feat(chat): the timeline shows the work instead of hiding it behind three clicks

[leonardoxr]

Problem

#5 made the omp payload rich — thinking, per-tool output, inline diffs, tool identity. Using it, the chat still read almost exactly like before. The reason was not the data: it was the timeline's default posture.

• MAX_VISIBLE_WORK_LOG_ENTRIES = 1 condensed every work group to its last row, including while the turn was still running.
• deriveTurnFolds hid the whole turn behind Worked for Xs the instant it settled.
• Thinking arrived as a one-line clickable stub; output and diffs lived in the row's disclosure.

Net effect: reading what the agent actually did took three clicks (open the fold → open +N previous → open the row). My own verification screenshots for #5 needed exactly those three clicks to prove the feature existed — a signal I ignored at the time.

Change

Five adjustments to what is visible by default. The model becomes: the current turn is fully visible; history folds away.

1. Thinking renders as prose in the flow, not as a tool row — its own timeline row (markdown, dimmed, italic), clamped to three lines with a Show thinking disclosure. Tool grouping breaks around it.
2. The current turn shows every work row. Condensation applies only to turns that are neither running nor the latest.
3. Short tool output (≤600 chars, ≤8 lines) renders under its row without a click. Larger output stays behind the disclosure.
4. Small diffs (≤2 files, ≤80 lines) render inline; bigger ones stay lazy so shiki is not paid for off-screen work.
5. The latest turn's fold starts open, keeping its duration header while leaving content on screen. The toggle set now records a deviation from each fold's default, so one set both opens an old turn and closes the current one.

The default-open rule deliberately does not apply in the window between sending a message and the server creating its turn: latestTurn still points at the previous settled turn there, so expanding it would pop content open and snap it shut a moment later. An existing test guards that window; it caught this during development.

Mobile gets the same five behaviors adapted to its feed, including height estimates so rows do not jump, and a plain +/- diff renderer (no new dependency). Expansion state lives in the feed's map rather than row-local state because that list recycles rows.

Before / after

Same thread, same data — only the default posture changed.

Before — thinking as one-line stubs, tool rows with no payload, output and diff behind the disclosure:

After — thinking as clamped prose, read output inline, colored diff inline, every tool row of the current turn present:

Another thread — two thinking blocks reading as prose, intent-titled tool rows, subagent CTA, all with zero clicks:

Verification

• 131 web tests (timeline logic + rows + session-logic) and 21 mobile threadActivity tests green; typecheck clean on web and mobile. New cases cover: thinking lifted out of tool groups, current turn uncondensed vs. older turns condensed, the XOR fold toggle, the inline output/diff thresholds, and the post-send no-flash window.
• Driven in a real browser against existing omp threads (no model calls spent): thinking, per-tool output and a colored inline diff all visible with zero clicks.

Perf notes: the condensation that protects long threads is unchanged for history — folds still hide older turns entirely, and LegendList still virtualizes. Inline diffs are threshold-gated so shiki stays lazy for anything substantial. No new animation; the thinking clamp is a static class swap.

Known, not fixed here: rows persisted before #5's title fix still read Tool because the summary was written at ingestion time. Deriving the title in the projection would repair old threads retroactively — separate change if wanted.

Built with Claude Fable 5 on Oh My Pi.