Add MCP annotations (readOnlyHint/destructiveHint) to tool definitions

Summary

The Stagehand tool definitions don't include MCP annotations (readOnlyHint, destructiveHint, idempotentHint). These are optional hints in the MCP spec that help LLMs decide how to handle tools safely.

Suggested annotations

Tool	Suggested annotation	Reason
browserbase_stagehand_observe	readOnlyHint: true	Inspects page, no side effects
browserbase_stagehand_screenshot	readOnlyHint: true	Captures state, no side effects
browserbase_stagehand_extract	readOnlyHint: true	Extracts data, doesn't modify page
browserbase_stagehand_navigate	idempotentHint: true	Same URL = same result
browserbase_stagehand_act	destructiveHint: false	Interacts with page but typically not data-destructive

Example

In src/tools/observe.ts:

const observeSchema: ToolSchema<typeof ObserveInputSchema> = {
  name: "browserbase_stagehand_observe",
  description: "Observe the current page and return a list of actionable elements.",
  annotations: {
    readOnlyHint: true,
  },
  inputSchema: ObserveInputSchema,
};

Context

I ran a benchmark of tool definition quality across popular MCP servers. This server scores 94/100 (A grade) — the basics are solid. The main gap is missing annotations. Servers like Exa and Kubernetes MCP use annotations extensively.

Full benchmark: https://wildrunai.com/blog/mcp-quality-benchmark
Validation tool: https://wildrunai.com/tools/mcp-validator