Flow and taxonomy

This page is the complete lifecycle and taxonomy reference for Agent UI. It uses a specification style: fields, constraints, lifecycle stages, surfaces, team-workbench rules, and validation cases are listed directly. Background research is tracked in Source index.

Core contract

Agent UI is a projection protocol for agent workbenches. A compatible client consumes ordered runtime facts and projects them into user-visible surfaces without becoming the owner of those facts.

Compatible clients MUST:

• Preserve runtime event order for the active run.
• Keep final answer text separate from reasoning, tools, actions, artifacts, evidence, status, diagnostics, and team events.
• Keep running process visible while work is active; archive completed process into collapsed summaries by default.
• Preserve team ownership when the runtime exposes coordinators, teammates, child sessions, or remote agents.
• Route user writes through the owning runtime, policy, artifact, evidence, session, or team-control API.
• Render unknown or missing facts as unknown, unavailable, stale, or blocked instead of guessing from prose.
• Support progressive hydration for old sessions, child sessions, remote tasks, and long runs.

Compatible clients MUST NOT:

• Infer tool success, permission grants, artifact kind, evidence verdicts, approval state, or teammate completion from assistant prose.
• Duplicate the same runtime fact as both expanded inline process and expanded timeline detail on the same screen.
• Append final completion text after already streamed text without reconciliation.
• Treat UI collapse state, selected tab, active pane, or local draft as runtime truth.
• Treat internal worker notifications as real user messages, even if a source transport delivers them through a user-role channel.
• Collapse all workers into one anonymous assistant when the runtime exposes agent, task, thread, or team identity.

Lifecycle overview

The canonical flow is:

session/thread open
  -> composer draft
  -> listener bind
  -> submit or queue or steer
  -> run accepted
  -> active turn stream
  -> tool/action/team loop
  -> answer/artifact/evidence production
  -> final reconciliation
  -> timeline archive
  -> hydration/repair/replay

The flow applies to web apps, IDEs, desktop apps, terminals, and embedded assistants. The visual layout may differ; the projection rules do not.

Event envelope

Agent UI event adapters SHOULD normalize source events into an envelope with these fields. The public JSON schema is intentionally permissive so products can add fields, but these names are the portable classification layer.

Field	Required	Constraints	Purpose
type	Yes	String event class.	Drives reducer behavior.
sequence	Recommended	Monotonic within a run, thread, task, or child-agent stream.	Preserves active-run order and repair.
timestamp	Recommended	Producer timestamp when available.	Timeline, latency, replay.
sessionId	Recommended	Stable id.	Session and tab projection.
threadId	Recommended	Stable id when different from session.	Conversation recovery and branching.
runId	Recommended	Stable id for active work.	Runtime status and cancellation.
turnId	Recommended	Stable id for user turn or model turn.	Message/process grouping.
messageId	Conditional	Required for message parts.	Text reconciliation.
partId	Conditional	Required for ordered message parts when available.	Interleaved rendering.
taskId	Conditional	Required for background tasks, work items, or subagent tasks.	Task capsules, work board, task center.
agentId	Conditional	Required when the fact belongs to a child agent, teammate, or remote agent.	Team roster, delegation graph, transcript zoom.
parentSessionId	Conditional	Required for child sessions when available.	Parent/child lineage and hydration.
parentThreadId	Conditional	Required for child threads when available.	Delegation graph and replay.
toolCallId	Conditional	Required for tool lifecycle events.	Tool state and progress.
actionId	Conditional	Required for human-in-the-loop events.	Approval/input resolution.
artifactId	Conditional	Required for artifact events.	Artifact workspace routing.
evidenceId	Conditional	Required for evidence events.	Evidence/replay/review routing.
owner	Recommended	One of the fact owner values below.	Prevents UI ownership drift.
scope	Recommended	One of the scope values below.	Defines blast radius and persistence.
phase	Recommended	One of the phase values below.	Drives status and grouping.
surface	Optional	Primary intended surface.	Rendering hint, not fact owner.
persistence	Optional	One of the persistence values below.	Hydration and retention hint.
control	Optional	User action connected to the fact.	Controlled write mapping.
topology	Optional	One of the team topology values below.	Explains how work is organized.
payload	Optional	Structured data.	Producer-owned fact body.
refs	Optional	Array or map of ids/refs.	Link to artifact, evidence, transcript, files, or raw diagnostics.
rawEventRef	Optional	Safe reference, not raw secret-bearing payload.	Debug and replay without polluting UI.

