Drupal AI Learners Club Examines the Learning Gap in Agent-Written Code

During a Drupal AI Learners Club session on 17 July 2026, Amber Matz and Angie Byron hosted a demonstration of Decodie by Owen Bush. Bush, a senior developer at Lullabot, described the open-source project as a way to create structured learning entries during AI-assisted coding sessions. The entries record what was built, the concepts involved, the approach selected, and alternatives considered.

Bush framed the project as a response to the gap between accepting working code and understanding why an agent chose a particular implementation. He said explanations about design patterns, architectural choices, and alternatives considered can remain in an agent conversation and become difficult to revisit or share. Decodie stores that reasoning locally so developers and teams can review it after the coding session ends.

The tool writes structured JSON files to a .decodie directory, where entries can be committed with a repository or kept local. Its seven skills cover observing decisions during active work, analysing existing code, preparing project overviews, explaining selected code, answering questions, verifying references, and flagging stale entries. Bush said the observe skill records meaningful decisions during the same agent session, while the analysis tools can document an existing file, directory, or project.

The wider toolset includes a Visual Studio Code extension, a local web interface, a DDEV add-on, a GitHub Action, and a GitHub bot. The editor extension displays entries beside relevant code, while the web interface groups them into lessons and tracks learning progress. The GitHub Action can generate entries from pull-request changes, although Bush said it may miss some rationale available during the original coding session. The bot can answer questions about code during review.

Decodie refers to functions by signature rather than line number so entries can remain connected when code moves. Bush cautioned that larger changes can still break those references, and the verification tools can mark affected entries as stale. During the question-and-answer segment, he acknowledged that inaccurate or hallucinated explanations remain possible and said the current tools do not provide a refinement step or direct entry editing. He also described multi-repository aggregation as a proof of concept rather than an available service. For Drupal teams using coding agents, the project offers a method for retaining implementation reasoning, but the session did not establish that generated explanations can replace mentoring, documentation, or code review.