Agentic Code Dataset 22: 22 Real Claude Code Sessions

December 2025 collection of 22 complete Claude Code session recordings. The dataset contains 17,487 conversation turns, 50 hours of runtime, and over 13 million total tokens. It was created by the author 'novita' to validate Suffix Decoding's applicability in Agentic Coding scenarios.

Use Cases

• Validate suffix decoding techniques based on recorded agentic coding sessions.
• Analyze conversational patterns in AI-assisted coding based on 17,487 turns.
• Benchmark LLM performance in extended coding tasks based on 50 hours of runtime.
• Study token usage and efficiency in agentic workflows based on over 13 million tokens.

Strengths

• Contains 22 complete, real-world coding session recordings.
• Provides detailed session-scale statistics, including min, max, and average conversation turns.
• Includes precise token counts (6,996,619 input, 6,094,906 output) and runtime metrics.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and specific file formats are unknown, which may limit suitability assessment.

Provenance

Source

huggingface, author 'novita'.

Collection Method

Collected from 22 complete Claude Code session recordings.

Time Range

Collected in December 2025.

Freshness

Last updated 2025-12-31 08:54:20.

Geography

null