Add behavioral evals for web tool selection (google_web_search vs web_fetch)

[PewterZz]

Summary

The eval suite has no tests covering how the agent chooses between google_web_search and web_fetch. These are the two web tools available to the agent but they serve different purposes:

• google_web_search — for open-ended queries where the agent needs to find information
• web_fetch — for fetching a specific URL the user has already provided

Without evals, regressions in this distinction go undetected.

Expected behavior

• When asked for current information with no URL, agent should use google_web_search
• When given a specific URL, agent should use web_fetch, not search
• When the answer is in local files, agent should use neither web tool

[agarwal-tanmay-work]

I'm working on the behavioral evaluation coverage for web tools as part of my GSoC 2026 contribution. I have a PR incoming that addresses the gaps identified here.

[agarwal-tanmay-work]

I'm working on the behavioral evaluation coverage for web tools as part of my GSoC 2026 contribution. I have a PR incoming that addresses the gaps identified here.