@@ -34,6 +34,10 @@ applies_to:
3434 - packages/agent/src/process/detached-process-registry-live.ts
3535 - packages/agent/src/process/detached-process-registry.ts
3636 - packages/agent/src/projections/compaction.ts
37+ - packages/agent/src/projections/agent-lifecycle.ts
38+ - packages/agent/src/display-view/semantic.ts
39+ - packages/client-common/src/utils/actor-work.ts
40+ - packages/client-common/src/utils/model-request-progress.ts
3741 - packages/agent/tests/session-work-status.test.ts
3842 - cli/src/**
3943 - desktop/src/**
@@ -267,6 +271,26 @@ while that interval is open. Transient request progress controls loading and pre
267271owns, reconstructs, or adjusts the work timer. This keeps live and replayed completed timing
268272identical even when a chain crosses several model requests.
269273
274+ Completed root turns attribute the selected model display name independently from optional native
275+ decode telemetry. Work summaries therefore identify both local and cloud models. A summary includes
276+ decode tokens per second only when every contributing request in the chain supplied a valid native
277+ measurement; clients never estimate missing cloud throughput.
278+
279+ Clients distinguish an active root chain from an open productive interval. Authoritative model
280+ loading and request progress takes precedence when present. When a root is waiting for a provider
281+ that exposes no granular request progress and no worker keeps the productive interval open, clients
282+ show a generic model-wait state rather than presenting the paused accumulated duration as a running
283+ Working timer. Providers do not receive synthetic progress for presentation purposes.
284+
285+ Thinking activity follows the semantic response lifecycle. Thinking start opens the activity,
286+ thinking chunks refine its copy, and thinking end closes it immediately unless another
287+ independently active derived activity takes precedence. Assistant-message start and streamed answer
288+ text also close stale or interleaved thinking activity, so the live rail never claims the model is
289+ thinking while it is actively emitting its answer. The lifecycle projection stores only the
290+ authoritative response and tool lifecycle fields; display activity is derived from those fields at
291+ the display-view boundary rather than retained as a second synchronized copy. Clients never infer
292+ or retain thinking state from rendered text.
293+
270294Closing the final subscription removes the display registration. There is no separate close RPC.
271295
272296## Subscription protocol
0 commit comments