[tracking][dogfood] Build a full Agent Harness interface with Proto UI

Goal

Build a real Agent Harness interface as an internal-HEAD Proto UI dogfood application. Every eligible reusable or behavioral UI surface must be authored through Proto UI prototypes, compositions, modules, host capabilities, adapters, and design-language outputs.

This program does not adopt assistant-ui, CopilotKit, AG-UI UI runtimes, AI Elements, Radix, React Aria, shadcn primitives, or another component/headless Agent UI stack. Competitors and their issue trackers are research evidence only.

The Harness application remains the App Maker and owns Agent-domain data and service behavior such as sessions, threads, runs, messages, tool calls, approvals, artifacts, persistence, transport, retries, and execution. Those domain facts may drive application-local Proto UI prototypes. They must not be promoted into neutral Base subjects merely because the dogfood application needs them.

Completion definition

A Harness surface is dogfooded when:

• reusable interaction semantics, state transitions, accessibility projection, focus, scrolling, input, dismissal, and visual feedback are owned by Proto UI;
• application code supplies Agent-domain data, content, service callbacks, and persistence, rather than recreating a parallel UI state machine;
• no third-party component, headless primitive, chat runtime, or Agent UI framework owns an eligible UI path;
• any missing general capability is represented by a linked Proto UI Prototype, Composition, Module, Host Capability, Adapter, or spec Issue;
• every matrix row reaches a final state: dogfooded, app-local-proto, ready, research, blocked, native/static, or infrastructure-exempt.

This parent closes only when no ready, research, or blocked row remains and all required evidence is current.

Architectural boundary

Proto UI-owned interface

At minimum, inventory and dogfood:

• application shell, responsive navigation, sidebar, panels, tabs, menus, dialogs, sheets, and pane controls;
• session/thread list, search, rename, status, unread/running indicators, and empty/loading/error states;
• transcript viewport, message layout, streaming/busy feedback, reasoning disclosure, tool invocation, human approval/input, retry, and jump-to-latest behavior;
• Composer input, send/stop actions, keyboard policy, validation/error feedback, and attachment presentation;
• plan/todo, file/subagent hierarchy, artifact shelf, code/log view, diff review, Inspector, and related workbench controls;
• design-language styling and tokens without an application-owned parallel design system.

App-local Proto UI prototypes

Agent-specific subjects should begin here unless independent cross-application evidence later justifies extraction:

• HarnessMessage and message-part projection;
• ToolInvocation;
• ApprovalRequest / HumanInputRequest;
• ReasoningTrace;
• RunStatus / AgentLane;
• PlanView;
• ArtifactCard / DiffReview;
• branch/checkpoint navigation.

These prototypes may own application-domain projection and semantic events, but must compose official Proto UI controls and must not duplicate generic focus, keyboard, scrolling, collection, text-control, overlay, or accessibility state machines.

Native/static and infrastructure exemptions

The matrix may explicitly exempt:

• Agent backend execution, inference, transport, persistence, authentication, and service orchestration;
• Markdown parsing, syntax highlighting, diff computation, terminal emulation, or document-model engines when they do not own the surrounding reusable UI contract;
• plain static document content where the host already owns the complete behavior;
• adapter/demo infrastructure whose purpose is to compare Proto UI hosts.

Every exemption needs a reason, owner, and re-review trigger. An exemption must not hide user-facing keyboard, focus, scrolling, dismissal, selection, form, or ARIA behavior.

Agent Harness Dogfood Coverage Matrix

M0 must publish one authoritative matrix. Each row records:

• stable surface ID and application path;
• user job and current implementation owner;
• target class: official-prototype, composition, app-local-proto, native/static, or infrastructure-exempt;
• P/T/M/HC/Adapter/profile IDs and design-language entry where applicable;
• application-domain state owner and semantic events consumed/emitted;
• primary production host and React/Vue/WC equivalence evidence required for extracted reusable work;
• dependency Issue, owner, difficulty, milestone, state, and evidence links;
• temporary escape hatch and the public Issue that removes it.

Initial row families:

Family	Representative surfaces
Shell	app frame, responsive navigation, sidebar, workspace panes, mobile sheet
Sessions	thread list, search, rename, status, archive/delete actions
Run	run header, mode/status, cancel/retry, usage/trace summary
Transcript	viewport, Message, message parts, follow-tail, history prepend, jump-to-latest
Agent activity	reasoning, steps, tool invocation, subagent lane, progress
Human gates	approval, edit-and-approve, reject, generic questionnaire/form input
Composer	multiline editor, send/stop, IME-safe keyboard policy, errors, attachments
Workspace	plan/todo, file tree, artifacts, code/log, diff review, Inspector
Scale	virtualized lists, stable item identity, streaming update batching
Accessibility	live announcements, busy regions, focus restoration, keyboard order, zoom/reflow
Advanced host surfaces	terminal, document/editor, sandbox/preview

Report category totals separately. Do not publish a single completion percentage that hides blocked or exempt rows.

Program milestones

M0 — Inventory and baseline

• publish the coverage matrix and application surface inventory;
• define the consumer/import boundary and primary production host;
• capture baseline keyboard, accessibility, responsive-layout, streaming, long-list, remount/dispose, theme, and bundle evidence;
• split independently reviewable child Issues without duplicating existing #374, #388, #389, or #495.

