Follow mode, for exploring architecture

[nhemsley]

Enhanced "Follow Agent" with Codebase Exploration Mode and Timeline Scrubbing**

Hello,

I'd like to propose three enhancements to the "Follow Agent" functionality, somewhat interrelated.:

1. Codebase Exploration Mode

When working with Claude to implement features based on existing patterns, It would be handy to use the AI's exploration abilities to understand complex relationships (e.g., how Zed's collab server, .proto files, and events interconnect). An "Exploration Mode" would let the AI:

• Systematically read and document the codebase
• Build a structured knowledge graph (hierarchical/DAG)
• Update the map as it discovers new relationships
• Present findings in a navigable, scrobbable timeline, or something similar

2. Timeline Scrobbing for Follow Agent

Following the AI in real-time is valuable but cognitively demanding. A timeline interface would let us:

• Record all AI actions with timestamps
• ScScrob through the work history
• Jump to specific decisions or changes
• Review at our own pace

3. Interactive Timeline Queries

While scrubbing through the timeline, we could select any line or symbol and ask the AI for clarification. Click on a function call → "Why did you choose this approach?" or select a variable → "What does this represent?" The AI would have the full context of that moment in time to provide accurate explanations about its reasoning and decisions.

Example: Claude explores how Zed's channel system works, builds a map of proto→handler→client flow, and you can scrub through both the exploration and implementation to understand its reasoning, asking questions at any point to clarify specific choices.

These features would transform "Follow Agent" from passive observation into an active learning and review system.

[nathansobo]

Hey! Thanks so much for sharing! I love this idea, and I think there's also a correspondence with the timeline of the unfolding conversation. We're working on the primitives to do this properly, in an early research phase now with active development starting in a couple weeks. It may take some time because I think we need a more a more abstract and scalable foundation than Zed's existing collab implementation really supports.

What brings you to share? Thanks again!

[nhemsley]

OK, thanks!

One of the joys of using zed & claude, is the follow mode, however once you look at a section, follow mode disengages, the AI charges forth, and you no longer see the flow. Still amazing!

WRT me sharing, I have been playing with AI, and discovering alot of the pain points:
ai one-shotting
ai overwriting, previous code, without committing.
ai backing out git commits, without pushing to main, or creating a new branch.
etc. etc.

I've always had an interst in developer tools, so using AI with zed has given me a bunch of ideas per se..

I have previously worked on something related: https://github.com/fluid-notion-systems/ize. Which actually may relate to this, but I had to shelve development for a while, and none of the version tracking code is actually there. just the research documents. (all written by claude in zed..).