Skip to content

Real-task acceptance benchmark: native tools vs this MCP server #67

Description

@cf-pages

Targeted: 0.5.0. Baseline: release/0.3.0.

Current evidence (dogfood, Aider harness) mostly submits pre-baked patches — it proves the mechanical pipeline, not editing capability; the SWE-bench harness is still preflight-only. Nothing today can falsify "the model spends rounds failing to land an edit".

  • 30–50 real tasks, same model + same prompt, native tool loop vs this MCP server.
  • Metrics: final test pass rate, first-patch success rate, rounds to green, regressions introduced, wall time.
  • Output lands in reports/; reuse the benchmarks/real_workloads.py skeleton where possible.
  • This is the data source for the 0.5.0 acceptance thresholds (first-patch success > 80%, overall failure rate < 3%) and for the structured-only default decision.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions