The modern developer's day is often defined by a frustrating cycle of waiting. You write a feature, commit the changes, push to a remote branch, and open a pull request. Then, the silence begins. You wait for a reviewer to find a window in their schedule, wait for the inline comments to appear, and then enter a fragmented dance of jumping between a GitHub comment thread and your IDE to implement requested changes. This asynchronous loop is the industry standard, but for the team behind the Zed editor, it has always felt like a fundamental disconnect between how software is actually discussed and how it is recorded.

The Architecture of Delta Streams and Stable Identifiers

To solve this friction, the Zed team developed DeltaDB, a version control system designed to treat the conversation and the code as a single, unified stream. While Git operates primarily on the concept of snapshots—capturing the state of a project at a specific commit—DeltaDB focuses on the delta stream. This approach records every single operation that occurs between commits, effectively turning the history of a file into a continuous movie rather than a series of still photographs.

Central to this system is the implementation of stable identifiers. In a traditional text editor or version control system, a reference to a piece of code is usually tied to a line number. However, as soon as a developer adds a new function at the top of a file, every subsequent line number shifts, breaking the link between a previous discussion and the actual code. DeltaDB assigns a unique, persistent ID to each delta. This means that even if a block of code moves from line 10 to line 500, the identifier remains constant.

This technical shift enables a bidirectional map between intent and execution. A developer can click on a line of code and instantly trace it back to the specific conversation—whether with a human teammate or an AI agent—that generated it. Conversely, a user can look at a chat message from three days ago and jump directly to the current state of the code that resulted from that specific prompt. For AI agents, these identifiers are critical. An agent can now query the context of a specific code block by calling the previous agent that worked on it, asking why a particular logic path was chosen without needing to re-scan the entire codebase.

The Collapse of the Code Review Cycle

The introduction of DeltaDB does more than just improve traceability; it fundamentally challenges the necessity of the pull request. The traditional PR exists because the discussion about the code happens in a different place than the coding itself. We use PRs to bridge the gap between the private worktree and the shared repository, using comments to retroactively explain decisions that were made hours or days prior.

DeltaDB eliminates this gap by integrating a conflict-free replicated worktree. This allows multiple humans and AI agents to edit the same files simultaneously across different machines in real-time. Because the conversation is baked into the version control stream, the review happens while the code is being written. When an agent modifies a file via the terminal, the human collaborator sees the change and the accompanying reasoning instantly. There is no need to wait for a push to start a discussion because the discussion is the mechanism of the push.

In this environment, the role of Git and Continuous Integration (CI) shifts. They are no longer the primary venues for human collaboration or the gatekeepers of social consensus. Instead, Git and CI are relegated to their most efficient roles: running automated tests, verifying builds, and managing the final connection to the external production environment. The social act of software engineering—the debating, the refining, and the reviewing—moves entirely into the conversation stream. The result is a workflow where the code is not a static artifact to be reviewed, but a living output of a continuous dialogue.

Zed is preparing to roll out DeltaDB to a select group of early users in the coming weeks, moving the industry closer to a world where the boundary between talking about code and writing it finally disappears.