10. WorkBuddy Benchmark — UX Spec Proposal

Source: hands-on walkthrough of WorkBuddy v5.3.5 (macOS, Electron) on 2026-07-25, captured via CDP screenshots. This doc records what WorkBuddy does UX-wise, and specs which patterns PI-Desktop should adopt, adapt, or reject. Baseline shell IA: 01-ui-ia.

1. What WorkBuddy is (observed)

A consumer/prosumer "AI work companion": chat-first desktop app whose unit of work is a Task (a conversation with a goal), organized under Spaces / Projects, extended by a marketplace of Experts (personas) · Skills · Connectors, plus Automations (scheduled tasks). It is persona- and template-heavy; PI-Desktop is developer- and workspace-heavy. The value here is interaction patterns, not product scope.

2. Observed IA

+----------------------------+------------------------------------------+
| Sidebar                    | Main pane                                |
|  收起 · 搜索 · 筛选 (top)   |  Per-destination header (tabs/actions)   |
|  新建任务                   |  Content                                 |
|  助理 (local assistant)     |                                          |
|  项目 (Projects)            |  Chat destination:                       |
|  专家·技能·连接器 (flyout)   |   transcript + floating composer          |
|  自动化 (Automations)       |   composer: + · 权限选择 · 模型选择 ·      |
|  更多 (flyout: 文件/知识库/  |            语音 · 发送                    |
|        灵感)                |   AI-content disclaimer footer            |
|  任务 (N)  — recent tasks   |                                          |
|  空间 (N)  — spaces         |                                          |
|  Footer: avatar · 通知 · 分享|                                          |
+----------------------------+------------------------------------------+

Notable per-screen details:

• Chat / Task: title in header with search / share / history / panel icons; per-message action row (copy, like/dislike, TTS, retry, share, more) plus a token-cost chip (共消耗 ◇2.85) and model badge inline; global disclaimer "内容由 AI 生成，请核实重要信息" under the composer.
• Composer: single pill; placeholder teaches syntax ("@ 引用对话文件， / 调用技能与指令"); left cluster + and a permission-mode dropdown ("默认权限"); right cluster model selector ("Auto"/named model), mic, send.
• 助理 (Assistant): a persistent "local assistant" chat with a connection status in the header ("已连接: 微信小程序" + gear) — the assistant is presented as an entity with integrations, not just a thread.
• 项目 (Projects): landing page with hero (title, one-line value prop, illustration), primary CTA "新建项目", "我的项目" card list with kebab menus, then a "从模版创建" template gallery (PRD flow, market research, team knowledge base, delivery, bug tracking).
• 自动化 (Automations): tabs 定时任务 / 运行记录; empty state (icon + one-liner + primary CTA); below it a template gallery of ~12 recipes (daily AI news, weekly report, meeting prep…) so the empty state still offers one-click value.
• 专家·技能·连接器: sidebar item opens a 3-way flyout; the page has segmented tabs (专家/技能/连接器), search, "我的专家", featured-scenario banner cards, category chip row, sort (综合/最热/最新), and rich persona cards (avatar, tagline, capability tags).
• 更多: overflow flyout for secondary destinations (我的文件, 腾讯文档, ima知识库, 乐享知识库, 灵感) — keeps the rail at ~6 primary items.
• 我的文件 (My Files): not a file manager — an artifact browser for task outputs. Tabs 任务成果 / 云端网盘 (cloud-drive sync); type filter, search across "文件、任务或工作空间", favorites toggle; a table (name / type / updated-by / updated-at / size) whose rows are grouped by the task that produced them (group header shows task name + count, e.g. "未分组 · 1个任务" → task → main.js / styles.css / index.html). Every deliverable an agent produces is findable later without reopening the conversation.
• 空间 (Spaces): a sidebar section between 任务 and the footer. A Space is bound to a local directory (hovering the space "tools" reveals a tooltip such as /Users/example/Projects/tools) and expands inline to list the tasks run inside it; row actions on hover: + (new task in this space) and … (manage). Tasks in both lists carry status badges — green dot = running, red ! = failed. Cloud/collab Projects (项目) also appear in this section, so "Space" is the umbrella for "place work happens", local folder or shared project alike.
• Onboarding: the first assistant message is a structured interview (name/style for the agent; how to address you, city, current focus) — personalization by conversation instead of a form wizard.

3. Adopt (spec changes for PI-Desktop)

Each item below is a concrete proposal; decisions go to the decisions log.

3.1 Permission mode in the composer

WorkBuddy puts the permission posture next to the send button. PI-Desktop exposes the same control as a composer chip (current effective mode: Ask every time / Accept edits / Auto) opening a small menu. The menu shows only those three modes and marks the effective selection directly; it does not surface global-default or inherit provenance. Choosing a mode creates a per-session override. This complements, rather than replaces, the inline permission cards in 03-permission-ux (D132).

3.2 Per-message meta: cost + model

Adopt the inline token/cost chip and model badge on assistant messages (collapsed by default into the message action row; hover to expand input/output/cached breakdown). Data already exists in the runtime usage events.