Minimal example:

{
  "type": "agent.spawned",
  "sequence": 42,
  "sessionId": "session-lead",
  "threadId": "thread-lead",
  "taskId": "task-research",
  "agentId": "researcher@delivery-team",
  "parentSessionId": "session-lead",
  "parentThreadId": "thread-lead",
  "owner": "agent",
  "scope": "agent",
  "phase": "acting",
  "surface": "team_roster",
  "persistence": "snapshot",
  "topology": "coordinator_team",
  "payload": {
    "agentName": "researcher",
    "teamName": "delivery-team",
    "role": "researcher",
    "status": "running"
  }
}

Taxonomy dimensions

Fact owner

owner describes who writes the fact. UI projection may render the fact but must not become its writer.

Owner	Writer	Examples
runtime	Agent runtime or protocol adapter.	Run lifecycle, queue, interrupts, final outcome.
model	Model response adapter.	Text deltas, reasoning summaries, tool call requests.
tool	Tool runtime or tool adapter.	Tool input, progress, output, errors.
action	Runtime or policy action manager.	Approval requests, structured user input, plan decisions.
artifact	Artifact service.	Artifact id, preview, version, diff, export state.
evidence	Evidence, review, or replay service.	Citations, traces, verdicts, replay ids, review decisions.
context	Context, memory, or retrieval service.	Context refs, budgets, missing context, compaction.
policy	Policy, permission, sandbox, or security service.	Risk, permission mode, sandbox, waiver, retention.
task	Task scheduler, work-board service, or subagent runtime.	Queue item, background teammate, work item, subagent task.
agent	Agent runtime, team runtime, or remote-agent adapter.	Teammate identity, availability, handoff, transcript refs.
session	Session/history service.	Thread metadata, hydration cursor, stale state.
diagnostics	Runtime or client diagnostics channel.	Non-user-facing debug and performance records.
ui_projection	Client UI controller only.	Collapse state, focused tab, local draft, selected artifact.

Scope

scope describes the smallest stable entity affected by the fact.

Scope	Use when
application	Global health, provider availability, or account state.
workspace	Workspace-local context, policy, or artifact store state.
team	A named team, roster, team preset, or team memory boundary.
session	A chat session, task session, or app tab.
thread	A conversation thread that can be resumed or branched.
run	One runtime execution boundary.
turn	One user turn or model turn inside a thread.
message	One user, assistant, system, tool, or notification message.
part	One ordered message part.
task	A queue item, work item, background job, or subagent task.
agent	A child agent, collaborator, teammate, or remote agent.
tool_call	One tool invocation.
action_request	One human-in-the-loop request.
artifact	One durable deliverable.
evidence	One evidence, trace, replay, review, or citation entity.

Phase

phase describes where the fact belongs in the run lifecycle.

