Add Harbor and ClawBench MCP benchmarks

[rgarcia]

Summary

• add a source-pinned Harbor smoke benchmark for stock Claude Code and Codex agents in Hypeman
• verify native MCP calls through paired ATIF observations without a telemetry proxy
• add an opt-in MCP toolset allowlist for controlled benchmark deployments
• add a ClawBench Kernel MCP arm that reuses one pre-created, instrumented stealth browser and disables lifecycle and managed-auth toolsets
• verify clean stop propagation while accepting the expected interrupted terminal tool observation
• pin Harbor 0.21.0 with harbor-hypeman 0.1.1 for reproducible runs

Validation

• bun test (249 passed)
• production image build
• Harbor smoke verifier tests (4 passed)
• ClawBench control-arm tests (6 passed)
• Kernel MCP arm on stealth Kernel browsers: 3/3 intercepted=1, judge_match=1, reward=1, infra_ok=1
• median agent-start-to-interception: 74.0s (51.7–89.9s)
• median first-browser-tool-to-interception: 59.6s (38.3–65.4s)
• median stop propagation: 56ms (53–100ms)
• median Kernel MCP calls: 9 (6–10)
• exact session reuse, project scope, source identity, replay retention, and browser deletion verified in every attempt

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
mcp	Ready	Preview	Aug 22, 2026 9:50pm

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 7475b8a. Configure here.}