Agent panel: expand a turn live while the agent works, then collapse it into Thinking / Commands / Edits sections when done

[Ekleog-Polygon]

What are you proposing?

Render each agent turn in two phases, so a long conversation stays scannable
without losing any detail.

Phase 1 — while the agent is working (live)

Show the full turn as an interleaved, expanded transcript, exactly as it
happens:

• Thinking streams live.
• File edits appear inline as they're made, previewing the first ~20 lines
  of the diff with a "see more" at the bottom to expand the rest.
• Command runs appear inline, showing the last ~20 lines of output (tail)
  with a "see more" at the top to reveal earlier output.

This is the current live experience, just with previews instead of full dumps.

Phase 2 — once the final answer is ready (settled)

The turn collapses into three top-level collapsible sections plus the final
message:

1. Thinking — auto-collapses to a header. Expanding it shows the full
   interleaved transcript from Phase 1 (reasoning with the command/edit
   sub-cards inline, in order), because that context explains why the agent
   did what it did. The nested command/edit sub-cards inside it are collapsed
   to headers by default.
2. Commands run — a new section: one collapsed sub-card per command.
3. Files edited — a new section: one collapsed sub-card per edited file.

So a completed turn reads as:

▸ Thinking
▸ Commands run
▸ Files edited
<final message>

Everything stays one click away, but scrolling back through a long conversation
shows a clean, uniform summary of each turn instead of hundreds of interleaved
reasoning tokens, diffs, and command outputs.

This would be the default behavior. A setting could let people who prefer the
current always-expanded stream opt back into it.

Why does this matter?

In longer sessions the panel becomes very hard to navigate — you scroll past
huge expanded thinking blocks (expanded by default since #51525 / #52536), many
small read/edit cards, and command output just to find each turn's answer or to
remember "what actually changed." Collapsing a finished turn into Thinking /
Commands / Edits gives a stable overview and makes it trivial to answer "what
files did it touch?" or "what commands did it run?" at a glance — while keeping
the live, fully-expanded view intact while the agent is actually working.

Prior art

The "expand while active, collapse when done" pattern is well established:

• Zed already does this for thinking (agent_ui: Auto-expand and then collapse thinking blocks #51525) — this proposal generalizes
  that same pattern to the whole turn and adds the two filtered sections.
• Cursor: reasoning shows as a "Thought for Xs" bubble that auto-collapses
  the moment the response finishes; compact mode collapses diffs.
• Claude Code: reasoning collapses into a summary group when done (Ctrl+O
  for full), command output folds to +N lines (ctrl+o to expand).
• Codex CLI: folds command output to … +N lines (Ctrl+T for the full
  transcript) and has a proposed "semantic narrative view" over raw logs
  (Feature: Semantic narrative view as an alternative to raw execution logs openai/codex#13544).
• OpenCode: /thinking and /details toggles.

How this relates to existing requests

This unifies several open Feature Requests rather than competing with them:

• Option to hide agent thinking #55914 (hide thinking) and Regression: Claude "Thinking" blocks always expand by default in Claude Agent Panel (#22373) #52536 (a setting for thinking) → the Thinking
  section auto-collapsing when done.
• Group repeated Agent tool calls in the panel #56955 (group repeated tool calls) → the Commands section.
• any chance of introducing an option to have code edits collapsed by default? #33401 (collapse edits by default) and AI: Agent-chat panel code folding settings #39457 (folding) → the Files edited
  section.
• Agent UI: Collapsible Agent Responses #34158 (collapsible agent responses) → the per-turn collapse model.

It also builds directly on the existing agent.expand_edit_card and
agent.expand_terminal_card settings — adding (a) a live ~20-line preview state,
(b) the collapse-on-completion transition, and (c) the two type-filtered sections
on top of those primitives.

Notes

• The Thinking section intentionally also contains the command/edit sub-cards
  inline (duplicated with the dedicated sections) — that interleaving is the
  point: it preserves the order and rationale of the turn.
• The ~20-line preview counts are illustrative and could be configurable.
• Should apply to external ACP agents (claude-acp, codex-acp, …), which render
  in the same AcpThreadView as the native agent.

Zed version

Zed 1.4.4 1d0b482638588e5507184d91631e30596339ee1b (Linux)