Phase	Meaning	Typical events
draft	User is composing local input.	state.delta for persisted draft only.
submitted	UI has submitted or queued input.	run.started, queue.changed.
accepted	Runtime accepted work.	run.status.
routing	Runtime is selecting model, mode, tool surface, worker, or teammate.	run.status, task.changed, team.changed.
preparing	Runtime is assembling context or request.	run.status, context.changed.
planning	Agent is producing or updating a plan.	plan.delta, plan.final.
reasoning	Reasoning/thinking is being produced.	reasoning.delta, reasoning.summary.
acting	Tool, command, browser, workflow, teammate, or subagent is running.	tool.*, task.changed, agent.changed.
waiting	Runtime is blocked on user, permission, dependency, teammate, or queue.	action.required, queue.changed, run.status, agent.changed.
reviewing	Reviewer or verifier is checking an outcome.	review.requested, review.completed, evidence.changed.
producing	Final text, artifact, or evidence is being produced.	text.delta, artifact.*, evidence.changed.
reconciling	Streamed and final facts are being reconciled.	text.final, messages.snapshot.
completed	Owning system reported success.	run.finished, tool.result, agent.completed.
failed	Owning system reported failure.	run.failed, tool.failed, artifact.failed, agent.completed.
cancelled	User or runtime cancelled work.	run.finished, task.changed.
interrupted	Work stopped and may resume from a bookmark.	run.finished, state.snapshot, agent.handoff.
archived	Completed detail is now represented by timeline summaries.	messages.snapshot, evidence.changed, worker.notification.
hydrating	Client is restoring state from snapshots/history.	session.hydrated, messages.snapshot.

Surface

surface is a rendering destination, not an owner.

Surface	Responsibility
composer	Draft, attachments, context chips, mode, queue/steer intent.
conversation	User messages and final assistant answer text.
inline_process	Active-run reasoning, tool progress, actions, and status in event order.
runtime_status	Accepted/routing/preparing/streaming/blocked/retrying/failed/done status.
tool_ui	Tool input summary, live progress, output preview, detail link.
hitl	Approval, reject, edit, answer, or structured input controls.
task_capsule	Running, queued, failed, needs-input, subagent, team, and background jobs.
artifact_workspace	Artifact preview, edit/canvas, version, diff, export, handoff.
timeline_evidence	Process archive, citations, verification, replay, review, audit.
session_tabs	Active, pinned, stale, unread, hydrated, running sessions.
diagnostics	Debug payloads, performance metrics, raw event refs.
team_roster	Human and agent teammates, role, source, capability, status, model, policy.
work_board	Human/agent work items, assignee, priority, blocker, progress, dependency.
delegation_graph	Parent/child tasks, coordinator-worker fanout, wait edges, handoff edges.
handoff_lane	Active owner transfer, handoff reason, resume target, memory boundary.
worker_notifications	Completed/failed/killed worker summaries, result refs, usage, duration.
review_lane	Reviewer/verifier verdicts, evidence links, requested fixes.
teammate_transcript	Zoomed teammate conversation, recent messages, tool activity, pending input.
background_teammate	Scheduled or triggered background agent as a teammate: wake reason, run record, sleep state.
remote_teammate	Remote agent card/capability, task status, input/auth needs, messages, artifacts.
team_policy	Per-teammate permission, approval, sandbox, plan mode, budget, and termination controls.

Persistence

Persistence	Meaning
ephemeral_live	Useful only while the run is active.
transcript	Rehydrates as part of conversation history.
snapshot	Stored as bounded recent state or session summary.
archive	Stored as timeline/process/team history.
artifact_store	Stored by artifact service.
evidence_pack	Stored by evidence/replay/review service.
diagnostics_log	Stored for debug or performance analysis only.
ui_local	Client-only local state.

User control

Control	Required write boundary
send	Runtime submit API.
queue	Runtime queue API.
steer	Runtime steer/resume API.
interrupt	Runtime interrupt API.
delegate	Runtime or team-control API that creates a teammate/subagent/remote task.
assign	Work-board or team-control API.
continue_agent	Runtime/team API that sends input to an existing teammate.
wait	Runtime/team API that waits for one or more teammates.
stop	Runtime/team API that stops a running teammate or task.
close	Runtime/team API that closes or archives a teammate.
request_review	Review/evidence service or runtime review API.
approve / reject	Runtime or policy action response API.
answer	Runtime action response API for structured input.
edit	Artifact service or runtime action API.
retry	Runtime retry, tool retry, artifact retry, or teammate retry API.
rollback	Artifact or session history API.
export	Artifact or evidence export API.
open_detail	Read-only session, artifact, evidence, transcript, or diagnostics API.

