AI coding assistants now perform multi-step workflows (file edits, shell commands, web fetches, commits) rather than only returning short snippets. This dataset captures that end-to-end interaction loop—conversation turns, explicit tool calls and results, agent “thinking” traces, git diffs and commit metadata—so researchers can study how agents behave, when they author code vs. humans, and how tool use affects outcomes at scale.
What Sets It Apart
- End-to-end session structure: data is split into named Parquet tables (conversations, sessions, commits, checkpoints, repositories) enabling joins from high-level session metadata down to per-turn content and file diffs. This makes longitudinal analyses (e.g., agent-authored commit trajectories) straightforward.
- Rich agent traces and attributions: includes assistant thinking blocks, tool_use/tool_result entries, model names (e.g., Claude Code, Gemini CLI, Codex), and an agent_percentage metric that quantifies the share of committed lines authored by the agent—useful for attribution and impact studies.
- Large, coding-focused scale: repository & session-level coverage (e.g., thousands of sessions and millions of conversation turns) plus 14k+ commits and detailed checkpoints, permitting both micro (turn-level) and macro (repo-level) analyses.
- Practical access & provenance: dataset links an arXiv paper (arXiv:2604.20779) and a project website, provides data collection and PII-redaction notes, and is distributed under an ODC-BY license with an access agreement required on Hugging Face.
Who it's for (and tradeoffs)
Great fit if you want empirical evidence about agent-assisted coding workflows, want to measure agent authorship/attribution, or need realistic tool-call traces and commit histories for model evaluation or human–AI collaboration research. Look elsewhere if you need runnable, fully deterministic testcases for code correctness (this dataset records real edits and metadata but not exhaustive function-level ground truth), or if you cannot accept datasets that require agreeing to access conditions and may include sensitive real-world content despite redaction steps. Note: analysis often requires joining multiple tables and handling large Parquet files—best used with data tools that scale (pandas, polars, or dedicated query engines).
