TraceLab Coding Agent Workload Traces

TraceLab is a sanitized, real-world workload trace of Claude Code and Codex. Each JSONL row represents one LLM invocation and retains the serving-relevant structure needed to study sessions, token accounting, prompt-cache behavior, tool calls, and timing.

• 665,453 agent steps
• 8,058 sessions
• 743,819 normalized tool calls
• 52 pseudonymous users
• Providers: Claude Code and Codex
• Release: v0.0.2

Project links:

• Paper: TraceLab: Characterizing Coding Agent Workloads for LLM Serving
• GitHub repository
• Interactive website
• Pinned GitHub release v0.0.2

Load the dataset

from datasets import load_dataset

dataset = load_dataset("UW-SyFI/TraceLab", split="train")
print(dataset.num_rows)
print(dataset.features)

The default config is the rounds table. Load the normalized child tables separately and join them to rounds with round_pk:

tool_calls = load_dataset("UW-SyFI/TraceLab", "tool_calls", split="train")
timing_events = load_dataset("UW-SyFI/TraceLab", "timing_events", split="train")

Pin the published snapshot for reproducible studies:

dataset = load_dataset("UW-SyFI/TraceLab", revision="v0.0.2", split="train")

The repository also includes the query-ready DuckDB database as an auxiliary artifact:

from huggingface_hub import hf_hub_download
import duckdb

database_path = hf_hub_download(
    repo_id="UW-SyFI/TraceLab",
    repo_type="dataset",
    filename="artifacts/v0.0.2/syfi_coding_trace.duckdb",
    revision="v0.0.2",
)
connection = duckdb.connect(database_path, read_only=True)
print(connection.execute("SELECT provider, count(*) FROM rounds GROUP BY provider").fetchall())

Data format

The Viewer-friendly Parquet representation contains three relational tables with stable schemas:

• default / rounds: 665,453 LLM invocations with pseudonymous provider/session/round identifiers, model, token counts, cache-prefix accounting, source store, and trace_key.
• tool_calls: 743,819 normalized tool calls keyed to rounds.round_pk.
• timing_events: 2,688,829 ordered events keyed to rounds.round_pk.

Important token fields include:

• input_tokens_total: total input tokens reported for the invocation.
• prefix_tokens: provider-reported cached input tokens.
• newly_append_tokens: non-cached input tokens appended for the invocation.
• output_tokens: generated output tokens.

tool_calls contains privacy-safe structural metadata such as tool name, pseudonymous tool-call id, timestamps, input/result character counts, latency, error status, executable labels, and command skeletons. timing_events.event_index preserves event order within each round. Full tool inputs and outputs are not published.

The original sanitized nested JSONL and the query-ready DuckDB remain available as auxiliary files. They are byte-identical to the GitHub v0.0.2 release assets. The Parquet tables are deterministic relational exports of that DuckDB for Hugging Face Data Studio and load_dataset().

Sanitization and responsible use

The public trace is generated by TraceLab's sanitization pipeline. User, project, session, round, turn, tool-call, and process-session identifiers are replaced with stable pseudonyms. Local paths, working directories, source session paths, tool inputs, prompts, responses, and full tool outputs are removed. Custom executable and integration names are replaced with stable generic labels unless they belong to an explicitly reviewed public allowlist.

This is a trace of real developer activity. Do not attempt to re-identify contributors or infer their projects from the remaining workload structure.

Known limitations

• Consecutive round_index values are consecutive normalized records, but should not always be interpreted as a single reconstructable linear conversation lineage.
• A small fraction of adjacent transitions report prefix_tokens larger than the immediately reconstructable previous context. For synthetic session-local KV-cache replay, conservatively cap the next reusable prefix at previous.input_tokens_total + previous.output_tokens. See TraceLab issue #22.
• Provider tokenization, message framing, cache boundaries, retries, resumes, branches, compaction, and historical extractor behavior can affect adjacent-round accounting.
• The sanitized analytical rows intentionally do not contain the message text required for a conversational Agent Trace Viewer.

Files and integrity

File	SHA-256
data/v0.0.2/rounds/train.parquet	cb8c2b1d23307090c8057c747f955663bacde16a49f84a3c5071a1e2e96a7ab8
data/v0.0.2/tool_calls/train.parquet	7186e0d86d3d870e5b374cc29678a4754fcb789ebee4f5ad86461dd02db6f68b
data/v0.0.2/timing_events/train.parquet	a207499989d4f004a6807698c33ad6da57cd30fd45404188bf8593fc4d3b9296
data/v0.0.2/syfi_coding_trace.jsonl.gz	11ce51ec0a25e3d1d95b025bca2f7d1647e47571eb7cc968acd5fc64d4b4fb65
artifacts/v0.0.2/syfi_coding_trace.duckdb	a7bab286bc640844560850965ccf47975cf66407154132abaab90f27ec9be744

The JSONL and DuckDB files are byte-identical to the assets published in the TraceLab v0.0.2 GitHub release.

License

The released dataset is licensed under Creative Commons Attribution 4.0 International. Please credit TraceLab (SyFI Lab, University of Washington) and link to https://tracelab.cs.washington.edu/. TraceLab source code is separately licensed under Apache-2.0 in the GitHub repository.

Citation

@misc{zhu2026tracelabcharacterizingcodingagent,
  title         = {TraceLab: Characterizing Coding Agent Workloads for LLM Serving},
  author        = {Kan Zhu and Mathew Jacob and Chenxi Ma and Yi Pan and Stephanie Wang and Arvind Krishnamurthy and Baris Kasikci},
  year          = {2026},
  eprint        = {2606.30560},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  url           = {https://arxiv.org/abs/2606.30560}
}