A coding agent’s useful behavior is often visible not in its final answer, but in the sequence of commands, tool outputs, retries, and recovery steps that led there. This collection makes those interactions inspectable, while its single-developer, codec-focused origin is also a reminder that realistic traces are not automatically representative ones.
What Sets It Apart
- Three aligned views—conversation trajectories, parent-to-subagent delegation pairs, and Codex-formatted traces—support different analysis and browsing workflows.
- The 437 runs include tool calls and their outputs, including failures and recoveries; this offers a more realistic picture of agent work than clean, success-only demonstrations.
- The 314 delegation examples expose both the task brief and subagent report, making them useful for studying how work is split and summarized.
- The sessions are mostly Windows-based C#/C++ development and video-codec reverse engineering, so the data provides depth in a narrow setting rather than broad task coverage.
Who It’s For—and the Trade-Off
A good fit for researchers examining tool-use sequences, agent delegation, or coding-agent trace formats, and for regression tests based on real sessions. Look elsewhere for a broad coding benchmark or a clean set of verified demonstrations: traces are not curated for correctness, context and some outputs are missing or truncated, and the session-based test split shares parent-thread goals with training data. Treat embedded commands and third-party tool output as untrusted, and check reuse rights for material inside the logs.
