Cross-Repo Changes with AI Coding Agents

Why cross-repo changes go wrong

Before changing code, an engineer asks:

• Where is this behaviour implemented?
• Which repositories are involved?
• What depends on the component I'm changing?
• Does a similar implementation already exist?
• What else could be affected?

An agent that sees one repository can't answer most of these. So it does the visible half of the change and leaves the rest broken.

A safe workflow: Understand → Plan → Build

1. Understand

Before writing code, find every repository the change touches. Look for consumers of the API, users of the shared type, and subscribers to the event.

2. Plan

Write the change down per repository, in order: shared type first, then the producer, then each consumer. Review it before any code is written. For example:

1. shared: add optional `currency` field to Order type
2. api: populate `currency` in GET /orders
3. web: display currency in order summary
4. mobile: same as web (separate repo)

3. Build

Let the agent execute the approved plan, one repository at a time, with the related repositories still visible as context.

Keep conversations separate per repository

If you work across several repositories in one agent conversation, context from one leaks into another, and the agent may apply one service's conventions to a different one. A cleaner pattern is one workstream per repository, with shared read access to the rest of the system.

Doing this with CodeZero

CodeZero is built around this workflow. It runs in VS Code alongside the agent you already use (Claude Code, Codex or Copilot) and doesn't replace it or change its credentials. CodeZero owns the context, and your agent owns execution:

• Understand: browse every indexed repository, trace features, and ask questions answered with evidence from other repositories.
• Plan: a plan can name several repositories. CodeZero keeps them in scope for the agent and shows the dependencies in Tasks.
• Build: choose how changes are handled. Planned mode always plans first and blocks edits until you approve. Normal mode asks you to confirm each edit. Auto mode plans only when a change is large or ambiguous.
• Separate workstreams: each repository keeps its own chat, while organisation-wide context is available in all of them.
• Safe writes: indexed repositories are read-only. If a plan needs to edit one you haven't cloned, CodeZero offers to clone it first.