SyFI TraceLab v0.0.2

Highlights

• Expanded the sanitized public corpus to 665,453 agent steps across 8,058 sessions and 52 pseudonymous users, including 743,819 normalized tool calls.
• Added the Bash command breakdown: executable popularity for Claude and Codex, raw and lifecycle-aware latency rankings, command-structure skeletons, and privacy-safe executable labels.
• Improved Codex command timing by treating write_stdin continuations as part of the originating exec_command lifecycle.
• Preserved historical rounds during multi-host recollection while filtering empty replay artifacts.
• Refreshed session, generation, tool-call, prefix-cache, and cost analyses, including English and Chinese web narratives.
• Privacy hardening: custom and integration-specific tool names are pseudonymized as stable custom_tool_N labels.

Data assets

• syfi_coding_trace.jsonl.gz — compressed sanitized JSONL
• syfi_coding_trace.duckdb — query-ready DuckDB database

SHA-256

• JSONL.GZ: 11ce51ec0a25e3d1d95b025bca2f7d1647e47571eb7cc968acd5fc64d4b4fb65
• DuckDB: a7bab286bc640844560850965ccf47975cf66407154132abaab90f27ec9be744