Standard event classes

The event class list is open to extension. These classes are the portable baseline.

Class	Purpose
session.opened	Client or runtime established a session/thread surface.
session.hydrated	Snapshot or history window has been applied.
session.updated	Title, status, unread, pinned, stale, or cursor changed.
session.closed	Session/tab closed or frozen by UI.
run.started	Runtime execution boundary started or was accepted.
run.status	Runtime status phase changed.
run.finished	Runtime finished with success, cancellation, or interrupt.
run.failed	Runtime failed and may expose retryability/diagnostics.
plan.delta / plan.final	Plan text or structured plan changed.
text.delta / text.final	Assistant answer text streamed and reconciled.
reasoning.delta / reasoning.summary	Reasoning or thinking streamed or summarized outside final text.
tool.started	Tool call boundary established.
tool.args	Tool input became available or changed while streaming.
tool.progress	Tool progress or partial output reference changed.
tool.output.delta	Tool output streamed in a typed channel.
tool.result	Tool completed successfully with output or output reference.
tool.failed	Tool completed with an error.
action.required / action.resolved	Runtime paused for user or policy decision, then resumed.
queue.changed	Queued turns, steer state, or queue order changed.
task.changed	Background task, subagent, work item, or team task changed.
agent.changed	Child agent, teammate, collaborator, or remote-agent status changed.
agent.spawned / agent.completed	Teammate/worker/remote collaborator started or returned.
agent.handoff	Active agent or task owner changed with an explicit reason and resume target.
team.changed	Team roster, selected team, work board, team memory, or team policy changed.
worker.notification	Worker result, failure, kill, usage, or summary notification arrived.
review.requested / review.completed	Reviewer/verifier request or verdict changed.
context.changed	Context selection, budget, retrieval, or missing context changed.
context.compaction.started / context.compaction.completed	Memory/context compaction boundary.
permission.changed	Policy, sandbox, approval, waiver, or risk projection changed.
artifact.created / artifact.updated	Artifact was created or updated.
artifact.preview.ready	Artifact preview became available.
artifact.version.created / artifact.diff.ready	Artifact version or diff became available.
artifact.export.started / artifact.export.completed	Artifact export lifecycle.
artifact.failed / artifact.deleted	Artifact failure or deletion.
artifact.changed	Collapsed adapter event when specific artifact events are unavailable.
evidence.changed	Citation, trace, verification, replay, review, audit, or source map changed.
state.snapshot / state.delta	External application state or durable runtime state synced.
messages.snapshot	Message history window or repair snapshot synced.
diagnostic.changed	Safe diagnostics changed.
metric.changed	Performance or responsiveness metric changed.

Message part taxonomy

Message parts are ordered. A client SHOULD render active parts in arrival order and may archive completed process into a collapsed timeline.

Part	Owner	Default surface	Rule
user_text	session or runtime	conversation	Show optimistically after submit and reconcile ids later.
assistant_text	model or runtime	conversation	Final answer only.
reasoning_summary	model or runtime	inline_process	Visible while running; collapsed after completion.
reasoning_detail	model or runtime	inline_process	Show only when provider policy and user setting allow it.
plan	model or runtime	inline_process or hitl	Use plan decision UI if approval is required.
runtime_status	runtime	runtime_status	Compact status, not prose.
tool_call	model or tool	tool_ui	Show safe input summary and running state.
tool_result	tool	tool_ui	Show preview/ref; load full output on demand.
action_required	action or policy	hitl	Explicit CTA with stable request id.
artifact_ref	artifact	artifact_workspace	Compact card in conversation/process; body in workspace.
evidence_ref	evidence	timeline_evidence	Citation/trace/review link only when backed by evidence facts.
context_event	context	inline_process or diagnostics	Budget, retrieval, missing context, or compaction summary.
agent_task	task	task_capsule or work_board	Subagent/team/background job summary.
agent_roster	agent	team_roster	Teammates, roles, status, and capability boundaries.
agent_handoff	agent	handoff_lane or delegation_graph	Active-agent transfer without rewriting transcript authorship.
worker_notification	agent or task	worker_notifications or timeline_evidence	Internal worker result; not a user message and not final prose.
remote_agent_status	agent or task	remote_teammate	Remote task, input/auth need, artifact update, or terminal state.
file_change	artifact or tool	artifact_workspace or timeline_evidence	Diff/review surface, not final prose.
error	Owning system	runtime_status, tool_ui, or task_capsule	Recoverable diagnostic and next action.
diagnostic	diagnostics	diagnostics	Hidden by default from normal transcript.

Tool lifecycle

A tool call SHOULD progress through these states when the source protocol provides them.

State	Event mapping	UI rule
input-streaming	tool.args	Show partial safe input only if useful and non-sensitive.
input-available	tool.args	Show stable safe input summary.
running	tool.started or tool.progress	Keep visible and expanded while active.
progress	tool.progress or tool.output.delta	Update row, progress bar, preview, or partial output ref.
output-available	tool.result	Collapse to completed row with output preview/ref.
output-error	tool.failed	Show recoverable error row.
cancelled	tool.failed or run.finished with cancellation.	Show quiet cancellation unless action is needed.

Large output MUST be offloaded to a detail ref or artifact/evidence ref. Raw secrets MUST be redacted before entering projection state.

Human-in-the-loop taxonomy

Action type	Owner	Required UI
tool_approval	policy or runtime	Approve/reject with tool name, scope, risk, and safe input.
plan_decision	runtime	Accept, reject, or request edits to a proposed plan.
teammate_plan_decision	agent or runtime	Identify requesting teammate and route response to that teammate, not the parent conversation.
structured_input	action	Form or options with stable schema and request id.
clarification	runtime	Answer prompt without pretending the run finished.
permission_grant	policy	Time/scope-limited permission state after runtime confirmation.
delegated_permission	policy or runtime	Approval card must show the delegated teammate/subagent that requested it.
credential_request	policy or secret store	Never expose secret values in message text or diagnostics.
artifact_review	artifact or evidence	Review diff, version, export, or handoff result.

The UI MUST mark an action resolved only after the owning runtime or policy system confirms the resolution.

Task and multi-agent taxonomy

Agent UI SHOULD classify long-running and multi-agent work separately from the conversation transcript.

Item	Scope	Default surface	Status values
Queued turn	task	task_capsule	queued, promoted, removed, started.
Work item	task	work_board	open, claimed, running, blocked, reviewing, done.
Background job	task	task_capsule or background_teammate	scheduled, sleeping, waking, running, blocked, paused, failed, completed.
Teammate/subagent	agent	team_roster, task_capsule, or session_tabs	spawning, running, waiting, needs_input, plan_ready, idle, failed, completed, killed, closed.
Coordinator team	team or agent	team_roster, delegation_graph, worker_notifications	forming, active, waiting, merging, done.
Specialist handoff	agent	handoff_lane or delegation_graph	handoff_requested, accepted, active, returned, resumed.
Review teammate	agent or evidence	review_lane	requested, reviewing, passed, changes_requested, failed.
Human/agent board	task or team	work_board	assigned, claimed, blocked, commented, done.
Remote teammate task	task or agent	remote_teammate	submitted, working, input_required, auth_required, completed, failed, canceled, rejected.

Conversation text may mention a task outcome, but task status itself belongs to task/runtime facts. Multi-agent status MUST NOT be flattened into a single assistant author if the source runtime exposes agent/task ownership.

