fix(mobile): stop the work log flickering during subagent runs and failing calls

[juliusmarminge]

Problem

Two kinds of jank in the mobile work log during a working turn, both visible in the reporter's recording of a Claude session:

1. The subagent row blinked out on every progress tick. Measured at 60 fps: a 5–7 frame blackout followed by a 10-frame fade-in, once per tool call the subagent made (30+ times in 36 s on a long-running agent). The batch row's anchor was the first task.progress activity. The server rewrites that activity in place with a new createdAt on every tick, and the row's activity-group id was the anchor's id — so the anchor flipped between task-progress:… and task-usage:…, LegendList remounted the row, and the "fresh row" entrance fade replayed from opacity 0.
2. The Thinking row thrashed the layout. In that session every Bash call failed within ~2 s. Each inProgress → failed transition switched the trailing group's shimmer off, which appended a separate Thinking row under it; the next call switched it back on and removed the row. Four insert/remove cycles in ten seconds, each with a layout transition on the list. Web doesn't have this problem because its live tool row and its Thinking row share one row id, so the handoff updates a single mounted row.

Beyond the flicker, the single-line "Kicked off 1 subagent · 1 working" row wasted the one thing a phone has room for: what the agent is doing right now.

Fix

• Agent task.started rows now anchor their batch (as on web). They have a fixed id and timestamp; progress ticks no longer change the row's identity.
• Spawn batches present as their own agent-spawn feed entry, keyed by the batch rather than the anchor activity. The card shows the agent's title (or "N subagents"), a status dot, and a status line that follows the newest member's progress detail ("Running Show git commit history") and updates in place; expanding lists the members. A working card is the turn's live activity, so no Thinking row sits under it. The launching Agent tool call folds into the card instead of rendering its own "Locate code" row above it.
• The trailing tool group's shimmering row and the Thinking row share LIVE_ACTIVITY_ROW_ID (mirrors web's live-activity-row), so a call failing changes that row's content instead of inserting a row.
• task.completed rows without a status default to completed (failed when error-toned), so a nested agent's terminal row settles its card.

isFreshRow still gates the entrance fade on createdAt; with the anchor fixed at spawn time that fade now plays once.

Verification

• vp test run apps/mobile/src/lib/threadActivity.test.ts (101 tests; 5 new: batch identity across ticks, card summary, agent tool call folding, single live slot across failing calls, thinking row under a settled card). tsc --noEmit for apps/mobile, targeted lint and format.
• iPhone 17e simulator, isolated server, one real Claude turn per build with the same prompt (five failing Bash calls, then one Explore subagent running ~25 shell commands). Pixel analysis of the subagent row at 60 fps: main 30 blackouts of 8–13 frames each in 36 s; this branch 0 across 1,595 frames while the agent worked.
• Also replayed the reporter's exact activity stream (232 activities, real timing) against the branch build; the card stays put through every tick.

Side by side during the subagent's run, main on the left:

https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/aa869a4a41b4e297/pair-subagent.mp4

Not addressed here: the message overlap in the reporter's screenshot (a user bubble drawn over the previous message's timestamp). That is a LegendList row-height issue independent of the work log; I could not reproduce it in these runs.

Implemented by Claude Fable 5 in the Claude Code harness.

Note

Fix work log flickering by unifying live-slot identity and rendering spawn batches as cards

• Replaces the old live-row logic in deriveThreadFeedPresentation and appendToolGroupRows with a single shared live-slot identifier, so a trailing tool group, a failed call, and a Thinking row no longer compete for separate live rows during a working turn.
• Folds the launcher tool call into the new ThreadAgentSpawnCard via collapseDerivedWorkLogEntries, which discards the matching non-task tool row so the spawn batch represents the launcher.
• Adds the AgentSpawnSummary model and agentSpawnSummary derivation to produce a batch title, live status from the newest working member, terminal outcome text, and per-member detail.
• ShimmerWorkContent gains a compact mode that uses caption text and skips the icon slot, fixing shimmer width calculation when no icon is shown.
• Behavioral Change: appendToolGroupRows now treats only a shimmering trailing group as occupying the live slot; earlier unfinished call groups remain separately grouped but no longer claim the live slot. A working spawn card suppresses the additional Thinking row for the unsettled turn.

^{Macroscope summarized 57ac2c2.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	13.6 KiB	13.6 KiB	+4 B (+0.0%)	15.1 KiB	✅
Codex	Thread snapshot wire	7.0 KiB	7.0 KiB	−3 B (−0.0%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.6 KiB	6.6 KiB	+7 B (+0.1%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	57.1 KiB	57.1 KiB	0 B (0.0%)	66.4 KiB	✅
Codex	Live turn messages	10	10	0 (0.0%)	21	✅
Claude	Total thread wire	13.6 KiB	13.6 KiB	+6 B (+0.0%)	15.1 KiB	✅
Claude	Thread snapshot wire	7.0 KiB	7.0 KiB	−12 B (−0.2%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.5 KiB	6.6 KiB	+18 B (+0.3%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	57.9 KiB	57.9 KiB	0 B (0.0%)	66.4 KiB	✅
Claude	Live turn messages	10	10	0 (0.0%)	21	✅

Baseline: ab67795 · PR result: 57ac2c2 · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 113.8 KiB
• Claude decoded thread snapshot: 114.5 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — The PR makes broad production changes to mobile work-log aggregation, row identity, animations, and nested-agent rendering, including a new expandable agent-spawn card. These changes affect existing subagent and failing-call paths beyond a narrowly isolated fix, so human review is warranted.

^{You can add or adjust custom eligibility rules. Learn more.}