生态洞察 · 002-deepseek-harness-v015
DeepSeek Harness v0.1.5 观察:平台收拢基础体验,生态分层正式成形
DSH Insights · 观察报告 #002 · 2026-09-11 · 数据快照可复核(/data/)· 兼容矩阵 = 权威集 10,534 个插件中已发布 npm 且带 web client 的 2,431 个 × 4 档 shell 逐插件实测 · 启发式评估,非安全审计
核心判断
DeepSeek Harness v0.1.5 是 dsh 平台史上一道分水岭:它标志着「基础体验内置化、差异化上移插件层」的生态分层正式成形。 文件树/预览进内核、左右侧栏标准化扩展入口、模型与 harness 共同训练 —— 平台在收拢地基的同时,给插件层留出了更明确、也更高门槛的位置。这个版本对生态的"破坏"几乎不在 API 层(兼容实测零破坏),而在定位层、沟通层与分发层:1522 个基础体验类插件需要重新回答"我为什么存在",却没有从官方得到任何回答这一问题的指引,而 npx 入口的 Node 静默坑则在无声损耗整个生态的新用户漏斗。
一、这个版本的核心价值:三条主线
1. 模型 × 平台共同进化成为产品形态。 V4.1 Flash 针对标准/PTC/极简三种模式专项训练,配合「保留 KV Cache 更新系统提示词」—— 模型不再是外挂而是被训练进 harness 的交互协议。这定义了 dsh 与通用 agent 框架的差异:模型行为与平台功能的耦合迭代。对插件作者的含义:靠近模型-平台协议的扩展点,比堆 UI 功能有更长的半衰期。
2. Agent 组织结构成为一等公民。 父子 agent 双向通信、排队/插话/停止、Agent Teams 共享任务列表(实验性、默认关闭)。多 agent 协作从"自己拼"变成平台原语 —— 这是插件生态的下一波机会带:团队编排、进度观测、成员间工具路由都还没有成熟插件。
3. 扩展入口标准化。 左右两侧 Sidebar 的标准注册位 + 多标签/分栏/全屏展示。入口的标准化通常先于生态的繁荣(参考浏览器扩展的历史),它降低了安装意愿门槛,但也把"占位"价值收归平台。
二、兼容性:一次被数据证明的平稳升级
权威集 10,534 个插件中,6,859 个尚未发布 npm(无法实测),3,675 个已发布者中 1,242 个为 CLI/TUI 形态(无 client bundle)—— 余下 2,431 个 web 形态已发布插件的最新版 client bundle,对四档 shell 的模块表逐插件实测:
| shell | ok | broken | conditional |
|---|---|---|---|
| 0.0.1-rc.5(旧 latest tag) | 2332 | 81 | 0 |
| 0.1.2-rc.1(主流存量线) | 2266 | 144 | 3 |
| 0.1.5-alpha.2 / rc.1 | 2266 | 144 | 3 |
0.1.2 → 0.1.5:零迁移(0 治愈、0 新破坏)。 模块表在两个版本间完全一致 —— 在 5 个 release、3 个含 breaking 的密集发布节奏下,平台守住了"不改删模块"的纪律,这是对生态最重要的承诺,值得被明确看见。
144 个存量 broken 里 132 个(92%)坏在同一个从未进入任何已发布 shell 模块表的模块(@deepseek-ai/dsh-client-runtime/client)。这与其说是兼容问题,不如说是生态知识传播问题:132 个作者在互不知情的情况下踩了同一个坑,而修法(迁移到 0.1.2-alpha.2 起收录的同名 API dsh-client-store)是确定且廉价的。一个坑被独立踩 132 次,说明生态缺少"修法知识"的分发层 —— 比模块本身更值得建设。
三、定位层冲击:1522 个插件需要重新回答"我是谁"
「侧栏 / 工作区」「文件浏览 / 预览」两类目共 1522 个插件(权威集 14%) 与本次内置功能正面重叠,其中头部三家(dsh-better-sidebar 4.3 万、dsh-univer-office 2.7 万、dsh-context 1.5 万周下载)承载着生态约 8.5 万/周的基础体验流量。
判断:这不是收编,是分层。 基础浏览/预览属于"用户预期开箱即得"的地基,平台收拢它对用户净收益为正;插件层的生存空间整体上移一格 —— 更深的格式支持、编辑能力、外部系统集成、与模型协议耦合的工作流。历史经验(浏览器扩展、VS Code)表明:入口标准化之后的三年,通常是插件生态总量的扩张期而非收缩期。但个体确实面临洗牌:与内置同质的插件,窗口期只到用户发现"内置已经有了"为止。
但上述乐观判断成立有一个前提:开发者相信投入还能回收 —— 而当前恰恰缺这一环。被重叠类目的作者面临的是三重不确定:不知道自己的下一块功能何时被吸收、不知道内置边界在哪里、不知道官方如何看待自己刚刚验证过的需求(头部侧栏/预览插件用数月积累证明了这些是真实需求,平台随即将其收为基础能力,却没有对这段"需求验证劳动"给出任何姿态 —— 哪怕一句致谢或推荐迁移方向)。
打击积极性的是不确定性,不是内置化本身。 理性的插件作者会在下次投入前先问"这会不会是下一个被吸收的",答案没人能给 —— 于是减少投入、转向观望,长尾创新枯竭,最终"一切皆插件"的愿景反被自己的成功侵蚀。这不是假设:吸收带头部作者的作品下周还会更新吗,是本站后续会持续跟踪的先行指标。
对官方的一个提醒:公告预告的「内置插件管理面板」是下一块落地区域,建议提前公布与现有插件管理类工具的边界,给生态留出调整时间。
四、分发层暗坑:整个生态的新用户漏斗在无声漏水
@deepseek-ai/dsh 0.1.5 的 CLI 入口以 if (import.meta.main) 守卫,该属性在较旧 Node 上为 undefined —— 所有命令零输出、退出码 0。实测:Node 22.14 / 23.11 / 24.0 / 24.1 静默失败,24.21 / 25.9 / 26.8 正常(精确引入版本待官方确认,见官方 Discussions #6124,社区已 3 例独立确认)。
