Non-coder first-day user feedback: plugin crash loop, retry count, port lockup, MCP UX

[Lescure1004]

Background

Non-programmer using DSH since day one of the Developer Preview (I build my
own project with it). Love the direction. Real issues below, most → least severe.

1. Plugins get auto-triggered by the model and crash-loop the UI (severe)

Installing the community plugin dsh-antv-infographic made the model use
infographic fences on its own. When the fence format was wrong ("first line
must be infographic <template-name>"), the render error fed back into the
conversation → model retries → error again → loop until the browser crashed.
Rendering also failed outright on Safari; Edge barely worked. Uninstalling
fixed everything.

Suggest: report a fence render error once — don't feed it back to the model
repeatedly; and publish an official browser support matrix.

2. Only 2 retries before giving up (UX)

On network hiccups, a request retries exactly twice, then stops with an error.
2 feels too few (sometimes 4–5 attempts would recover). I later learned the
count is configurable in a settings file — but no ordinary user would know.
Suggest: raise the default, or surface where to configure it in the UI.

3. Killing the terminal leaves the app unable to restart (UX)

I close the terminal window instead of pressing Ctrl+C. Next launch then fails
(port 3080 still held by the leftover process) and looks like the app is
broken. Suggest: detect the occupied port at startup and show "an instance is
already running" instead of a raw error; document the proper exit method.

4. Feature request: a user-facing way to add MCP servers

I expected to reuse Claude's existing MCP servers in DSH. The framework
supports MCP natively (@deepseek-ai/dsh-mcp-client, stdio + Streamable
HTTP), but the Web build doesn't load it by default, the GUI has no config
entry, and the user docs never mention it. To a non-developer it's unusable.
Suggest: an easy entry point like Claude's (claude mcp add-style command or
a Settings page section).

Closing

Great product and lively plugin ecosystem overall — just sharing real usage
feedback to help the preview improve.

[xiaoyuyu6420]

For issue #1 (the crash loop): break the loop first, then decide whether to keep the plugin.

1. Stop the loop — remove or disable dsh-antv-infographic (e.g. dsh plugin remove dsh-antv-infographic in your profile, or toggle it off in Settings → Plugins). The model auto-triggering the fence stops immediately, and the UI unblocks. No conversation data is touched by removing a plugin — your sessions stay intact.

2. If you want to keep it: the loop only starts when the fence format is wrong, so make the model emit the fence exactly as the plugin expects — first line must be infographic <template-name>. When a render error does happen, telling the model "stop using infographic fences" breaks the retry cycle much faster than letting it guess.

3. Worth reporting to the plugin author (or in a plugin feedback thread): a community plugin whose fences the model starts using on its own is an aggressive prompt-injection profile — plugins shouldn't change model behavior unless explicitly requested.

The rest of your list (2–5) reads like solid upstream feedback, especially the crash-loop one — hope the maintainers pick it up.