Dalia 1.0.0-beta3 Introduces Haiku Retrieval and Expands Developer Workflow Controls

Dalia version 1.0.0-beta3 introduces a redesigned retrieval system and expanded workflow controls for AI-assisted Drupal development. The release, shared by Mateu Aguiló Bosch, replaces the previous embedding-based retrieval model with a system powered by Claude Haiku, aiming to improve both speed and accuracy when retrieving Drupal documentation and project context. Full release details are available in the official changelog.

The shift to Haiku-powered retrieval marks a structural change in how Dalia handles retrieval-augmented generation (RAG). Instead of relying on locally executed embedding models, the system now queries contextual knowledge dynamically, allowing more relevant documentation to be injected into prompts during code generation. This change addresses a known limitation in AI-assisted Drupal development, where mixed and outdated training data can produce inconsistent outputs.

The update introduces granular control over retrieval behaviour. New per-phase RAG controls allow developers to determine when contextual retrieval is applied, particularly during planning and execution stages. This enables more precise management of how documentation influences generated output, aligning with Dalia’s broader design principle of keeping the developer in control of critical decisions throughout the pipeline.

Error handling has been significantly expanded. The system now surfaces authentication failures and token limits through explicit modal dialogs, replacing earlier silent failures. These changes improve transparency during execution, ensuring that developers can identify and respond to issues without relying on logs or indirect signals.

Several features focus on improving reliability and reducing redundant work. Mandatory module discovery introduces automated checks against Drupal.org to identify existing contributed modules before implementation begins, preventing duplication. Verification checklists are now generated alongside outputs, separating functional requirements from implementation guidance to support structured validation. Code duplication detection further enhances output quality by flagging repeated patterns during generation.

The release also expands control within the specification workflow. Developers can now stop and resume the agent during specification conversations, not just execution phases. This allows iterative refinement of requirements before committing to implementation. Supporting interface updates, including activity indicators and planning notifications, provide clearer visibility into system state and progress.

Additional improvements affect usability and interface consistency. Dashboard updates now include feature branch names and timestamps, while a consistent stop control is maintained across views. The system also introduces environment checks such as Playwright browser validation, ensuring that testing dependencies are correctly installed for browser-based validation workflows.

These changes build on Dalia’s structured development model, which operates across multiple phases including specification, planning, task breakdown, execution, validation and review. The platform integrates components such as the AI Task Manager and Self-Review to orchestrate tasks and enable local code review workflows.

The release also includes multiple bug fixes addressing retrieval failures, session inconsistencies, interface rendering issues and parsing errors. These updates contribute to stabilising the platform as it approaches a production-ready state.

Dalia remains in beta and is currently distributed as a development tool built around existing Claude Code subscriptions. The project is developed by Mateu Aguiló Bosch and sponsored by Lullabot, reflecting ongoing investment in AI-assisted Drupal development workflows. Its architecture continues to emphasise controlled AI-assisted development, where automation supports but does not replace developer decision-making as the project moves toward a stable 1.0 release.