Google Conductor: Persistent Context for AI-Assisted Development in Gemini CLI

Google has released Conductor, a new preview extension for Gemini CLI that introduces a structured, context-driven approach to AI-assisted software development. The extension addresses a fundamental limitation of chat-based coding tools: the loss of project context across sessions. Instead of relying on conversation history, Conductor externalizes context into persistent Markdown files stored directly in the repository.

This approach changes how developers interact with AI agents. Rather than treating the AI as a stateless assistant that forgets between sessions, Conductor treats the repository itself as the source of truth. Product goals, architectural constraints, technology choices, and workflow preferences are defined once and remain available to both humans and AI agents.

From Chat to Repository: The Context Problem

Most AI-assisted development tools maintain context in conversation logs. When a developer closes the terminal or switches tasks, that context disappears. For short, isolated tasks this is manageable, but for larger features or long-running projects, it becomes a bottleneck. Developers repeat explanations, restate constraints, and reestablish architectural decisions with every new session.

Conductor solves this by storing context in the repository alongside code. These Markdown files are version-controlled, reviewable, and shareable. They become part of the project's documentation and can be updated as the project evolves. This makes the AI's behavior more deterministic and the development process more reproducible.

Planning-First Workflow

Conductor encourages a planning-first approach. Instead of moving directly from prompt to code, developers define specifications and implementation plans before invoking code generation. These artifacts remain part of the codebase throughout the feature lifecycle.

This workflow is particularly valuable for:

• Feature development that requires understanding existing architecture
• Refactoring projects where constraints must be preserved
• Established codebases with complex dependencies

By planning first, developers can review and approve the AI's proposed approach before any code is generated. This reduces the risk of off-target implementations and makes AI contributions easier to integrate into existing code review processes.

Tracks: Discrete Units of Work

A core concept in Conductor is the track, which represents a discrete unit of work. Each track contains:

• A written specification describing what needs to be built
• A task-oriented plan broken into phases and subtasks
• Progress tracking directly in the plan file

Implementation proceeds only after the plan is reviewed. Because state is stored in the repository, work can be paused, resumed, or modified without losing context. Multiple tracks can be active simultaneously, and tracks can reference each other for coordination.

Early users have highlighted the track-based workflow as a practical improvement over ad-hoc prompting. Devin Dickerson, an engineering and product leader at Forrester, noted:

"My favorite thing about this extension is the tracks concept. I had been using the open source version of Conductor on my own prior to this release, and I ended up building my own feature slices. I can throw that away now that tracks are baked in."

Team-Wide Configuration and Standards

Conductor supports team-wide configuration, allowing projects to define shared standards once and apply them consistently to all AI-assisted contributions. This includes:

• Testing strategies (e.g., test-driven development requirements)
• Coding conventions and style guides
• Workflow preferences (e.g., commit message formats, review processes)
• Architectural patterns and anti-patterns

These standards are enforced across all tracks and AI interactions, ensuring consistency regardless of which developer initiates the work or which AI agent performs the implementation. This is particularly valuable for teams with multiple contributors or those working across different environments.

Test-Driven Development Emphasis

The extension explicitly supports test-driven development (TDD) workflows. Developers can define test requirements in the specification, and the AI will generate tests before implementation code. This approach ensures that generated code meets defined requirements and provides immediate feedback on correctness.

Navid Farazmand described his experience:

"I immediately tried to create something similar with .md files when Gemini CLI came out. Conductor is an order of magnitude better—especially the test-driven-development approach it takes."

Implementation Details

Conductor is implemented as a CLI extension that integrates with Gemini CLI. It operates through a set of commands that manage tracks, specifications, and implementation plans. The extension is open source and available on GitHub.

Key technical features:

• Markdown-based specifications: Human-readable and editable with any text editor
• Version control integration: Tracks are Git-tracked files, enabling history and review
• Phase-based execution: Work is broken into verifiable phases with clear completion criteria
• State persistence: All state is stored in the repository, not in memory or conversation logs

Getting Started

Conductor is available as a preview extension and can be installed from its public GitHub repository. Installation involves cloning the repository and configuring the extension with Gemini CLI.

Basic workflow:

1. Define a project context file with goals and constraints
2. Create a track for a specific feature or task
3. Write a specification in the track file
4. Generate an implementation plan
5. Review and approve the plan
6. Execute the plan phase by phase

Future Development

Google has positioned this release as an initial step. Future iterations will likely include:

• Enhanced track dependencies and relationships
• Integration with additional AI models and tools
• Improved visualization of track status and progress
• Better integration with CI/CD pipelines

The preview status means the API and workflow may evolve based on community feedback.

Broader Context

Conductor represents a shift in how AI-assisted development tools are designed. Rather than optimizing for conversational fluency, it prioritizes persistence, reviewability, and reproducibility. This aligns with professional software development practices where documentation, planning, and code review are standard.

The approach also addresses concerns about AI-generated code quality and maintainability. By requiring explicit specifications and plans, Conductor makes AI contributions more transparent and easier to audit. The track system provides a clear audit trail of what was requested, planned, and implemented.

For teams adopting AI-assisted development, Conductor offers a structured path that integrates with existing workflows rather than requiring a complete process overhaul. The emphasis on repository-stored context means that even if the AI tooling changes, the documentation and planning artifacts remain valuable.

The extension is part of Google's broader investment in AI-assisted development tools, following releases like Genkit Extension for Gemini CLI and Gemini CLI itself. Conductor's focus on context management and planning fills a gap in the current tooling ecosystem.

Conclusion

Conductor addresses the context loss problem in AI-assisted development by externalizing state into the repository. Its track-based workflow and planning-first approach provide structure for complex tasks, while team-wide configuration supports consistency across contributors. As a preview extension, it offers a glimpse into how AI development tools might evolve to better integrate with professional software engineering practices.

Developers interested in experimenting with Conductor can find it on GitHub, where installation instructions and examples are provided. The project welcomes feedback as it moves toward a more stable release.

Robert Krzaczyński is a software engineer with experience in web applications and AI applications in healthcare. He holds degrees in Control Engineering and Robotics and Computer Science.