Session Logs for "Rediscovering Architectural Decision Records: How Persistent Design Context Improves LLM Code Generation"

Raw session logs from a controlled experiment comparing three conditions of LLM-assisted code generation: (1) vanilla Claude Code (no ADR), (2) sqlew (automated ADR from inception), and (3) vanilla-to-sqlew (retroactive ADR adoption mid-project). The experiment involved 12.5 turns of TypeScript workflow approval system development using Claude Code CLI (Opus 4.6) in Docker environments with zero human design intervention (YOLO mode). Contents: session-log-sqlew — JSONL session logs from the sqlew condition (19 turns) session-log-vanilla — JSONL session logs from the vanilla condition (24 turns, incl. supplementary Turn 12.6) session-log-vanilla-to-sqlew — JSONL session logs from the vanilla-to-sqlew condition (4 turns, Turn 11.5–12.5) Each JSONL file contains timestamped records of user prompts, assistant responses (including Chain-of-Thought thinking blocks), and tool usage. Related Resources: Experiment results with full commit history: https://github.com/sqlew-io/sqlew-efficacy-study Replication kit (Docker environments, prompts, analysis methodology): https://github.com/sqlew-io/sqlew-efficacy-replication sqlew tool: https://github.com/sqlew-io/sqlew