Runtime profile test cases

Use these cases when a product source is compatible with Agent Runtime, Lime AgentRuntime Profile, or an equivalent runtime spine. The goal is not to test the runtime implementation itself; the goal is to prove that Agent UI projects runtime facts without creating a second source of truth.

Canonical projection chain

RuntimeEvent / ThreadReadModel / TaskSnapshot / EvidencePack
  -> Agent UI adapter
  -> projection store
  -> status, task, tool, HITL, timeline, evidence, review, replay, and team surfaces
  -> controlled writes back to runtime, artifact, policy, or evidence owners

Agent UI tests fail when UI state invents status, approval, tool success, evidence verdict, known gaps, or task completion that cannot be traced back to runtime, artifact, policy, or evidence facts.

Source fixtures

When Agent Runtime fixtures are available, map them into Agent UI projection assertions:

Runtime fixture	UI surfaces to verify	Required projection outcome
submit-turn-event.json	Runtime status, message shell, task capsule	Accepted work appears before first text and preserves sessionId/threadId/turnId.
tool-approval-action-required-event.json	Human-in-the-loop, task attention, tool UI	Approval card uses actionId/toolCallId; UI does not run the tool optimistically.
task-retry-attempt-failed-event.json	Task capsule, timeline evidence	Failed attempt remains visible and retry state does not overwrite history.
routing-single-candidate-event.json	Runtime status, model chip, diagnostics	Single-candidate routing is explained as runtime fact, not final answer prose.
evidence-export-event.json	Timeline/evidence, review, replay	Evidence, replay, and review refs point to the same fact source.
thread-read-snapshot.json	Session hydration, status, task capsule	The UI can hydrate directly from snapshot without recomputing runtime truth.

Identity preservation

ID	Case	Input	Expected result
AUI-AR-ID-001	Preserve runtime spine ids	Source event carries runtimeId/sessionId/threadId/turnId	Projection keeps these ids on status, task, and timeline records.
AUI-AR-ID-002	Preserve task/run/attempt ids	Source event carries taskId/runId/attemptId	Task capsule and timeline can link the active attempt and prior attempts.
AUI-AR-ID-003	Preserve tool/action ids	Source event carries toolCallId/actionId	Tool row, approval card, and evidence timeline join through the same ids.
AUI-AR-ID-004	Preserve evidence trace ids	Source event carries evidenceId/traceId/evidencePackRef	Evidence surface links to durable details without copying the full payload.
AUI-AR-ID-005	Preserve parent/child lineage	Source event carries parentSessionId/parentThreadId/subagentId	Delegation graph and teammate transcript retain lineage after hydration.

Status and read model projection

ID	Case	Source facts	Expected result
AUI-AR-READ-001	Accepted before first text	turn.submitted or accepted read model	Runtime status shows accepted/preparing before text.delta.
AUI-AR-READ-002	Running turn	turn.started / active turn in read model	Status and task capsule show running; no final answer is fabricated.
AUI-AR-READ-003	Waiting on permission	action.required in read model	HITL surface appears and task attention state becomes waiting/needs-input.
AUI-AR-READ-004	Completed turn	turn.completed plus snapshot update	Status reconciles to completed and process details archive to timeline.
AUI-AR-READ-005	Failed turn	turn.failed with failure category	Failure is visible in status, task capsule, and timeline; final answer does not claim success.
AUI-AR-READ-006	Missing source field	Snapshot lacks optional routing or evidence summary	UI renders unknown/unavailable and keeps diagnostics safe.

Tool approval and controlled write cases

ID	Case	User action	Expected result
AUI-AR-ACTION-001	Approval request appears	Runtime emits action.required	UI shows approve/reject/respond controls with stable actionId.
AUI-AR-ACTION-002	Approval write is controlled	User clicks approve	UI calls runtime action response API and waits for action.resolved.
AUI-AR-ACTION-003	Denial does not execute tool	User clicks reject	UI waits for runtime denial/tool failure fact; no optimistic tool.result appears.
AUI-AR-ACTION-004	Duplicate response is safe	User repeats response	UI remains idempotent and does not create a second resolved fact.
AUI-AR-ACTION-005	Tool output is offloaded	Tool result includes large output ref	UI shows summary/ref and loads detail on demand.