Adopted in D103 (tokens-only first cut): completed assistant turns show a model badge plus a compact Codex-style context inspector under the answer. The inspector shows remaining context percentage in a small ring and hover/focus reveals a remaining-token header, used-capacity meter, exact provider input/output/cache/reasoning usage, and generation tokens per second. The generation rate is a completed-turn snapshot rather than a live streaming counter. The context-window total comes from the same pi-ai model metadata used by the agent sidecar, with provider metadata and the default window only as fallbacks for unknown models. Each unique tool type has an explicit estimated badge plus aggregated argument/result footprint, call count, and cumulative duration. Repeated calls are grouped by tool name. Tool estimates are separate from exact provider totals because providers do not expose per-tool context allocation. Currency pricing remains deferred. Regenerate is available as a quiet action chip next to Copy and rewrites the current turn in place (D105); when multiple variants exist, a ChatGPT-style current / total pager on the root user turn restores archived branches (D109). Completed assistant rows also expose Fork and reversible Edit, while omitting Delete; both edits diverge into isolated sessions so the source cache/runtime stays untouched (D134).

3.3 Template galleries on empty states

WorkBuddy never ships a dead empty state: Automations and Projects both pair the empty hero with click-to-instantiate templates. Spec: Scheduled page empty state gains 4–6 developer recipes (nightly test run, dependency-update digest, PR review sweep, changelog draft); the Settings Project archive keeps its table but gains a "Start from template" row when empty.

3.4 Teaching placeholder in the composer

Replace the static placeholder with a syntax-teaching one: "Describe a task — @ to reference files, / for commands". The empty home keeps this direct-entry surface focused on the hero, optional onboarding, and bottom composer, without adding developer starter cards or a contextual quick-action layer (D204/D206).

3.5 Overflow "More" flyout for secondary destinations

Keep the sidebar rail at ≤6 primary items. As destinations grow (Logs, Files, Knowledge, future panels), park them under a "More" flyout instead of lengthening the rail.

3.6 Conversational onboarding (first-run)

Complement the checklist onboarding (05-onboarding) with a first assistant turn that interviews the user (project, preferred tone, guardrails) and writes results to workspace memory/config — a form disguised as a chat.

3.7 Artifacts view (from 我的文件)

Sessions produce files the user later can't find without scrolling the transcript. Adopted first step in D128: clicking a file artifact creates a path-keyed, closeable work-panel tab, and successful workspace Write/Edit artifacts open Review. Adopted in D179: the transcript also places a message-scoped review card directly after each successful file mutation; its status, +/− totals, and expandable hunks stay attached to that tool row rather than becoming a global footer entry. These renderer tabs and cards are transient views, not a second persistence model; the host-owned artifacts table remains authoritative. A future dedicated Artifacts destination may list files grouped by session with name / kind / session / time columns and Finder/diff actions. Skip the cloud-drive tab (no cloud storage in scope).

3.8 Workspace-scoped session tree (from 空间)

WorkBuddy nests tasks under the folder-bound Space they ran in, with + (new task here) on hover and status badges (running dot / failure mark) per task. Adopted and extended in D093: PI-Desktop renders every retained, path-keyed project as a compact, independently collapsible group, followed by a Temporary sessions group for path-less sessions. Each directory title is the single disclosure target (chevron, folder, label, and remaining row hit area), while + and overflow actions appear on hover/focus. The active project is still the only selected host workspace; retained groups do not create parallel workspace singletons. This supersedes D088's one-current-project sidebar limitation while preserving its exact-path and Temporary boundaries. Refined in D135: each conversation row uses a fixed leading status slot with distinct color and geometry: accent-blue selection ring, warning-orange breathing in-progress dot, success-green check, and error-red alert. In-progress outranks selection, selection outranks the latest terminal outcome, and reduced motion keeps the in-progress dot static.

3.9 Transcript density and user-plate alignment

WorkBuddy's task transcript keeps user turns as compact right-side plates and assistant turns as full-width transparent prose, with quiet hover actions under each turn. Adopted in D101: PI-Desktop keeps the developer-tool restraint (no mascot, no like/dislike) but densifies row spacing, right-aligns the user plate at min(78%, 560px), softens the plate border/shadow, and shows copy chips only on hover/focus-within. Streaming assistant answers use a thin accent left rule instead of a heavy pulse frame.

4. Adapt with caution

• Assistant-with-integrations header ("已连接: X"): good pattern for showing MCP/connector health per session; adapt as a small connector status cluster in the context panel, not the chat header.
• Persona marketplace: out of scope as a store, but the card grammar (avatar/icon, one-line tagline, capability tags, category chips, sort) is the right template for the Plugins page as the catalog grows.
• Voice input / TTS: note as future; not MVP.

5. Reject

• Mascot artwork floating over the transcript — conflicts with the restrained developer-tool aesthetic (07-ui-design-system).
• Like/dislike + share on every message — no backend to serve; keep the action row to copy / retry / (cost chip).
• Consumer template content (bedtime stories, wallpapers) — replaced by developer recipes in §3.3.
• Blanket AI disclaimer footer — redundant in a tool whose every output is inspectable; permission UX already carries the trust surface.

6. Open questions

1. Does the composer permission chip write through to Settings or stay session-scoped? (Proposed: session-scoped, Settings unchanged.)
2. Cost chip currency: tokens only, or provider-priced estimate?
3. Scheduled templates: hardcoded or plugin-contributable?