Publish a schema for --json output and MCP tool results (outputSchema)

[cpeaustriajc]

What I'd like

A published schema for what agent-browser returns, the way agent-browser.schema.json already covers the config file:

1. A JSON Schema for the --json envelope (success, data, error, and error fields such as code) and for each command's data, versioned with each release.
2. An outputSchema on each MCP tool, describing the structuredContent it already returns.

Why

We wrap agent-browser in a test-runner engine. Every reply arrives as untyped JSON, so today each consumer reverse-engineers the shapes and writes its own parser. When a shape changes, nothing tells us. It happened in #785 (the refs dict left snapshot --json in the Rust rewrite) and #1024 (snapshot -i changed shape on some pages). A published schema would let a consumer validate replies at runtime and fail loudly, naming the field, on the first upgrade that changes a shape, instead of failing later somewhere unrelated.

What I checked (0.41.1)

• The docs describe the --json output piecemeal: data.result and data.unserializable for eval, data.webmcp, targetId in tab list --json, code: "tab_gone" on errors. There is no single page that describes the envelope or each command's data.
• The npm package ships the native binary and a launcher only: no main, types or .d.ts.
• agent-browser mcp --tools all lists 64 tools. Every one has a strict inputSchema, and none has an outputSchema. Yet agent-browser mcp call agent_browser_tools_profiles '{}' returns structuredContent, and the commands doc refers to structuredContent.response.data.unserializable. So the structured shape exists but isn't declared.
• 能否提供python或者api调用支持 #85 confirms that the CLI's --json and MCP over stdio are the supported programmatic interfaces, so this would cover both.

Happy to help with a draft if a format is preferred (a JSON Schema per command, or one generated from the Rust types).