影响的是生态整体而非 dsh 本体:一个潜在用户按公告执行 npx @deepseek-ai/dsh web,得到"什么都没发生",他的归因大概率不是"我的 Node 旧了"而是"dsh 不能用"—— 生态里所有插件的潜在装机量都从这同一个漏斗流过。修复成本极低(声明 engines + 启动守卫),收益是整个漏斗。
五、公告效应:温和的正向
DeepSeek Harness 官方公告发布当日(09-10),权威集按仓库创建日的新增:09-08 82 → 09-09 78 → 09-10(公告日)98,周内次高(次于 09-03 的 102)。公告带来了真实的创作者流入,但生态的主增量仍来自日常惯性 —— 单日脉冲不是这个生态的增长模式,持续的开发者体验才是。
六、给三方的建议
给 dsh 团队:① 发版即移正 dist-tag(npm latest 曾长期指向远古 0.0.1-rc.5);② 补 engines 声明与启动守卫,消灭静默失败(#6124);③ 若未来强制校验 engines.dsh,需先定义预发布语义 —— 标准字面解读会把 >=0.1.0-rc.6 判为不兼容 0.1.1-rc.x,与作者意图相反(已有作者因此弃用声明);④ 把"不改删模块表"的纪律写进公开承诺 —— 本版已经做到了,说出来价值更大;⑤ 内置化路线图提前 1~2 个版本预告吸收范围,并发布「内置边界宣言」:平台只收"用户预期开箱即得"的基础层,深度格式、编辑、跨系统集成永久留给插件 —— 给生态一颗定心丸; ⑥ 对被重叠的头部插件给出官方姿态:推荐替代、邀请共建、或至少公开致谢其需求验证价值;一个「被吸收插件迁移指南」的成本远低于流失一批头部作者的代价。
给插件作者:基础体验同质区(浏览/预览/简单侧栏)的窗口正在关闭,向"内置不做"的方向迁移:深度格式、编辑、跨系统集成、模型协议耦合的编排;依赖运行时能力探测而非静态版本声明;132 个存量坑的修法是确定的一行级迁移,早修早安。
给用户:升级兼容风险实测为零,放心升;升级前先 node --version(< 24.21 会静默失败,升 Node 即解)。
关于本站:DSH Insights(dsh-insights.com)是一个独立的 dsh 生态观察站:插件健康评级、实测兼容矩阵、修法案例库与定期的生态报告,数据、规则与管线全部开源,每条结论可回溯到原始数据。插件作者可以在详情页核对每一项评分依据、免费领取 README 健康徽章;本文的全部数字同样来自它的公开快照。
*数据与方法:兼容矩阵 = 最新版 tarball 的 client bundle 静态 require 提取 × shell 模块表三态判定(ok/conditional/broken),全缓存、每日增量;下载为 npm 周窗口;类目为六维评估体系的分类信号。原始数据:/data/compat-observed.json、/data/downloads.json、/data/fixes.json。DSH Insights 为独立观察站,与 DeepSeek 官方无隶属关系;判断部分基于启发式数据,欢迎反驳(提 issue 即可)。*
DeepSeek Harness v0.1.5, Observed: The Platform Absorbs Base Experiences, and the Ecosystem Stratifies
DSH Insights · Observation Report #002 · 2026-09-11 · Snapshot data verifiable (/data/) · Compatibility matrix = 2,431 npm-published web-form plugins out of the 10,534-plugin authoritative set × 4 shell versions, per-plugin load-tested · Heuristic evaluation, not a security audit
Core Judgment
v0.1.5 is a watershed in dsh's platform history: it marks the formal stratification of the ecosystem into "base experiences built in, differentiation pushed up to the plugin layer." File tree/preview into the kernel, standardized sidebar extension entries on both sides, models trained jointly with the harness — the platform is consolidating its foundation while leaving the plugin layer a clearer, and higher-bar, position. The damage this release does to the ecosystem is almost never at the API layer (measured compatibility damage: zero); it sits in the positioning layer, the communication layer, and the distribution layer: 1,522 base-experience plugins must re-answer "why do I exist" — with no guidance from the official side to help answer it — while the Node silent-failure trap at the npx entry quietly drains the entire ecosystem's new-user funnel.
