Claw-Eval Rollout
August 16, 2026 ยท View on GitHub
Canonical local data layout:
benchmarks/claw-eval/dataset/tasksbenchmarks/claw-eval/dataset/general
The adapter should use the local mirrored dataset above as its runtime source of truth. Do not introduce a second dataset/upstream/ task tree.
Phases:
- scaffold
- task.yaml loader
- text-only pilot tasks
- service-backed pilot tasks
- broader split/category rollout
This adapter should remain schema-separated from benchmarks/pinchbench/.
Mock Service Split
The plugin split changes tool exposure, not backend semantics.
- Original behavior still lives in upstream FastAPI services under
claw-eval/mock_services/. claw-eval-mock-toolsis the aggregate OpenClaw wrapper over those HTTP endpoints.- Per-service plugins such as
claw-eval-mock-tools-notes,-helpdesk, and-crmare thin shims that only re-export tools matching a name prefix.
Implication:
- If service base URLs are correct, the split plugins should behave the same as the original mock services.
- The new risks are adapter-layer risks:
- wrong tool-name to plugin mapping
- wrong plugin install/enable plan
- wrong service URL wiring
That is why the adapter rollout starts with plugin-closure resolution before any runtime mutation.