Task, routing, and limit cases

ID	Case	Source facts	Expected result
AUI-AR-TASK-001	Retry keeps history	task.attempt.failed -> task.retrying -> task.attempt.started	Task capsule shows retrying/current attempt while timeline keeps failed attempt.
AUI-AR-TASK-002	Blocked task is visible	quota.blocked or routing.not_possible	Task capsule shows blocked/failed with reason, not infinite running.
AUI-AR-TASK-003	Single model candidate	routing.single_candidate	Runtime status/model chip explains selected model and decision source.
AUI-AR-TASK-004	Cost/limit state	cost.estimated, rate_limit.hit, or limit summary	UI shows cost/limit state as diagnostics or status, not final prose.
AUI-AR-TASK-005	Subagent lineage	subagent.spawned or parent/child task snapshot	Team/delegation surfaces show child owner and parent task.

Evidence, replay, and review cases

ID	Case	Source facts	Expected result
AUI-AR-EVID-001	Evidence export progress	evidence.changed with pending/exporting status	Timeline/evidence surface shows progress without blocking text streaming.
AUI-AR-EVID-002	Evidence pack ready	evidence.changed with evidencePackRef	Evidence surface links durable pack details.
AUI-AR-EVID-003	Replay and review share source	replayRef and reviewRef reference same pack/source ids	Replay/review lanes do not recompute their own status truth.
AUI-AR-EVID-004	Known gaps are not guessed	Runtime has no matching telemetry	UI shows unavailable/empty summary, not a fabricated unlinked gap.
AUI-AR-EVID-005	Failed tool is auditable	Tool failure has toolCallId and evidence ref	Timeline links failure to tool row and evidence detail.

Session hydration cases

ID	Case	Input	Expected result
AUI-AR-HYDRATE-001	Hydrate from thread snapshot	ThreadReadModel snapshot	Shell, status, pending actions, queued turns, and recent messages render without replaying all events.
AUI-AR-HYDRATE-002	Repair from event stream	Snapshot stale, event stream available	Projection rebuilds read model state and marks stale sections until repaired.
AUI-AR-HYDRATE-003	Evidence lazy load	Evidence refs exist but payload is not loaded	Timeline shows refs; payload loads only when requested.
AUI-AR-HYDRATE-004	Parent/child ids survive	Old session has subagent lineage	Delegation graph and teammate transcript can still resolve parent/child ids.

Governance failure cases

These are explicit failures for compatible Agent UI implementations:

1. UI parses assistant prose to infer tool success, approval state, model routing, or task completion.
2. UI projection store becomes the owner of runtimeStatus, evidence verdict, permission grant, or artifact contents.
3. Evidence, replay, and review lanes display contradictory status because each rebuilt facts separately.
4. A missing runtime field is silently replaced with a fabricated value instead of unknown, unavailable, stale, or safe diagnostics.
5. Action controls mark success before runtime confirmation.
6. Background, subagent, or remote work is flattened into one assistant transcript when runtime exposes ownership and lineage.

Minimum validation set

For an Agent Runtime-compatible source, run at least:

1. Identity preservation: AUI-AR-ID-001 through AUI-AR-ID-004.
2. Status/read model projection: AUI-AR-READ-001 through AUI-AR-READ-006.
3. HITL controlled writes: AUI-AR-ACTION-001 through AUI-AR-ACTION-004.
4. Evidence consistency: AUI-AR-EVID-001 through AUI-AR-EVID-004.
5. Hydration: AUI-AR-HYDRATE-001 and AUI-AR-HYDRATE-003.

These cases are the Agent UI counterpart to AgentRuntime profile tests: Runtime proves the facts exist; UI proves those facts are projected honestly.