Skip to content

Source index

This page is the traceability register for Agent UI. Use the source ids below when changing the specification, acceptance scenarios, or schemas. The standard may adapt ideas from these sources, but it does not copy their APIs.

Last reviewed: 2026-05-11.

Citation format

Use this format in design notes or change logs:

text
[SRC-ID] -> requirement or behavior that changed

Example:

text
[SRC-AI-SDK-PARTS] supports ordered typed message parts; Agent UI requires active-run part order preservation.

External sources

Source idSourceEvidence usedAgent UI requirements informed
SRC-AGENTSKILLS-SPECAgent Skills specificationConcise protocol style: directory/field tables, constraints, examples, progressive disclosure, validation.Agent UI spec pages should use explicit fields, constraints, minimal examples, and validation checklists.
SRC-AGUI-EVENTSAG-UI EventsStreaming event categories, base event properties, run lifecycle, text message streaming, tool/state/activity event categories.Agent UI should classify lifecycle, text, tool, state, activity, and special events instead of treating the transcript as the only stream.
SRC-AI-SDK-PARTSVercel AI SDK UIMessageUIMessage uses parts for UI rendering and metadata.Message parts are typed and ordered; final answer text must not absorb process facts.
SRC-AI-SDK-V5-TOOLSVercel AI SDK 5.0 migration guideTool states include granular streaming lifecycle states such as input streaming, input available, output available, and output error.Tool lifecycle requires input/progress/output/error states and live rendering while running.
SRC-ASSISTANT-UI-PARTSassistant-ui MessagePrimitive.PartsMessage parts render by type: text, reasoning, audio, tool-call, data.Clients need typed part renderers and fallbacks.
SRC-ASSISTANT-UI-COTassistant-ui Chain of Thought guideReasoning and tool-call parts can be grouped without losing the underlying part model.Reasoning/tool grouping is a projection mode; it must not reorder or duplicate active facts.
SRC-LANGGRAPH-STREAMINGLangGraph JS streamingStream modes include updates, messages, custom, tools, and debug; tool events include start, progress, end, and error.Agent UI should support multiple stream modes, typed tool events, progress, and diagnostics.
SRC-LANGGRAPH-INTERRUPTSLangGraph interruptsWorkflows can pause for human review and resume with structured commands.Human-in-the-loop actions need stable ids, visible pending state, and controlled resume.
SRC-OPENAI-CHATKIT-THREAD-EVENTSOpenAI ChatKit thread stream eventsThread metadata events, thread item added/updated/done/removed/replaced events, progress updates, client effects, stream options.Session/thread metadata, item lifecycle, transient progress, and cancel options should be first-class projection facts.
SRC-OPENAI-APPS-SDKOpenAI Apps SDK referenceTool descriptors, structured tool results, component resources, and widget bridge concepts.Rich tool UI should attach to structured tool/component boundaries, not be inferred from assistant text.
SRC-COPILOTKITCopilotKit docsFrontend actions, generative UI, shared state, and human-in-the-loop patterns.UI tools and shared state updates need controlled runtime boundaries.
SRC-CLAUDE-ARTIFACTSClaude Artifacts helpSubstantial generated content opens in a dedicated artifact area outside the main conversation.Durable deliverables belong in Artifact Workspace with conversation references only.
SRC-A2A-PROTOCOLAgent2Agent Protocol specificationAgent Card, task lifecycle, messages, parts, artifacts, input-required/auth-required style remote progress.Remote teammates should map remote agent/task truth into Agent UI surfaces without replacing the local runtime owner model.
SRC-PAPERCLIP-HEARTBEATPaperclip heartbeat protocolBackground wake/sleep, heartbeat, task/background coordination patterns.Background agent work should be represented as scheduled/triggered teammate work; Paperclip-style hierarchy language is not normative for Agent UI.
SRC-VITEPRESS-PUBLICVitePress asset handling and deploy docspublic directory assets are copied to site root and repository-base deployments require base-aware URLs.Standalone runnable demos live under docs/public/examples/ and docs pages link to them instead of embedding demo components.

Local implementation research

These local sources were used as implementation references while drafting the flow taxonomy. They are not public normative dependencies, but they explain why certain UI behaviors are required for a desktop workbench.

