DeepSeek Harness |
| Extensibility | New flows require control-plane changes | Plugins per company needs; workflows customizable as the business grows |
| Long jobs | Goal Loop review rounds | Native subagent / workflow, one less layer |
| Cost | Control plane idles more | Same job, fewer turns, fewer tokens |
| Images | OCR first, then into chat | Native multimodal vision; PDFs read as files |
| Data dir | ~/.agent-pi | %APPDATA%\agent-pi-dsh-desktop |
故障排查
Troubleshooting
SmartScreen 拦截安装
安装包未签名。选「更多信息」→「仍要运行」。
升级后行为没变 / app.asar 被占用
覆盖安装前没有完全退出——关到托盘不算。退出托盘里的 Agent Pi DSH，关掉占用安装目录的编辑器，再装一次。
贴了图但模型「看不见」
当前会话模型是 Flash / Pro（纯文本）。新开会话选 deepseek-v4-flash-vision-exp；已有会话沿用建会话时记下的模型。
对话导入没进知识库
只贴文件不说话不会入库。发送时带上「把这个 PDF 准确整理完整内容，做成知识包再入库。」，然后在右侧对 …-知识包/ 右键一键导入。
知识库预览里表格是 HTML
那是旧稿。在知识库点「全部重建」（或对该条再解析一次），<table> 会收成 Markdown 表。
目录菜单有点慢
上百工人的父会话里目录菜单仍可能慢，这是已知项；历史窗口本身不会卡死。
SmartScreen blocks the installer
The installer is unsigned. Choose “More info” → “Run anyway”.
Upgrade seems not to apply / app.asar locked
The app was not fully quit — closing to the tray is not enough. Quit Agent Pi DSH from the tray, close editors holding the install directory, and reinstall.
The model “can't see” pasted images
The session is on Flash / Pro (text-only). Start a new session with deepseek-v4-flash-vision-exp; existing sessions keep the model they were created with.
Chat import never reached the KB
Attaching a file without the prompt imports nothing. Send “把这个 PDF 准确整理完整内容，做成知识包再入库。”, then right-click the …-知识包/ folder on the files rail to import.
KB preview shows HTML tables
That's an older entry. Hit “Rebuild all” in the KB (or re-parse the entry) and <table> becomes a Markdown table.
The catalog menu feels slow
In parent sessions with a hundred workers the catalog menu can still be slow — a known trade-off; the history window itself will not freeze.