feat(ui): compact or collapse DevSpace tool calls in ChatGPT Web #338

Problem

Long DevSpace sessions in ChatGPT Web can generate many consecutive MCP tool calls (for example open_workspace, read, exec_command, write_stdin, patches, verification calls, etc.). Each call/result is rendered as a separate visible block in the conversation.

For multi-step coding or research tasks this makes the ChatGPT page very long and visually noisy. The actual assistant answer can become difficult to locate, and scrolling back through a session is unnecessarily expensive.

I searched the current issues/PRs for similar requests around collapsible/compact tool calls, ChatGPT result cards, and repeated MCP output, but did not find an issue directly addressing this UX problem. Related work such as #128 and #142 reduces repeated context / simplifies MCP orchestration, but does not solve the accumulated visible tool-call blocks in ChatGPT Web.

Requested behavior

It would be useful if completed DevSpace tool activity could be presented in a compact, collapsible form in ChatGPT Web.

A possible UX:

• Consecutive DevSpace calls from one assistant turn are grouped into one compact section, e.g. DevSpace · 12 calls.
• The group is collapsed by default after the calls finish.
• Clicking it expands individual calls and full outputs for debugging.
• Errors, approval prompts, review checkpoints, and other user-action-required states remain visible/expanded.
• Large command output is collapsed or summarized by default, while the full raw result remains available on demand.
• Optionally group by workspaceId so activity from different workspaces remains distinguishable.

Even a simpler first step—making each completed DevSpace result card individually collapsible by default—would substantially improve long ChatGPT sessions.

Host limitation / fallback

I realize the native ChatGPT tool-call shell may be controlled by the host rather than by DevSpace itself. If DevSpace cannot directly collapse native host-rendered tool rows, possible fallbacks might be:

1. Use the MCP Apps / result-card UI to expose only a compact summary by default and place detailed output behind an internal disclosure/expand control.
2. Reduce repeated model-visible/result-card content after a workspace has already been established.
3. Add a configurable compact-output mode such as ui.compactToolCalls / ui.compactResults for ChatGPT-facing sessions.
4. Where technically possible, aggregate sequential low-level operations into a single activity/result card without changing tool semantics or model-visible correctness.

Reproduction

1. Connect DevSpace to ChatGPT Web.
2. Ask ChatGPT to perform a non-trivial task involving many file inspections, command executions, patches, and verification steps.
3. After several turns, dozens of DevSpace tool-call/result blocks accumulate in the page.
4. The conversation becomes much longer than the human-readable assistant/user content itself.

Environment

• Host: ChatGPT Web
• DevSpace: v1.0.8
• Typical workflow: ChatGPT -> DevSpace MCP -> local Windows workspace / Codex

Goal

Keep the full audit/debug information available, but make normal long-running DevSpace conversations substantially easier to read and navigate without changing execution semantics.