Runtime execution alignment

Agent UI does not require every runtime to use the same internal taxonomy. It does require the UI projection to avoid inventing extra execution categories when the runtime already exposes a smaller set of durable entities. A compatible adapter SHOULD map runtime entities as follows:

Runtime entity	Agent UI mapping	Required fields	UI rule
Foreground agent turn	run.*, queue.changed, task.changed with runtimeEntity=agent_turn.	sessionId, threadId, turnId or runId, runtimeStatus, optional queuedTurnCount.	Show in runtime status, composer queue, and task capsule; do not create a separate teammate.
Child subagent or teammate turn	agent.spawned, agent.changed, task.changed, team.changed with runtimeEntity=subagent_turn.	agentId, agentName, parentSessionId, optional parentThreadId, teamName, runtimeStatus, latestTurnStatus.	Show as a teammate/child session with lineage; preserve coordinator-worker separation.
Durable automation job	task.changed and, when agent-owned, agent.changed with runtimeEntity=automation_job.	taskId, schedule or wake reason, runtimeStatus, last run ref, optional agentId.	Show as background teammate or task capsule; do not create a fourth local runtime taxonomy.
Remote protocol task	agent.changed, task.changed, artifact.changed, action.required with runtimeEntity=external_task.	agentId, remoteTaskId, remote capability/card ref, runtimeStatus, input/auth needs.	Preserve remote truth while projecting into remote_teammate.
Board work item	task.changed or team.changed with runtimeEntity=work_item.	workItemId, owner/assignee, status, blocker, artifact/evidence refs.	Show in work_board; assignment writes through board/team APIs, not runtime status hacks.

Adapters SHOULD preserve queue and parallelism facts such as teamPhase, teamParallelBudget, teamActiveCount, teamQueuedCount, providerConcurrencyGroup, providerParallelBudget, queueReason, and retryableOverload when the runtime provides them.

Team topology taxonomy

Agent UI SHOULD describe the topology that produced the work, because different topologies require different control surfaces. The primary metaphor is a team/workbench: teammates coordinate work, hand off ownership, request review, and preserve evidence. Hierarchy-first metaphors are implementation-specific and are not a required Agent UI topology.

Topology	Required UI projection	Typical facts
solo_run	One active run with process, tools, artifacts, and evidence.	run.*, tool.*, artifact.*.
coordinator_team	Coordinator plus roster, worker tasks, worker result notifications, and synthesis boundary.	team.changed, agent.spawned, worker.notification.
parallel_workers	Fanout/fanin graph, per-worker state, wait controls, partial failures, merge/retry controls.	agent.spawned, task.changed, worker.notification, agent.completed.
specialist_handoff	Active owner indicator, handoff reason, resume target, memory/context boundary.	agent.handoff, agent.changed, messages.snapshot.
review_team	Reviewer/verifier lane, verdicts, evidence links, requested fixes.	review.requested, review.completed, evidence.changed.
human_agent_board	Board of human and agent assignees with comments, blockers, claims, progress.	team.changed, task.changed, state.delta.
background_teammate	Wake reason, schedule, background run record, pause/resume, sleep state.	task.changed, agent.changed, evidence.changed.
remote_teammate	Remote Agent Card/capability, task lifecycle, input/auth required, messages, artifacts.	agent.changed, task.changed, artifact.changed, action.required.

Topology is metadata about how work is organized. It does not replace owner, scope, phase, surface, or persistence.

Team workbench contract

A compatible team workbench SHOULD project these facts when the runtime exposes them:

Fact	Required fields	UI rule
Teammate identity	agentId, agentName, teamName, role/source, optional color/avatar, model, policy.	Show in team_roster; do not merge into anonymous assistant.
Parent/child lineage	parentSessionId, parentThreadId, child sessionId/threadId, taskId, spawn reason.	Preserve in delegation_graph, history, evidence, and replay.
Worker notification	taskId, agentId, status, summary, result ref, optional usage/duration/tool count.	Show in worker_notifications; it may be archived, but must not masquerade as user speech.
Teammate status	runtimeStatus, latestTurnStatus, queue counts, plus needs_input, plan_ready, idle, shutdown_requested, closed when available.	Promote needs_input and plan_ready above normal running state; show queue/parallel limits without treating them as failure.
Teammate transcript	Transcript ref or bounded recent messages, tool activity, pending input queue.	Open on demand in teammate_transcript; keep bounded for performance.
Work board item	Title, assignee, status, blocker, priority, dependencies, artifact/evidence refs.	Show in work_board; user assignment writes through board/team API.
Handoff	From, to, reason, resume target, memory/context boundary, accepted time.	Show in handoff_lane; do not rewrite past message authorship.
Review verdict	Reviewer, target, verdict, evidence refs, requested fixes.	Show in review_lane; final answer may summarize, but verdict belongs to evidence/review facts.
Delegated approval	Action id, requesting teammate, policy scope, safe input, parent routing.	Approval card must identify the requesting teammate/subagent.
Remote teammate	Agent Card/capability ref, remote task id, status, messages, artifact updates, auth/input needs.	Map into remote_teammate; do not treat one idle tick as terminal completion.
Background teammate	Schedule/wake trigger, current run, last run record, pause/terminate controls.	Show as teammate-owned background work, not as a separate hierarchy.

