fix(desktop): collapse sub-agent trees and stream markdown in real time

[HUQIANTAO]

Summary

Two related desktop UX fixes for the main-v2 build, both surfacing during long sub-agent (/explore, /research) runs.

1. /explore's step tree could not be collapsed

The parent task tool card's chevron only hid its args/output; the nested sub-agent calls (read_file / grep / ls / glob / web_fetch) always rendered underneath. A 50-step /explore therefore flooded the transcript with its full step list, with no way to fold it back. Reproduce:

/explore 找到所有调用 Foo 的地方

→ 50 read_file cards pile up, no way to collapse the whole block to one line.

The chevron now toggles the whole subagent tree. Behavior:

• Open by default while the sub-agent is still running, so the user can watch it progress.
• Closed by default once it settles, so a finished explore folds to a single summary line ("12 steps") on first paint.
• A single click re-folds or re-expands the whole tree — and the user can still expand/collapse individual sub-calls independently.
• Writer tools (edit_file / write_file / multi_edit) keep their diff-rendered bodies, which were never affected by this bug.

2. Assistant markdown rendered all at once on turn_done

During streaming we deliberately showed raw text to avoid re-parsing markdown on every token, but the trade-off was that headings, lists, code blocks, and links all snapped into place on the closing message event. Watching a 2 000-token response was a wall of monospace-looking text that suddenly reflowed into a formatted document.

Switch to react-markdown for the live bubble too, with useDeferredValue marking the re-parse as a low-priority update:

• The cursor and scroll stay smooth because React commits the streaming flag (which controls the cursor) at high priority, before the heavy parse catches up.
• The deferred value converges on the full text when the stream ends, so the final render is byte-identical to the previous completion-time render.
• The AssistantMessage memo is preserved, so a per-token re-render still touches only the live bubble, not the backlog — the streaming-perf fix from fix: show sent desktop messages immediately #3056 stays intact.

Files

File	Change
desktop/frontend/src/components/Message.tsx	Replace the streaming ? raw : <Markdown> branch with a single <Markdown text={deferred}/> + cursor; add useDeferredValue and an explanatory comment.
desktop/frontend/src/components/ToolCard.tsx	Make the chevron's open state cover both hasBody and hasNested; default open to item.status === "running" && hasNested.

Diff stat

desktop/frontend/src/components/Message.tsx    | 26 ++++++++++++++------------
desktop/frontend/src/components/ToolCard.tsx   | 15 +++++++++++----
2 files changed, 25 insertions(+), 16 deletions(-)

Verification

• pnpm typecheck → clean.
• pnpm build → clean (✓ built in 1.86s).
• Manual: open the desktop client, run /explore over a 20+ file repo, confirm the steps tree collapses to one line on completion and re-expands on chevron click. Run a regular chat turn and confirm markdown formatting appears progressively as the response streams, with the cursor still blinking at the end.
• No new dependencies; no locale / i18n / CSS surface change.