Skip to content

Benchmark Kernel MCP with ClawBench through Harbor - #162

Open
rgarcia wants to merge 22 commits into
mainfrom
hypeship/harbor-mcp-benchmark
Open

Benchmark Kernel MCP with ClawBench through Harbor#162
rgarcia wants to merge 22 commits into
mainfrom
hypeship/harbor-mcp-benchmark

Conversation

@rgarcia

@rgarciargarcia commented Aug 21, 2026

Copy link
Copy Markdown
Contributor

Summary

  • add KERNEL_MCP_ENABLED_TOOLSETS so self-hosted deployments can positively select tool families; the benchmark exposes only get_connection_context and execute_playwright_code
  • steer execute_playwright_code to return a relevant accessibility snapshot or compact page state after every action
  • build kernel-mcp-server from the current Git SHA, run it locally inside the Harbor/Hypeman task, and record that source identity in the verifier manifest
  • adapt either one ClawBench task or the complete suite, with configurable Harbor concurrency, while reusing ClawBench's exact pre-created stealth Kernel browser and supplied PurelyMail credentials
  • verify ATIF call/observation pairing, project scope, exact session reuse, the DOM-only tool allowlist, no managed-auth/lifecycle/Playwright-MCP/direct-HTTP use, ClawBench interception and agent stop timing, replay finalization, and browser deletion

Validation

  • bun test — 249 passed
  • python3 -m unittest benchmarks/harbor/clawbench/test_control.py — 6 passed
  • production image build, TypeScript check, Python compilation, and shell syntax checks pass
  • GPT-5.6 Luna full 129-task run: updated description 24/129 reward, original description 19/129, Playwright MCP 15/129
  • updated Kernel MCP used 49% fewer browser calls and at least 45% less agent-model spend than Playwright MCP; all 387 browsers across the three-arm run retained replays and were deletion-verified

@vercel

vercelBot commented Aug 21, 2026

Copy link
Copy Markdown

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

ProjectDeploymentActionsUpdated (UTC)
mcpReadyReadyPreviewAug 24, 2026 12:01am

Comment threadbenchmarks/harbor/run-smoke.sh Outdated
Comment threadbenchmarks/harbor/clawbench/prepare-control.py
Comment threadbenchmarks/harbor/clawbench/prepare-control.py Outdated
@rgarciargarcia changed the title Add Harbor MCP benchmark harnessAdd Harbor and ClawBench MCP benchmarksAug 22, 2026
Comment threadbenchmarks/harbor/clawbench/verify-control.py
The smoke task duplicated what the ClawBench arm already proves. Drop its
task definition, runner, verifier, fixtures, and MCP config, and drop
stale ignore entries nothing writes. Document only the ClawBench flow.
@rgarciargarcia changed the title Add Harbor and ClawBench MCP benchmarksRun ClawBench through Harbor with a source-pinned local Kernel MCP serverAug 23, 2026
@rgarciargarcia changed the title Run ClawBench through Harbor with a source-pinned local Kernel MCP serverBenchmark Kernel MCP with ClawBench through HarborAug 23, 2026

@cursorcursorBot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

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

Fix All in Cursor

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

Want higher recall? High effort reviews run extra passes and find more bugs. A team admin can switch effort levels in the Cursor dashboard.

Reviewed by Cursor Bugbot for commit 9e91f52. Configure here.

Comment threadbenchmarks/harbor/clawbench/run-control.sh
Sign up for freeto join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant

@rgarcia