Research paper 14
A Longitudinal Corpus of Human-Codex Software Work
This paper describes a longitudinal corpus of Codex-assisted software work drawn from local JSONL session logs. The current local parse on 2026-07-21 enumerates 7,514 session files under active and archived Codex roots. Of those, 7,183 sessions contain usable token counters and 331 do not. The first observed timestamp is 2026-02-28T02:45:36.711Z and the latest observed timestamp in this pass is 2026-07-21T07:19:35.322Z. Summing the last cumulative token counter per token-instrumented session yields 20,786,706,498 total tokens, including 20,704,900,411 input tokens, 19,684,191,488 cached input tokens, 81,046,087 output tokens, and 28,573,367 reasoning output tokens. The same parse counts 22,923 user messages, 79,110 assistant messages, 245,480 function calls, and 238,745 shell command calls. The contribution is a data-descriptor method for studying high-volume human-agent software work without mistaking private log volume for public validity, cost, productivity, or generalization.
- Paper
- 14
- Authors
- A.G. Mauro and C.A. Harris
- Date
- 2026-07-21
- Collection
- Standing Framework Research
Abstract
This paper describes a longitudinal corpus of Codex-assisted software work drawn from local JSONL session logs. The current local parse on 2026-07-21 enumerates 7,514 session files under active and archived Codex roots. Of those, 7,183 sessions contain usable token counters and 331 do not. The first observed timestamp is 2026-02-28T02:45:36.711Z and the latest observed timestamp in this pass is 2026-07-21T07:19:35.322Z. Summing the last cumulative token counter per token-instrumented session yields 20,786,706,498 total tokens, including 20,704,900,411 input tokens, 19,684,191,488 cached input tokens, 81,046,087 output tokens, and 28,573,367 reasoning output tokens. The same parse counts 22,923 user messages, 79,110 assistant messages, 245,480 function calls, and 238,745 shell command calls. The contribution is a data-descriptor method for studying high-volume human-agent software work without mistaking private log volume for public validity, cost, productivity, or generalization.
← Back to research papers