Handle manual form submissions in evals-cli

[swissspidy]

Summary

Fixes #66

When WebMCP forms do not use the toolautosubmit attribute, the browser activates the form and fires a toolactivated DOM event rather than automatically submitting. In a headless evals-cli session, this previously caused the browser evaluation to hang indefinitely.

This PR addresses the issue by handling toolactivated events gracefully, adding tool execution result verification to evals, and reflecting result comparisons in both the console and HTML reports.

Changes

1. Graceful toolactivated Handling in BrowserToolRegistry:

   • Listens for the toolactivated DOM event when executing browser tools.
   • If no auto-submission occurs within a 1-second timeout after toolactivated fires, resolves the tool execution with "pending form submission".
   • Cleans up unused listToolsFromPage function.

2. Tool Execution Result Matching (result in expectedCall):

   • Added an optional result?: unknown; field to FunctionCall and ToolCall types.
   • Updated VercelBackend to capture the tool execution return value on each executed ToolCall.
   • Updated functionCallOutcome in utils.ts to evaluate expected.result against actual.result using matchesArgument (supports exact match, objects, regex $pattern, $contains, etc.).

3. Report and Console Enhancements:

   • Added a "Result" comparison row to the HTML report's Step Details table.
   • Added a "Details" column to the console summary table to clearly display failure reasons (such as result mismatches).

4. Example Evals for French Bistro Demo:

   • Updated evals.json to assert "result": "pending form submission" when testing manual form submission behavior.
   • Added evals-toolautosubmit.json to assert "result": { "$contains": "welcoming you" } when testing auto-submission behavior.

Examples

Expecting manual form submission:

Expected auto form submission:

Expecting manual form submission on a form with toolautosubmit results in an error:

Expecting auto form submission ohn a form without toolautosubmit results in an error: