Open Traces

Public traces are one supported input family in workledger, not the whole repository story.

They matter here because they are easy-to-share test cases for the normalize-and-rollup pipeline:

• agent messages
• trajectories
• span trees
• token and cost totals

The implemented contribution is still the same:

• normalize source records into ObservationSpan
• roll them into WorkUnit
• optionally classify and report on those work units

Supported Public Dataset Shapes

dataset	url	shape	support	demo
smolagents/gaia-traces	https://huggingface.co/datasets/smolagents/gaia-traces	messages / trajectories	supported	yes
kshitijthakkar/smoltrace-traces-20260130_053009	https://huggingface.co/datasets/kshitijthakkar/smoltrace-traces-20260130_053009	trace + spans	supported	yes
smolagents/codeagent-traces	https://huggingface.co/datasets/smolagents/codeagent-traces	messages / outcomes	planned	no

What The Current Adapters Actually Do

• gaia maps a message-style row into one root span plus per-message spans
• smoltrace maps trace-and-span rows into span-preserving ObservationSpan
• both preserve Hugging Face lineage in raw_payload_ref
• both populate dataset metadata in facets["hf"]

What workledger Adds After Ingestion

• ObservationSpan as the normalized execution record
• WorkUnit as the rolled unit of work people can inspect
• evidence bundles and lineage refs
• direct and allocated cost attached to rolled work
• optional review, policy, and economics layers after rollup

How workledger Fits

workledger is not a trace viewer or tracing backend. In this repo, public traces are just one way to feed the same local pipeline.

Adding More Adapters

New public trace formats should:

• map into ObservationSpan
• preserve stable source refs in raw_payload_ref
• store source-specific metadata in namespaced facets
• include a small fixture and a demo-sized path