dsh-inbox-paste

Readme

# dsh-inbox-paste

[中文文档](README.zh.md)

DSH **web bundle**: paste or pick files in the composer, store under `{session.cwd}/inbox/`, show thumbnails before/after send. Agents use `list_inbox` / `read_inbox` for **text**; **image understanding requires your own vision skill or model** (see below).

![Example: pasted screenshot analyzed via image-reader skill](docs/images/example-chat.png)

## Requirements

- [DeepSeek Harness](https://github.com/deepseek-ai/DeepSeek-Harness) with the **web** profile
- Node.js ≥ 22, `dsh` CLI

## Install

```sh
dsh plugin --profile web add github:JarszmLovesAI/dsh_upload_file
dsh web
```

Remove: `dsh plugin --profile web remove dsh-inbox-paste`

## Usage

1. **Paste** (Ctrl+V / Cmd+V) or use the **paperclip** in the composer.
2. Pending thumbnails appear above the input; remove with × before send (no disk write yet).
3. On send, files land in `inbox/`; the bubble can show thumbnails (see config).
4. Agent: `list_inbox` → `read_inbox` for text; for **images**, use the bundled [image-reader](skills/image-reader/README.md) skill or your own vision tooling.

### Upload ≠ multimodal understanding

This plugin does **not** use Harness native image attachments. The main model only gets path markers unless you add vision:

- Install the optional **image-reader** skill and set `ARK_API_KEY` in `~/.dsh/.env`, or
- Bring your own vision MCP / skill / multimodal model.

## Configuration

`~/.dsh/profiles/web/cordis.patch.yml` (adjust profile name):

```yaml
- id: inbox-paste
  config:
    uploadMaxBytes: 26214400
    inboxDirName: inbox
    allowedExtensions: []
    customizeUserBubble: true   # false = built-in user bubble (may show raw paths)
```

Restart `dsh web` after changes.

## vs `dsh-file-upload`

| | dsh-inbox-paste | dsh-file-upload |
|--|-----------------|-----------------|
| Paste | ✅ | ❌ |
| Storage | `{cwd}/inbox/` | `{cwd}/.dsh-uploads/` |
| Built-in PDF | ❌ | ✅ |

## Tests

```sh
npm test
```

## License

MIT