harbor run -d frontis/generated-basic-production-tracesBasic runnable replay Tasks mechanically generated from Frontis production Historical Trials.
harbor run -d frontis/generated-basic-production-tracesMVP Harbor Dataset mechanically generated from exported production Historical Trials. It contains 57 basic replay Tasks.
Each Task preserves source session/invocation lineage, selects one task episode by
max-replayable-tool-pairs-then-earliest-user-step.v1, excludes historical reasoning/final content, and exposes the
recorded Tool observations through a deterministic generic MCP replay fixture.
Only standalone, read/retrieval episodes with substantive observations are
eligible; operational setup calls and side-effect/file-delivery episodes are not
exposed as replay evidence.
Generation is mechanical and uses no AI: instruction extraction, episode
selection, classification, fixture construction, validation, and lineage are all
scripted (generation_ai_used=false). AI is reserved for the post-run semantic
Judge and cannot fill Task facts or see the historical final/Gold.
Run the published Dataset with stock Harbor:
harbor run \
-d frontis/generated-basic-production-traces@v0.1.0 \
-a opencode \
-m <provider/model> \
--allow-agent-host <provider-host-or-cidr> \
--ve REWARDKIT_JUDGE=<provider/model>
Configure the selected provider through its standard Harbor model environment or
adapter configuration (for example, the provider API key/base URL). The explicit
allowlist is required because Harbor restricts agent.run() to no-network
after the stock Agent finishes setup in the public environment. It should contain
only the selected model-provider host or CIDR; the replay fixture remains
reachable over loopback. No Task-specific local path, host alias, or custom
Harness patch is required.
This Dataset demonstrates scalable conversion and runnable replay fixtures. It is
not a claim that the production native Tool surface or side effects were exactly
recreated. See generation-report.json and each Task's
tests/source-lineage.json for selection and fidelity evidence.
generation-report.json also serves as the compact category index and rejection
ledger; no duplicate index format is generated.