fix(agent-activity): align the stopped note and show it after the agent's reply - #2441

🤖 [P3] Protect the explicit thread layout as well as the default. This test omits messageLayout, so it only exercises the timeline fallback. Replacing the production data-layout expression with a hard-coded timeline still passes all 62 agent-activity tests, although thread layout has different spacing. Add a case with messageLayout="thread" and assert data-layout="thread" to catch loss of the pass-through. The existing assertion remains useful: it fails against the previous implementation.

🤖 Added in 042a1f2: ActivityAccessory.test.tsx › "notes a stopped request in the response lane" now runs for both the default (asserts data-layout="timeline") and an explicit messageLayout="thread" (asserts data-layout="thread"). Verified the thread case fails when the production expression is hard-coded to timeline.

🤖 [P2] Preserve thread scope when relocating a stopped note. If an agent posts two consecutive top-level public replies in a thread, the latest reply uses continuation layout. Stopping work on the root then hides its note through repliedAgents, while this relocated MessageActivity receives the root message and the reply’s continuation layout. visible() consequently applies the channel-only filter and removes thread-scoped work, leaving no stopped note in either location. A targeted render reproduction passes with the reply layout set to thread and fails with continuation (missing Blossom · Stopped). Derive visibility from the request/thread scope independently of the host row geometry, and add a rendered continuation-reply regression.

🤖 Good catch, thanks. Fixed in 6b7fa2b. The relocated note now gets its visibility from the reply's thread scope (threadScoped), not from the row's layout. I added a rendered regression, keeps a thread request's stopped note on a %s reply, covering both thread and continuation reply layouts. The continuation case fails without the fix. CI is green on 6b7fa2b.

🤖 [P3] Add coverage for the timeline's paired relocation wiring. Setting repliedAgents hides the note on the request, so the latest public reply must also receive the matching stoppedRequest. The accessory tests construct both instances directly, the presentation test checks only maps, and the browser no-reply case does not exercise this host wiring. Dropping or mis-keying this lookup would therefore silently remove the note without failing those tests. Add a ChannelTimeline regression asserting both the request's replied agent and the latest public reply's request ID, analogous to the ThreadPanel nesting test.

🤖 Added in c936280: moves a request's stopped note to its agent's latest public reply in ChannelTimeline.viewport.test.tsx. It renders ChannelTimeline with a request, an earlier agent reply, an activity step and a latest agent reply. It then checks that the request gets repliedAgents=[agent], the latest reply gets stoppedRequest=request, and the earlier reply gets nothing. Dropping the stoppedRequest lookup makes the test fail.

🤖 [P3] Cover the response-accessory forwarding hop. The ChannelTimeline and ThreadPanel tests mock MessageRow, while accessory tests supply props directly. Removing this stoppedRequest forwarding line still passes all 2769 tests under messages, agent-activity and conversation. If this forwarding regresses, repliedAgents still suppresses the request note while the latest reply receives no stopped note. Add one MessageRow.test.tsx case using the response-accessory registry harness around lines 1753–1773: capture accessory props and assert that both repliedAgents and stoppedRequest arrive. This is a regression-coverage gap; the current forwarding is present.

🤖 Added in dfaaf99: MessageRow.test.tsx › "forwards replied agents and the stopped request to response accessories" registers a response-placement accessory, captures its props, and asserts both repliedAgents and stoppedRequest arrive. Verified it fails when either forwarding line in MessageRow.tsx is removed.

convention, must-fix unless a human already approved it: this file is marked // FOUNDATION (line 1). The repo AGENTS.md requires explicit human guidance before editing FOUNDATION files. This PR adds repliedAgents and stoppedRequest to the plugin-facing ComposerAccessoryProps, and the description does not record that guidance. Please get it and note it in the description, or keep these props host-internal.

question (UX): the stopped agent can show twice for a moment. When an agent that already replied is stopped, the request keeps its leaving bubble for the leave animation (leaving/retained, ~lines 244-262). The reply's moved note mounts at once, so the bubble and "X · Stopped" overlap briefly. Before, the note waited for the bubble to leave (bubble.length ? [] : …). This is cosmetic. Is it acceptable?

advice (perf): each agent's latest reply row also mounts a second MessageActivity. That instance subscribes to session.agentActivity and runs messageWork on every store change, plus the recorded-agents, names and profile hooks. In agent-heavy channels, this roughly doubles the activity work per agent row. Suggestion: mount the inner component only when the request has stopped work for that agent, or derive the stopped set once in the host presentation.

nit (simplicity): this repeats the messageLayout === "thread" || threadRootId rule that visible() uses as its default (~lines 171-173), but evaluates it against the reply. A short comment, or computing threadScoped once in the outer component for both calls, removes the duplicated rule.

advice (simplicity/perf): rootReplies repeats the computation that renderReplies(undefined) already runs (~line 1467): same rows, agents and rootId. The only difference is pinnedActivity, which affects only retained. It also runs the full activityRuns + activityStatus/requestWork pass just to read two maps. Suggestion: memoize the top-level presentation once and use it in both places, or extract only the public-reply map.