Agent Panel: allow ACP/tool/code transcript blocks to be collapsed by default

[myprestarocks]

Check for existing issues

• Completed

Describe the feature

Please add a user setting for the Agent Panel to collapse verbose external-agent transcript blocks by default, while keeping them expandable on click.

This should cover content that currently appears expanded inside ACP/external-agent conversations, especially:

• tool-call transcript blocks
• command/code-output transcript blocks
• markdown fenced code blocks emitted as assistant message content

Zed already has useful Agent Panel settings such as:

"agent": {
  "expand_edit_card": false,
  "expand_terminal_card": false,
  "thinking_display": "always_collapsed"
}

Those help for native edit/terminal/thinking cards, but external-agent ACP transcript blocks can still dominate the conversation view. In busy agent workflows, the user often wants to scan the natural-language parts first and expand code/output only when needed.

Suggested behavior

Add a setting along these lines:

"agent": {
  "collapse_tool_call_output_by_default": true,
  "collapse_markdown_code_blocks_by_default": true
}

Exact names are not important. The important behavior is:

• new verbose tool/code/transcript blocks start collapsed
• each block remains individually expandable
• the collapsed header shows enough context to identify it, for example tool name, language, command, first line, or output length
• existing expand_edit_card, expand_terminal_card, and thinking_display behavior remains unchanged
• this works for ACP/external agents as well as native Agent Panel content where applicable

Why this matters

When using multiple long-running agent threads, expanded command/code transcript blocks make the conversation hard to read. The important human-visible parts get pushed apart by implementation details and command output. Collapsed-by-default blocks would make the Agent Panel much easier to scan without losing access to the full output.

Current workaround

There does not appear to be a settings-only workaround for ACP transcript/code blocks. The existing collapse settings help with native cards, but not with ordinary ACP transcript/message content.

Zed version and platform

Observed on Zed 1.4.4 on Linux/Arch using external ACP agents.

[tredondo]

Yes please. I've been testing Claude Code via the VS Code extension vs. Zed ACP, and the experience is much better in VS Code thanks to tool blocks being collapsed by default.

In Zed, for long agent turns, the user has to:

• scroll up to the beginning of the agent response
• scroll down past length tool call block output, or line diffs
  • or collapse them along the way

Along with the lack of AskUserQuestion from ACP, this UX friction is what has kept me away from Zed lately.

[jbender0]

Yes, please! The long blocks of text are so distracting when I just need to see the overview. Having to wade through long scrolls to find the information I'm looking for strains my eyes.

[nertzy]

A third default state would make this much more useful than a binary collapsed/expanded choice:

• Collapsed - header only
• Preview - a fixed-height, internally scrollable output viewport
• Expanded - full output

This should apply consistently to every tool result that can be either tiny or huge, including native tools and ACP/external-agent transcript blocks.

For example:

{
  "agent": {
    "tool_output_display": "preview",
    "tool_output_preview_lines": 5,
  },
}

Exact setting names aside, preview mode should show approximately 4-5 lines by default, retain the complete output for scrolling, and offer a per-card control to collapse or fully expand it. For completed terminal-style output, anchoring the initial preview to the tail would surface the result or error; while a command is running, it should follow the latest output.

This is a display concern, separate from limiting output sent into model context as proposed in #59401. The full output remains available to the user without making every long tool call dominate the transcript.

[Achour]

Please fix this asap, I really want to start using Zed as my main ADE

[leviyevalex]

Agree this would be a great feature. Chat with agent feels cluttered at the moment

[mithun1999]

This would be a great feature. This is what preventing me from using Zed as my primary editor.

[kites262]

This feature would be really helpful.

[Git-I985]

Please do this; the extra blocks and UI elements are really cluttering up the interface and the agent panel.

[Achour]

zed team please do something

[cachebag]

Responding for reach. This makes it impossible for me to use Zed's agent panel.

[tredondo]

@yeskunall Any chance to get the team's attention?

[breathingcells]

And now, two days later, Zed has changed again:
the ACP Agent rendering has returned to showing the real commands.

Google's (fabulated?) explanation:

The Problem: Over those two days, Zed pushed a client-side update or an experimental flag
that forced tool calls into a collapsed "Summary Mode" to reduce thread clutter and save visual space.

The Resolution: Because many developers rely on seeing the raw shell commands, files being read,
or precise tool arguments to ensure the agent isn't executing destructive code, the change was swiftly rolled back

If this is true, thanks for the roll back!
In any case, I'm glad the real commands are back.

-----------Original post, 2 days ago---------------------

Something changed recently, and I hope this now has a control setting.

Context: External Agents: Claude Agent (external ACP-agent rendering)

Until recently (1.18.0?), when an agent ran a command, I could see that command.
This was especially helpful when the agent ran git commit: I could read the commit message.
Now it just shows "Commit", without any message.

The issue: Zed’s ACP client rendering shows only a short description instead of input.command.

Specifically (Windows 11), this prompt instruction produced nice readable git commit messages:

**Use the Bash tool with a heredoc for commit messages**. `git commit -F-` fed by `<<'EOF'` 
renders as readable text in the chat panel; a PowerShell here-string collapses to one unreadable line.

I understand why you'd want to HIDE command details, but it is also valuable to SHOW them.

Please make this an adjustable setting. Thanks!

[rahulnuthalapati]

yes, please add this functionality. its really hard to use claude in zed compared to vscode