1. The Release's Core Value: Three Storylines
1. Model × platform co-evolution becomes the product shape. V4.1 Flash is trained specifically for standard/PTC/minimal modes, plus "update the system prompt while preserving KV Cache" — the model is no longer bolted on but trained into the harness's interaction protocol. This defines dsh's difference from generic agent frameworks: coupled iteration of model behavior and platform capability. What it means for plugin authors: extension points close to the model-platform protocol have a longer half-life than piled-up UI features.
2. Agent organization becomes a first-class primitive. Bidirectional parent-child agent communication, queued messages/interruptions/stops, and Agent Teams' shared task lists (experimental, off by default). Multi-agent coordination goes from "roll your own" to platform primitive — this is the ecosystem's next opportunity band: team orchestration, progress observation, and inter-member tool routing have no mature plugins yet.
3. Standardized extension entries. Registered slots in both sidebars + tabs/splits/fullscreen presentation. Standardized entries historically precede ecosystem booms (see browser extensions) — they lower the installation-intent threshold, but they also fold "placement" value back into the platform.
2. Compatibility: A Smooth Upgrade, Proven by Data
Of the 10,534 authoritative-set plugins, 6,859 are not yet on npm (unmeasurable); of the 3,675 published ones, 1,242 are CLI/TUI-form (no client bundle). The remaining 2,431 web-form published plugins' latest client bundles were load-tested per-plugin against four shells' module tables:
| shell | ok | broken | conditional |
|---|---|---|---|
| 0.0.1-rc.5 (old latest tag) | 2332 | 81 | 0 |
| 0.1.2-rc.1 (mainline installed base) | 2266 | 144 | 3 |
| 0.1.5-alpha.2 / rc.1 | 2266 | 144 | 3 |
0.1.2 → 0.1.5: zero migration (0 healed, 0 newly broken). The module tables are identical across the two versions — under a dense cadence of 5 releases, 3 of them breaking, the platform kept the "never remove modules" discipline. That is the most important promise a platform can make an ecosystem, and it deserves to be seen.
Of the 144 long-broken plugins, 132 (92%) fail on a single module that has never been in any released shell's module table (@deepseek-ai/dsh-client-runtime/client). This is less a compatibility problem than an ecosystem knowledge-distribution problem: 132 authors hit the same pit independently, while the fix (migrate to dsh-client-store, seeded since 0.1.2-alpha.2 with the same-name API) is deterministic and cheap. A pit hit independently 132 times means the ecosystem lacks a distribution layer for fix knowledge — which is more worth building than the module itself.
3. The Positioning Shock: 1,522 Plugins Must Re-answer "Who Am I"
The "Sidebar / Workspace" and "File browsing / Preview" categories total 1,522 plugins (14% of the authoritative set) in direct overlap with this release's built-ins; the top three (dsh-better-sidebar 43k, dsh-univer-office 27k, dsh-context 15k weekly downloads) carry ~85k/week of the ecosystem's base-experience traffic.
Judgment: this is stratification, not absorption. Basic browsing/preview is foundation users expect out of the box; consolidating it is net-positive for users, and the plugin layer's viable space moves up one notch — deeper format support, editing, external-system integration, model-protocol-coupled workflows. History (browser extensions, VS Code) suggests the three years after entry standardization are usually an expansion period for plugin ecosystems, not contraction. But individuals do face a reshuffle: a plugin indistinguishable from the built-in has a window that ends the moment users notice "it's already built in."