Source idLocal sourceEvidence usedAgent UI requirements informed
SRC-LIME-ROADMAP-AGENTUILime docs/roadmap/agentui/Conversation/process/task/artifact/evidence architecture, event flow, hydration and performance constraints.Agent UI should cover the full workbench flow, not only thinking blocks.
SRC-CODEX-PROTOCOLLocal OpenAI Codex checkout: codex-rs/docs/protocol_v1.md and app-server protocol schemaTurn lifecycle, text deltas, plan deltas, approval/input requests, command output deltas, typed thread item categories.Standard taxonomy should include plan, reasoning, command execution, file changes, MCP/dynamic tools, web search, image view/generation, review mode, context compaction, and collaborative agent calls.
SRC-CLAUDECODE-LOCALLocal Claude Code checkout: message adapter and message rendering componentsStreaming assistant content blocks, thinking visibility policy, tool progress, tool result grouping, compact boundary handling.Active thinking/tool progress should stay visible, completed history should reduce noise, and tool results should not pollute final text.
SRC-CLAUDECODE-TEAMLocal Claude Code checkout: src/Task.ts, src/coordinator/coordinatorMode.ts, src/tasks/InProcessTeammateTask/types.ts, src/tasks/LocalAgentTask/LocalAgentTask.tsx, src/tasks/RemoteAgentTask/RemoteAgentTask.tsx, SDK hook schemas.Task types include local/remote agents and in-process teammates; coordinator mode treats worker results as internal notifications; teammate identity carries agentId, agentName, teamName, color, parent session, plan approval, permission, recent messages, pending user messages, idle/shutdown/progress.Agent UI v0.6 standardizes team roster, worker notifications, teammate transcript zoom, delegated plan/permission prompts, parent/child lineage, and coordinator synthesis vs worker result separation.
SRC-CODEX-COLLABLocal Codex checkout: codex-rs/tui/src/multi_agents.rs, codex-rs/protocol/src/protocol.rs, codex-rs/app-server-protocol/src/protocol/event_mapping.rs, codex-rs/analytics/src/*.Collaborative tools include spawn/send/resume/wait/close; TUI renders spawned/sent/waiting/closed history rows; subagents track source, parent thread id, nickname, role, depth; delegated approvals and subagent tool-call counts are tracked.Agent UI v0.6 adds controls for delegate/continue/wait/stop/close, parent/child thread metadata, delegated approval source, and parallel worker fanout/fanin surfaces.
SRC-LIME-TEAM-RUNTIMELime checkout: docs/aiprompts/task-agent-taxonomy.md, docs/aiprompts/state-history-telemetry.md, src-tauri/src/commands/aster_agent_cmd/subagent_runtime.rs, src-tauri/crates/agent/src/session_store.rs, src/components/agent/chat/teamWorkspaceRuntime.ts, src/lib/teamMemorySync.ts.Lime classifies execution into agent turn, subagent turn, and automation job; child subagent sessions expose role, profile, team preset, runtime status, queue and team phase; request telemetry joins session/thread/turn/pending/queued/subagent keys; team memory uses repo-scoped team.selection, team.subagents, and team.parent_context.Agent UI v0.6 aligns with Lime by treating teammates as child sessions/workbench facts, preserving request/session lineage, carrying runtimeEntity/queue/parallelism facts, and keeping background work within agent/subagent/automation boundaries instead of adding a fourth runtime taxonomy.
SRC-LIME-AGENTRUNTIME-PROFILEAgentRuntime project: docs/en/profiles/lime.md, docs/public/schemas/agentruntime-lime-profile-*.schema.json, and docs/public/fixtures/lime-profile/.Lime AgentRuntime Profile fixes the execution fact spine around RuntimeEvent, ThreadReadModel, TaskSnapshot, evidence export, runtime ids, tool approval, task retry, routing, and known-gap evidence behavior.Agent UI must project Agent Runtime facts without becoming a UI-owned truth source; runtime profile test cases verify id preservation, read model projection, action controls, evidence/replay/review consistency, and hydration.

Requirement traceability

Requirement areaPrimary sources
Ordered message partsSRC-AI-SDK-PARTS, SRC-ASSISTANT-UI-PARTS, SRC-ASSISTANT-UI-COT, SRC-CODEX-PROTOCOL
Running process visible, completed process archivedSRC-CLAUDECODE-LOCAL, SRC-CODEX-PROTOCOL, SRC-AI-SDK-V5-TOOLS
Tool lifecycle and progressSRC-AI-SDK-V5-TOOLS, SRC-LANGGRAPH-STREAMING, SRC-OPENAI-CHATKIT-THREAD-EVENTS
Human-in-the-loop actionsSRC-LANGGRAPH-INTERRUPTS, SRC-CODEX-PROTOCOL, SRC-COPILOTKIT
Session/thread hydrationSRC-OPENAI-CHATKIT-THREAD-EVENTS, SRC-LIME-ROADMAP-AGENTUI, SRC-CODEX-PROTOCOL
Artifact workspaceSRC-CLAUDE-ARTIFACTS, SRC-OPENAI-APPS-SDK, SRC-LIME-ROADMAP-AGENTUI
Evidence/timeline/replaySRC-LIME-ROADMAP-AGENTUI, SRC-CODEX-PROTOCOL, SRC-LIME-AGENTRUNTIME-PROFILE
Runtime profile projection testsSRC-LIME-AGENTRUNTIME-PROFILE, SRC-LIME-TEAM-RUNTIME, SRC-LIME-ROADMAP-AGENTUI
Spec writing styleSRC-AGENTSKILLS-SPEC
Team workbench and teammate identitySRC-CLAUDECODE-TEAM, SRC-CODEX-COLLAB, SRC-LIME-TEAM-RUNTIME
Remote teammate mappingSRC-A2A-PROTOCOL, SRC-CLAUDECODE-TEAM
Background teammate mappingSRC-PAPERCLIP-HEARTBEAT, SRC-LIME-TEAM-RUNTIME
Standalone runnable examplesSRC-VITEPRESS-PUBLIC

Draft runtime-first standard for agent interaction surfaces.