Google Releases Conductor: a context pushed Gemini CLI extension that shops data as Markdown and orchestrates agentic workflows

Google has launched Conductor, an open supply preview extension for Gemini CLI that turns AI code technology right into a structured, context pushed workflow. Conductor shops product data, technical choices, and work plans as versioned Markdown contained in the repository, then drives Gemini brokers from these information as a substitute of advert hoc chat prompts.

From chat based mostly coding to context pushed improvement

Most AI coding as we speak is session based mostly. You paste code right into a chat, describe the duty, and the context disappears when the session ends. Conductor treats that as a core downside.

As a substitute of ephemeral prompts, Conductor maintains a persistent context listing contained in the repo. It captures product objectives, constraints, tech stack, workflow guidelines, and elegance guides as Markdown. Gemini then reads these information on each run. This makes AI conduct repeatable throughout machines, shells, and crew members.

Conductor additionally enforces a easy lifecycle:

Context → Spec and Plan → Implement

The extension doesn’t leap immediately from a pure language request to code edits. It first creates a monitor, writes a spec, generates a plan, and solely then executes.

Putting in Conductor into Gemini CLI

Conductor runs as a Gemini CLI extension. Set up is one command:

gemini extensions set up https://github.com/gemini-cli-extensions/conductor --auto-update

The --auto-update flag is elective and retains the extension synchronized with the newest launch. After set up, Conductor instructions can be found inside Gemini CLI if you find yourself in a undertaking listing.

Undertaking setup with /conductor:setup

The workflow begins with undertaking stage setup:

This command runs an interactive session that builds the bottom context. Conductor asks concerning the product, customers, necessities, tech stack, and improvement practices. From these solutions it generates a conductor/ listing with a number of information, for instance:

• conductor/product.md
• conductor/product-guidelines.md
• conductor/tech-stack.md
• conductor/workflow.md
• conductor/code_styleguides/
• conductor/tracks.md

These artifacts outline how the AI ought to cause concerning the undertaking. They describe the goal customers, excessive stage options, accepted applied sciences, testing expectations, and coding conventions. They stay in Git with the remainder of the supply code, so adjustments to context are reviewable and auditable.

Tracks: spec and plan as firstclass artifacts

Conductor introduces tracks to symbolize items of labor equivalent to options or bug fixes. You create a monitor with:

or with a brief description:

/conductor:newTrack "Add darkish mode toggle to settings web page"

For every new monitor, Conductor creates a listing beneath conductor/tracks// containing:

• spec.md
• plan.md
• metadata.json

spec.md holds the detailed necessities and constraints for the monitor. plan.md accommodates a stepwise execution plan damaged into phases, duties, and subtasks. metadata.json shops identifiers and standing data.

Conductor helps draft spec and plan utilizing the present context information. The developer then edits and approves them. The essential level is that every one implementation should comply with a plan that’s specific and model managed.

Implementation with /conductor:implement

As soon as the plan is prepared, you hand management to the agent:

Conductor reads plan.md, selects the subsequent pending activity, and runs the configured workflow. Typical cycles embrace:

1. Examine related information and context.
2. Suggest code adjustments.
3. Run checks or checks in accordance with conductor/workflow.md.
4. Replace activity standing in plan.md and international tracks.md.

The extension additionally inserts checkpoints at section boundaries. At these factors Conductor pauses for human verification earlier than persevering with. This retains the agent from making use of giant, unreviewed refactors.

A number of operational instructions assist this circulation:

• /conductor:standing reveals monitor and activity progress.
• /conductor:evaluation helps validate accomplished work towards product and elegance tips.
• /conductor:revert makes use of Git to roll again a monitor, section, or activity.

Reverts are outlined when it comes to tracks, not uncooked commit hashes, which is less complicated to cause about in a multi change workflow.

Brownfield tasks and crew workflows

Conductor is designed to work on brownfield codebases, not solely contemporary tasks. Once you run /conductor:setup in an current repository, the context session turns into a approach to extract implicit data from the crew into specific Markdown. Over time, as extra tracks run, the context listing turns into a compact illustration of the system’s structure and constraints.

Group stage conduct is encoded in workflow.md, tech-stack.md, and elegance information information. Any engineer or AI agent that makes use of Conductor in that repo inherits the identical guidelines. That is helpful for imposing take a look at methods, linting expectations, or permitted frameworks throughout contributors.

As a result of context and plans are in Git, they are often code reviewed, mentioned, and altered with the identical course of as supply information.

Key Takeaways

• Conductor is a Gemini CLI extension for context-driven improvement: It’s an open supply, Apache 2.0 licensed extension that runs inside Gemini CLI and drives AI brokers from repository-local Markdown context as a substitute of advert hoc prompts.
• Undertaking context is saved as versioned Markdown beneath conductor/: Information like product.md, tech-stack.md, workflow.md, and code type guides outline product objectives, tech decisions, and workflow guidelines that the agent reads on every run.
• Work is organized into tracks with spec.md and plan.md: /conductor:newTrack creates a monitor listing containing spec.md, plan.md, and metadata.json, making necessities and execution plans specific, reviewable, and tied to Git.
• Implementation is managed through /conductor:implement and track-aware ops: The agent executes duties in accordance with plan.md, updates progress in tracks.md, and helps /conductor:standing, /conductor:evaluation, and /conductor:revert for progress inspection and Git-backed rollback.

Take a look at the Repo and Technical particulars. Additionally, be at liberty to comply with us on Twitter and don’t neglect to hitch our 100k+ ML SubReddit and Subscribe to our E-newsletter. Wait! are you on telegram? now you possibly can be part of us on telegram as properly.