Turn folding hides the substantive answer when a provider ends a turn with a short progress note (Grok)

Summary

deriveTurnFolds keeps exactly one entry visible per settled turn — the last assistant message, per deriveTerminalAssistantMessageIds. With providers that interleave short progress commentary with the actual answer, the last message of a turn is frequently a one-line "now I'll check X" note, so the real answer is folded away under Worked for Ns and the visible remainder is a fragment that reads like a non sequitur.

This is not the "collapse noisy tool calls" case the fold is designed for: the folded content is the response itself.

Environment

• Server t3@0.0.33 (headless t3 serve, Linux x64, Bun runtime)
• Desktop client T3Code(Alpha)/0.0.33, Electron 41.5.0, macOS
• Provider: Grok Build 1.0.4 (ACP over grok agent stdio), model grok-4.6

What happens

A Grok turn emits assistant messages as separate segments (assistant:…:segment:N) interleaved with tool calls. The turn commonly ends on a progress note rather than on the answer, e.g. turn a210c1f8:

segment	time	length	content
segment:5	10:07:34	3178 chars	the actual answer
segment:6	10:10:12	146 chars	a progress note ("looks like a shared session — checking where live dev redirects to")

segment:6 is the terminal message, so it stays visible; segment:5 is hidden behind Worked for ….

How common

One thread, 108 messages, 68 assistant messages across 36 turns. 19 turns had more than one assistant message; in 18 of those 19 the last message was not the longest. Terminal notes clustered at 77–146 chars while the answers they hid ran to 3178 chars.

Turn size distribution: 17 turns × 1 message, 14 × 2, 3 × 3, 1 × 5, 1 × 9.

Expected

The visible entry after folding should be the turn's substantive response. Folding should not be able to hide every assistant message that carries the answer.

Code

apps/web/src/components/chat/MessagesTimeline.logic.ts

• deriveTerminalAssistantMessageIds — lastAssistantMessageIdByResponseKey.set(responseKey, message.id) on every assistant message, so the winner is positional: whichever message came last.
• deriveTurnFolds — everything in the group except terminalEntry goes into hiddenEntryIds (only work entries with agentSpawn are exempt).

apps/mobile/src/lib/threadActivity.ts appears to carry the same rule, so mobile is affected identically.

Possible directions

1. Keep all assistant messages visible and fold only work entries — the fold then does what its docstring says ("commentary and tool activity") without depending on which message happens to be last.
2. Treat a trailing assistant message that is immediately followed by tool activity as commentary rather than as the terminal message.
3. Pick the terminal message by content rather than position.

Option 1 is the smallest change and does not need heuristics.

Notes

Reproduced against a live grok agent stdio session; the segmentation is the provider's normal output shape, not an error path.

Start with deriveTerminalAssistantMessageIds and deriveTurnFolds in apps/web/src/components/chat/MessagesTimeline.logic.ts, then compare the matching rule in apps/mobile/src/lib/threadActivity.ts. Reproduce a turn with multiple assistant segments and tool activity, and verify that folding no longer hides the substantive assistant response while still folding the intended work entries.