M1 — Fixed-layout all-Proto vertical slice

Build a bounded three-region Harness: session list, transcript + Composer, and artifact tabs. Include User/Assistant messages, streaming text, Code Block, always-visible reasoning, Tool Invocation, inline Approval, Send/Stop, Error/Retry, and code/log artifacts.

M1 intentionally defers pane resizing, virtualization, file intake, Tree, full terminal/editor/browser surfaces, and branch editing.

M2 — Transcript and human interaction maturity

• Message Scroller follow-tail and anchor preservation;
• bounded Message, Code Block, and Composer compositions;
• Collapsible-backed reasoning/tool details after #388 permits the required slice;
• Questionnaire/review compositions;
• application-local Agent activity prototypes.

M3 — Workbench capabilities

• Input/Field-dependent search and forms after #389 prerequisites;
• hierarchy/Tree, artifact workspace, diff review, plan/todo;
• relative geometry and Resizable workbench after #495 prerequisites;
• attachments, Image, and File Intake after the relevant capability decisions, including #374 where applicable.

M4 — Scale, responsive behavior, and enforcement

• virtualized/windowed Collection and long transcript/log evidence;
• narrow/mobile layouts, keyboard-only journeys, high contrast, reduced motion, zoom/reflow;
• static checks preventing third-party UI ownership, raw internal imports outside an allowlist, and app-private generic state-machine drift;
• repeated mount/unmount/dispose, session switching, streaming updates, and stale callback tests.

M5 — Advanced host surfaces

Research and separately admit Terminal, text-document/editor, and sandbox/preview host domains only after the core Harness is useful. Do not overload Textarea or Code Block with those responsibilities.

Existing dependencies and non-duplicates

• #377 — authoritative prototype classification and publication governance.
• #500 / merged #501 — initially-private @proto.ui/compositions-chatui ownership decision; Message and Code Block remain independently reviewable and need separate implementation authorization.
• #388 — Collapsible/Accordion semantic work; this program must not open a duplicate Base Collapsible proposal.
• #389 — Base Input / Text Control prerequisite path; this program consumes it rather than creating a parallel input protocol.
• #374 — Image View semantic boundary; Attachment preview work must not invent a shadow Image owner.
• #495 — relative geometry, constrained drag-to-value, and interruptible motion research; Resizable/workbench work must reconcile with it.
• #420 — website self-hosting program; useful dogfood governance precedent but independent scope.

Child-Issue rules

Do not assign this parent as one implementation task. Each child must name:

• exact matrix rows and user journeys;
• current Proto UI entities and missing prerequisite;
• semantic owner and negative boundary;
• browser/adapter/accessibility evidence;
• non-goals and dependency Issues;
• whether the child is research, proposal/checkpoint, app-local dogfood, or authorized implementation.

Creating a child Issue does not itself authorize implementation where the repository requires a maintainer semantic or implementation checkpoint.

Acceptance criteria

Inventory and architecture

• Every user-facing Harness surface has exactly one matrix row and target class.
• No eligible surface is owned by a third-party component/headless/chat/Agent UI library.
• Application code contains no duplicate generic keyboard, pointer, focus, collection, scroll, dismissal, text-control, form, or ARIA state machine.
• App-local Agent prototypes are clearly separated from neutral Base and reusable compositions.
• Every generalizable gap has a public dependency Issue and removal path for temporary escapes.

Functional journeys

• Create/select a session, send a prompt, observe streaming output, stop a run, retry an error, inspect a tool call, resolve an approval, and open an artifact entirely through Proto UI-owned controls.
• Switching sessions during streaming does not leak state, focus, listeners, or stale callbacks.
• Transcript follow-tail pauses on user scroll and recovers through an explicit jump-to-latest control.
• Long histories/logs remain responsive after the scale milestone.
• Responsive layouts retain reachable controls and preserve the active work context.

Accessibility and host evidence

• Keyboard order, activation, disclosure, selection, Escape/dismissal, focus restoration, visible focus, accessible names/roles/states, and announcements are verified.
• IME composition is not mistaken for Composer submission.
• Busy and live-region behavior avoids token-by-token announcement spam.
• 320px, 390px, desktop, zoom/reflow, forced-colors/high-contrast, and reduced-motion evidence is recorded where applicable.
• Extracted reusable slices have real WC/React/Vue equivalence evidence; non-Web claims require separate host evidence.

Enforcement

• Static checks reject forbidden third-party UI imports and untracked matrix rows.
• Ordinary app UI consumes approved public/private dogfood surfaces rather than raw Core/Runtime/Module APIs outside a reviewed integration allowlist.
• Focused contract and browser tests cover every new Proto UI capability introduced by this program.
• The final matrix contains no ready, research, or blocked row.

Out of scope

• Making this program a v0.3 release blocker.
• Treating Agent-domain data/runtime semantics as automatic Base candidates.
• Building an Agent workflow designer before the end-user Harness is useful.
• Reimplementing an LLM/Agent runtime inside Proto UI.
• Claiming that a static Code Block is a terminal or that Textarea is a production code editor.
• Manufacturing prototypes only to improve a dogfood percentage.

Start with M0: publish the authoritative coverage matrix and application surface inventory, then define the consumer/import boundary and primary production host. Review the listed milestones, dependencies, and child-issue rules before selecting a bounded child; the program is done only when every matrix row has a final state and the required functional, accessibility, host, and enforcement evidence is current.