Replication Package for Agentic Software Engineering Industrial Case Study

This dataset is the anonymized replication package for the industrial case study "Delivering an Enterprise Multi-Tenant SaaS System with Parallel LLM Coding Agents under Frozen Contracts: An Industrial Case Study" (submitted to Empirical Software Engineering). The study documents, with end-to-end traceability, the eight-day delivery of a commercial multi-tenant SaaS system built on an open-source framework baseline, in which one human arbitrated decisions while one architect agent and seven implementation agents — each an independently launched agent-CLI instance, not a subagent within a single session — produced all code and documentation, coordinated solely through frozen API contracts and executable work orders. The package contains anonymized process metadata only. It includes no business source code, secrets, domain names, account identifiers, or any conversation content; committer logins, project/framework names, and the business domain have been deterministically anonymized. Contents (data sources D01–D20, each indexed by SHA-256 in the paper's appendix): - Repository-mining metrics: commits, file-level numstat, daily cadence, 90-minute session clusters, phase milestones, per-component code size, document statistics (D01–D04, D07–D09, D12–D17). - Human-input event-stream telemetry — timestamps, target component, character counts, category; no input text — and its summary (D18). - Human gate/decision candidates and fix/defect classifications (D05, D06). - Merge-request platform events (D11). - Static-quality measures: lizard cyclomatic complexity and jscpd duplication (D19). - Process metrics of a prior, abandoned attempt that motivated the architecture, reported as background only (D20). - Analysis/extraction scripts and two analysis reports. Verification (anyone, no source code required). Run python3 verify.py: it recomputes the SHA-256 of every data file and compares it against the paper's data-source index, and re-derives the paper's headline figures (e.g., 179 commits; 488 human inputs = 407 dialog + 81 confirmations; 36,850 characters; 99 gate candidates = 49/23/27) directly from the data, printing a pass/fail report. Reproducibility note. The extraction scripts regenerate the data from the original private project repository and local session logs, which are not included (proprietary code cannot be shared) — standard for industrial case studies on proprietary systems. Run outside that environment, the scripts exit gracefully and point to verify.py. The shipped data/ is the verified, anonymized final product against which the paper's claims can be checked. Provenance window: 2026-06-03 to 2026-06-10 (+0800). Commit timestamps reflect engineering-activity windows, not human time-on-task. License: CC BY 4.0. Citation: Please cite the associated paper (DOI / arXiv to be added upon publication).