That optimism holds on one precondition, however: developers believe their investment will still pay off — and that link is exactly what's missing today. Authors in the overlap zone face a triple uncertainty: they don't know when their next feature gets absorbed, where the built-in boundary is, or how the platform views the demand they just finished validating (top sidebar/preview plugins spent months proving these were real needs; the platform then took them in as base capabilities — without any gesture toward that demand-validation labor, not even an acknowledgement or a suggested migration direction).
What kills motivation is the uncertainty, not the absorption itself. A rational plugin author will ask "is this the next thing to be absorbed?" before every investment, and nobody can answer — so they invest less, shift to waiting, long-tail innovation dries up, and the "everything is a plugin" vision is eroded by its own success. This is not hypothetical: whether the absorption-zone's top authors keep shipping updates next week is a leading indicator this observatory will keep tracking.
A note to the official team: the announced "built-in plugin management panel" is the next absorption zone. Please publish its boundary with existing plugin-management tooling ahead of time and leave the ecosystem room to adjust.
4. The Distribution Trap: The Whole Ecosystem's New-User Funnel Is Leaking Silently
The @deepseek-ai/dsh 0.1.5 CLI entry is guarded by if (import.meta.main), which is undefined on older Node — every command produces zero output and exits 0. Measured: Node 22.14 / 23.11 / 24.0 / 24.1 fail silently; 24.21 / 25.9 / 26.8 work (exact introduction version pending official confirmation; see Discussions #6124, 3 independent community confirmations).
This hurts the ecosystem, not just dsh itself: a prospective user runs npx @deepseek-ai/dsh web per the announcement, gets nothing, and most likely concludes "dsh doesn't work" rather than "my Node is old" — every plugin's potential installs flow through this one funnel. The fix is trivial (declare engines + a startup guard); the payoff is the entire funnel.
5. Announcement Effect: Mildly Positive
On announcement day (09-10), authoritative-set additions by repo creation date: 09-08 82 → 09-09 78 → 09-10: 98 — the week's second-highest (after 09-03's 102). The announcement brought real creator inflow, but the ecosystem's main growth still comes from daily inertia — single-day pulses are not this ecosystem's growth mode; sustained developer experience is.
6. Recommendations for Three Parties
To the dsh team: ① move dist-tags the moment you ship (npm latest long pointed at the ancient 0.0.1-rc.5); ② add an engines declaration and a startup guard to kill silent failures (#6124); ③ if engines.dsh ever becomes enforced, define prerelease semantics first — literal semver reads >=0.1.0-rc.6 as incompatible with 0.1.1-rc.x, the opposite of author intent (one author already dropped the declaration over this); ④ make the "never remove modules" discipline a public commitment — you already kept it this release; saying so out loud is worth more; ⑤ pre-announce the absorption roadmap 1–2 releases ahead and publish a "built-in boundary manifesto": the platform takes only the open-out-of-the-box base layer; depth, editing, and integration permanently belong to plugins — give the ecosystem certainty; ⑥ take an official posture toward overlapped head plugins: recommend alternatives, invite co-building, or at least publicly acknowledge their demand-validation work. A migration guide costs far less than losing a cohort of head authors.
To plugin authors: the window for the base-experience commodity zone (browsing/preview/simple sidebars) is closing — move toward what the built-ins won't do: deep formats, editing, cross-system integration, protocol-coupled orchestration; prefer runtime capability detection over static version declarations; and the 132-pit fix is a deterministic one-line migration — the earlier, the better.
To users: measured upgrade compatibility risk is zero — upgrade with confidence; check node --version first (< 24.21 fails silently; a Node upgrade resolves it).
About this site: DSH Insights (dsh-insights.com) is an independent observatory of the dsh ecosystem: plugin health grades, a measured compatibility matrix, a known-fix case library, and periodic ecosystem reports — data, rules, and pipeline fully open source, every conclusion traceable to raw data. Plugin authors can audit every scoring basis on their detail page and pick up a free README health badge; every number in this report comes from its public snapshots.
*Data & method: the compatibility matrix = static require-literal extraction from each latest-published client bundle × shell module-table three-state verdict (ok/conditional/broken), fully cached, incrementally refreshed daily; downloads are npm weekly windows; categories come from the six-dimension scoring framework. Raw data: /data/compat-observed.json, /data/downloads.json, /data/fixes.json. DSH Insights is an independent observatory with no affiliation to DeepSeek; the judgment sections are heuristic and welcome rebuttal (open an issue).*