Agent Panel: collapsed tool-call cards still dominate the transcript

[Sergio-3]

Problem

Reading back a long agent session means scrolling through tool-call cards that occupy most of the vertical space even when collapsed, with the actual conversation (assistant text) scattered between them. Two independent causes:

1. A collapsed card hides only the output, never the input. A terminal card renders the complete command — for a heredoc (multi-line commit message, SQL, generated file content) that is easily 20+ lines while "collapsed". The screenshot below shows a single collapsed card at roughly 600 px. expand_terminal_card: false does not help; it only controls the output section's initial state.

2. Every tool call is its own full-width bordered card. A typical agent work stretch is 5–15 consecutive calls between two pieces of assistant text; each gets a card with border, header row and padding. The per-card overhead multiplies, and the assistant text — the part a human actually reads back — disappears between the cards.

Proposal

Two additive changes, independently useful:

(a) Truncate collapsed card input to one line. Collapsed state renders the first line of the command (or edit target) plus an ellipsis; expanding reveals the full input as today. This alone bounds every collapsed card to a constant height.

(b) An opt-in compact density mode that groups consecutive tool calls. Consecutive tool-call entries between two assistant messages render as a single borderless summary line, expandable into today's individual cards (which keep their own expand state, so drill-down is two-stage: group → card → full detail). Assistant text renders normally. Roughly:

Assistant: The migration is safe — only two rows change, both preserved verbatim.

  Ran 2 commands, read and edited migrate-checklists.mjs  +83 −40  ›

Assistant: All 75 records now carry the new format …

— one dim line replacing what is currently 3–4 full cards. This is the pattern claude.ai uses for its tool activity, and it makes long sessions skimmable without losing any drill-down capability.

(a) fixes the acute case even for users who prefer per-call cards; (b) is the structural fix for reviewability of long sessions.

[Vlaaaaaaad]

Related/duplicate of #60099 (proposal B)

[Sergio-3]

Good catch, thanks — my duplicate search missed it. Agreed that proposal (b) overlaps with #60099; I've added the grouping design from here as a comment over there so the discussion stays in one place.

Proposal (a) stands on its own though, even with aggregation in place: a collapsed card today still renders its full input — a heredoc command is 20+ lines while "collapsed" (screenshot above). Whatever happens to grouping, collapsed cards need a bounded height.