Coordinator synthesis and worker results are different facts. The coordinator may summarize worker output for the user, but the worker result, usage, and transcript refs remain available as task/agent/evidence facts.

Context, memory, and compaction taxonomy

Fact	Owner	UI rule
Context reference	context	Show as context chip, source ref, or tool/detail link.
Context budget	context	Show low-noise budget meter or warning.
Missing context	context	Show blocked or unknown; do not fabricate availability.
Retrieval result	context or tool	Keep source ids/citations separate from final text.
Memory write	context	Show only if user-visible policy requires it.
Team memory	context or team	Show repo/team scoped memory refs; redact secrets; keep explicit conflict precedence.
Compaction boundary	context	Show compact boundary or summary; preserve resume metadata.

Permission and security taxonomy

Fact	Owner	UI rule
Risk level	policy	Show when it affects user control or required approval.
Sandbox/access mode	policy	Show active constraints and escalation requests.
Per-teammate permission mode	policy or agent	Show when teammate permissions differ from coordinator permissions.
Approval state	policy or runtime	Never infer from prose; require action confirmation.
Delegated approval source	policy or runtime	Preserve the teammate/subagent that triggered the approval.
Secret-bearing payload	policy or secret store	Do not persist in projection state. Use refs/redaction.
Retention/waiver	policy	Show only durable policy facts or evidence refs.

Hydration and repair

A compatible client SHOULD hydrate in this order:

1. Render shell, tab, title, and lightweight snapshot.
2. Apply recent message window.
3. Apply current run status, queue, pending action, task summary, and team summary.
4. Apply compact process/timeline/artifact/evidence/team references.
5. Load teammate transcripts, full tool output, artifact body, evidence payload, remote task details, and older history on demand.

Hydration events MUST be stale-safe. If the user switches sessions, late hydration results must not overwrite the active view.

Validation checklist

A compatible implementation SHOULD pass these checks:

1. Listener binding occurs before submit.
2. Runtime accepted/routing/preparing status can render before first text.
3. Active message parts render in event/part order.
4. Running reasoning and tools remain visible; completed process collapses into summaries.
5. Final answer text remains clean and is reconciled without duplication.
6. Tool input/progress/output/error state does not enter final answer text.
7. Human-in-the-loop actions use stable ids and controlled write APIs.
8. Queue and steer are visually and semantically distinct.
9. Artifact body lives in Artifact Workspace, not in the transcript by default.
10. Evidence, replay, review, and citations link to evidence facts.
11. Multi-agent and background tasks use task/agent/team facts, not assistant prose.
12. Team topology is visible when the runtime exposes coordinator, worker, handoff, board, background, review, or remote-agent facts.
13. Worker notifications are distinguishable from real user messages and coordinator final text.
14. Parent/child session/thread/task ids survive hydration, replay, evidence export, and transcript zoom.
15. Delegated approvals identify the teammate/subagent that requested the action.
16. Context and compaction are surfaced as facts or boundaries, not hidden text mutations.
17. Missing facts render honest fallback states.
18. Old sessions hydrate progressively and safely ignore stale results.
19. Diagnostics and metrics stay outside normal conversation text.