Transcription: OpenAgents Episode 108 - GitHub File Explorer

Source: https://x.com/OpenAgentsInc/status/1806401750171373650 Wiki source: https://raw.githubusercontent.com/wiki/OpenAgentsInc/openagents/Video-Series.md Media title: OpenAgents - Episode 108: GitHub File Explorer We add a GitHub file explorer to o... Upload date: 20240627 Transcription model: mlx-whisper/mlx-community/whisper-tiny Generated at: 2026-06-01T18:00:02Z

Machine-generated transcript. Review speaker labels and wording before using this as quote-grade source material.

[00:00] Christopher David: Okay, we are now able to index and query code bases via GripTow that works great.

[00:05] Christopher David: For making the actual edits to files, we're going to need some sort of, I don't know,

[00:10] Christopher David: file explorer.

[00:13] Christopher David: This is the Cloud Artifacts version where they just kind of show you a file and they edit

[00:17] Christopher David: it and then kind of put it there which is nice.

[00:19] Christopher David: But again, we want this to be active on our actual GitHub repo.

[00:25] Christopher David: I don't want to have to keep copypacing into my GitHub repo.

[00:28] