[Adamaik]

Problem

Collapsing a turn's process also folded assistant narration (the body text a step emits between reasoning and tool calls). Once a long intermediate report scrolled past, it could not be read again: after the turn closed, isEarlierNarration() classified those body steps as process content, and the Reader rendered them inside ProcessFragment / RetiringContent disclosures. During a live turn the same text was also swallowed by the live fold (思考×N · 输出×M · 工具×K).

The reader also needs an explicit boundary between two narration blocks: previously nothing showed that a folded process run sat between them, so consecutive reports read as one block (or, in an earlier local iteration, body text was styled like annotation chrome).

Change

• Narration is always visible. Assistant body segments (text / image / other blocks) render as ordinary 16px primary body prose. They never enter a process disclosure, keep no annotation colour/indent/left rule, and only the final answer keeps copy/fork actions.
• The process rail moves out of the text. When collapsed reasoning/tool rows sat between two body blocks, the gap now carries a compact process-chain summary: a 1px --reader-rail rule plus the familiar 思考×N · 工具×K · 记录×M label and chevron. It counts only the rows folded between those two body blocks and expands the real rows in place when clicked (same disclosure language as the turn's folded-process row).
• hasProcessContent only advertises literal reasoning. Body-only steps no longer create an empty thinking disclosure; earlier narration is no longer foldable process content.
• Live fold chains break at narration. A new reasoning step may collapse the reasoning/tool run on its side of the narration, never the narration itself; fold summaries never count narration as 输出×N.
• Failure / interruption / blocked / unknown terminals and auto-fold-off behavior are unchanged.

Verification

• pnpm exec tsc -p tsconfig.json --noEmit — passes.
• node --import tsx/esm --test tests/*.test.ts — 188 tests, 187 pass, 0 fail, 1 skipped (the skip is pre-existing).
• Real-Reader browser fixture tests/browser-commentary.tsx via node scripts/verify-commentary-visible.mjs --mode=green — 11/11 checks pass, including:
  • completed/collapsed: report body visible without expanding, not inside any fold, final answer intact;
  • top disclosure expands reasoning/tool rows in source order between body blocks and collapses back without hiding the body;
  • open turn: narration not hidden by the live fold;
  • auto-fold off / failed / interrupted turns;
  • between-paragraph seam: 1px rail, 思考×N label, geometry strictly between adjacent body blocks, click expands reasoning and collapses back to seams.
• --mode=red reproduces the old behavior on both the 0.3.4 and 0.3.9 baselines (body absent from the DOM while collapsed, report text inside [data-reader-process], live fold counting it as 输出×1).

Stacking

This branch is stacked on #67 (cursor/better-display-021-rc2-ce8f), the still-draft Harness 0.2.1-rc.2 adaptation. Retarget to main once #67 merges; the commit diff here is only the narration/seam work on top of that branch.

[aa2246740]

Thanks! The feature commits from this PR (32b046a, 3e0119d, with your authorship) are on main through #76 (merge commit 5f77e6f). They were rebased onto current main without the Harness 0.2.1-rc.2 adaptation and lib/ rebuild commits, which are out of scope here, so GitHub won't mark this PR merged and I'm closing it by hand.

On top of your work we added:

• 53c0fe9: a live turn no longer draws a second summary row when the live-fold cell already counts the same steps ([0.3.11] 进行中的轮次里同一段过程出现两行相同摘要（live-fold 标题 + narration 接缝行重复统计） #73).
• 3b26887: the pinned selection stays on narration body blocks.

Real-host acceptance with real model turns passed: 4 reasoning runs and 3 tool calls, no duplicate seam in 9 live samples, and the closed seams expand in place. This ships in the next release.