apps/os: show the LLM request with its round, like mobile's Meta tab

Status summary: done — PR #2407 approved (Iterate linter + all bugbot/ review threads resolved), CI green, demo video + screenshots in the PR body. Includes the follow-ups: summary-status round headers and YAML results on both surfaces.

Follow-up from #2398 (merged). Misha (2026-08-04, with screenshot of a 9-round agent turn in the os feed):

where you see the model/llm request info above each "Ran code", let's move that into a "Meta" tab just like we did on apps/mobile. In fact, make the "Ran code" thing work just like in Mobile: "Ran code 9x" is expandable, expanding shows "Round 1", "Round 2", "Round 3" etc. Clicking "Round N" shows the same tabs as we have on mobile (Script | Approvals | Result | Meta I think, with some of them being optional). we don't need to use a shared component for now, but maybe make the structure similar across the two of them and use comments to point to each other so future agents understand that it's worth asking if I want to keep them in sync in future if changes are made

In apps/os today, an expanded activity is a flat rail of step buttons — one per llm request ("✦ openai/gpt-5.6-terra · 4.8k → 138 tok · 11.9 s") and one per code run ("Ran code · Started 14:52:33 · 5.7 s") — each opening a separate URL-backed inspector sheet. Mobile's activity card (#2398) instead groups the steps into ROUNDS (the llm step that writes a script + the code step that runs it), each a tabbed view Script | Approvals | Result | Meta.

Checklist#

• Promote groupActivityRounds from apps/mobile/src/lib/feed.ts into the shared agent-ui-reducer (packages/ui) so both surfaces group steps identically; mobile imports it from there (agent-ui-reducer.ts exports groupActivityRounds + AgentUiActivityRound; mobile's feed.ts re-exports)
• apps/os agent-feed.tsx: expanded activity renders "Round 1..N" toggle rows instead of the flat llm/code step buttons (new agent-activity-rounds.tsx: AgentActivityRounds/AgentActivityRoundRow; llm-only rounds render mobile's LlmStepView-style stat line instead)
• Expanding a round shows tabs Script | Result | Meta (structure mirrors mobile's CodeStepTabs; cross-pointing comments in both files) (RoundTabs in agent-activity-rounds.tsx; header comment points at mobile's activity-card.tsx and vice-versa via the shared reducer note)
• Script tab: the round's submitted code in the shared SourceCodeBlock
• Result tab (only once the run settled with a value or error): returned value via SerializedObjectCodeBlock (+ oversized-preview guard), error text
• Meta tab: one YAML doc (same shape as mobile's metaYaml — llm model / duration / tokens / outcome, code duration, replayed prompt:) rendered in SourceCodeBlock yaml; prompt replayed with replayLlmRequest over the events table (same query shape as llm-request-inspector-panel.tsx), fetched only while Meta is open (buildRoundMetaYaml in ~/lib/agent-round-meta-yaml.ts — lib module because component files must only export components for fast refresh; inactive base-ui tab panels unmount, so the query only runs while open)
• Keep the ?llmRequest / script-execution sheets as deep links: "Open full trace" affordances from the tabs; per-step inspector buttons go away with the flat rail ("Full trace" in Meta, "Execution trace" in Script — same data-testids as the old step rows)
• Live activity (AgentLiveActivity): settled rounds render as round rows; the currently-streaming llm step keeps LiveStepStream (rounds with a running code step auto-expand so the run stays watchable)
• Update agent-feed.test.tsx for the rounds layout + new tests for tab content (Meta yaml, optional Result) (12 tests passing; Meta yaml unit test on buildRoundMetaYaml)
• Round headers show the agent's summary status as of that round (Misha follow-up, 2026-08-04): reducer folds agent/summary-updated into state.summaryActivity and stamps activitySummary onto code steps (live while running, inherited at settle); os headers truncate, mobile round labels gain an ellipsized suffix (agent-ui-reducer.ts, agent-activity-rounds.tsx roundHeaderMeta, activity-card.tsx)
• Result tabs render YAML instead of JSON on both surfaces (same follow-up) (resultYaml in ~/lib/agent-round-meta-yaml.ts; mobile previewResultYaml keeps the 2000-char cap)

Assumptions (made while Misha was AFK-ish; flag if wrong)#

• No Approvals tab on web for now. Mobile's Approvals tab derives batches from root-stream approval events, which the os web feed doesn't have wired into StreamFeedView. The tab list keeps a commented slot so the surfaces stay structurally parallel; wiring approvals into the web feed is its own task.
• Rounds default collapsed when an activity has several; a single-round activity auto-expands its round (the user already clicked once to expand the activity).
• The standalone ?llmRequest sheet stays (deep-linkable, live streaming overlay, markdown mode, copy button — richer than the inline Meta). The inline Meta tab is the everyday path; the sheet is the full trace. Same for the script-execution sheet.
• No shared component between web and mobile (per the ask) — a shared pure grouping function in agent-ui-reducer is fine (it's data, not UI, and the reducer package is already the shared home for exactly this kind of interpretation).

Implementation log#

• Implemented in one pass; typecheck/lint/knip/format clean. Pre-existing unrelated failure on main: agent-prompt-budgets.test.ts (prompt 13 chars over ceiling) — left alone, flagged separately.
• Verified against local dev with a real 4-round agent turn; screenshots in PR #2407 (captured via throwaway gitignored spec specs/rounds-demo.ignoreme.spec.ts — pattern copied from agent-chat.spec.ts; gotcha: wait for the SETTLED activity row title exactly, the live toggle's title also starts with "Agent activity").
• Header suffix for code rounds is Started HH:MM:SS · duration (+ "Code failed" / "request failed" markers); llm-only rounds reuse the old stepLabel/stepMeta strings, so cancellation reasons read the same as the old flat rail.