Design and prototype Tier 2 agent-driven end-to-end tests

[ivklgn]

Problem

In-process MCP integration tests (Layer A) exist, but there is no coverage that drives real agents (e.g. Claude Code, Cursor) against the MCP server. That layer needs a design pass on scope and CI cost before implementation.

Expected result

A designed, opt-in agent-driven test harness plus one end-to-end scenario implemented as a proof of concept.

Impact

Adds confidence that the server behaves correctly under real agent usage, catching integration regressions the in-process tests can't.

Scope

• Design the harness and decide which agents/scenarios are in scope.
• Decide the CI strategy (opt-in / nightly vs. per-PR) given external dependencies, flakiness, and cost.
• Implement one scenario as a proof of concept.

Files: a new test harness under internal/mcp/integration/ or a new e2e/ directory

Acceptance: at least one agent-driven scenario runs behind a documented, opt-in harness.

Status: needs a design pass on scope + CI strategy.