Round 1: The core

Build the agents catalogue, the link builder and the CLI. Plain ES modules, Node.js 20 or later, **no
dependency**. `ARCHITECTURE.md` in the repository is the reference: the names, formats and exit codes
below are fixed; everything else is yours.

## The agents

| id | name | kind | value |
| --- | --- | --- | --- |
| `claude` | `Claude` | `url` | `https://claude.ai/new?q=` + encoded prompt |
| `chatgpt` | `ChatGPT` | `url` | `https://chatgpt.com/?q=` + encoded prompt |
| `perplexity` | `Perplexity` | `url` | `https://www.perplexity.ai/search?q=` + encoded prompt |
| `le-chat` | `Le Chat` | `url` | `https://chat.mistral.ai/chat?q=` + encoded prompt |
| `cursor` | `Cursor` | `deeplink` | `cursor://anysphere.cursor-deeplink/prompt?text=` + encoded prompt |
| `claude-code` | `Claude Code` | `command` | `claude ` + quoted prompt |
| `codex` | `Codex CLI` | `command` | `codex ` + quoted prompt |

- Keep this order: it is the order of `list` and, from round 2, of the picker.
- **Encoded** is exactly `encodeURIComponent(prompt)`.
- **Quoted** is the prompt between single quotes for a POSIX shell, each `'` written `'\''`
  (`it's` gives `'it'\''s'`). Newlines, `$`, backticks, `\` and `"` stay as they are.
- A `url` or `deeplink` value is at most **8,000 characters**, measured on the final value. A `command`
  has no limit.

## The core: `src/harnesses.mjs`

- `agents`: the catalogue, an array of `{ id, name, kind }` in the order above (more fields allowed).
- `MAX_LINK_LENGTH`: `8000`.
- `linkFor(id, prompt)` returns `{ agent, kind, value }`, or throws an `Error` whose `code` is
  `"UNKNOWN_AGENT"`, `"EMPTY_PROMPT"` or `"TOO_LONG"`. Round 2's button imports it: keep it free of
  Node built-ins.

## The CLI: `node bin/connect-agent.mjs`

| Command | Prints on stdout | Exit |
| --- | --- | --- |
| `list` | one line per agent, `<id>\t<name>\t<kind>` (tab-separated), in catalogue order | 0 |
| `list --json` | a JSON array of `{ "id", "name", "kind" }`, in catalogue order | 0 |
| `link <id> <prompt>` | one line of JSON, exactly the keys `agent`, `kind`, `value` | 0 |
| `link <id> --file <path>` | the same, the prompt being the file's content read as UTF-8 | 0 |

- The prompt is used **exactly as given**: never trimmed or changed (a file's final newline included).
- **Exit 2**, usage errors: no command or an unknown one; an unknown agent (ids are case-sensitive); no
  prompt; an empty or whitespace-only prompt (from the argument or the file); `--file` with no path; both
  a prompt and `--file`; a file that cannot be read; any extra argument.
- **Exit 3**: the `url` or `deeplink` value would be longer than 8,000 characters.
- On an error, print nothing on stdout and one line on stderr starting with `error: `, saying what was
  wrong and how to fix it.

Add your own tests under `test/` (`npm test` runs `node --test`), and document the CLI and how to add
an agent in the README.

## How to run it

`node bin/connect-agent.mjs list`, then `node bin/connect-agent.mjs link claude "Explain this repo"` and
open the `value` in a browser; `node bin/connect-agent.mjs link claude-code "it's done"` and paste the
`value` in a shell (`echo` instead of `claude` to see what it receives).

- C1 (correctness): try every agent with a prompt holding `é`, an emoji, a newline, `&`, `#`, `'` and
  `"`; open two of the links and paste one command into `bash -c 'printf %s …'`.
- C2 (robustness): run `link` with no prompt, `"   "`, an unknown agent, a missing `--file`, an extra
  argument and a 9,000-character prompt. Read each message and exit code.
- C3 (code quality): read `src/harnesses.mjs` and `bin/connect-agent.mjs`; count what adding an agent
  takes.
- C4 (docs): follow the README from a fresh clone.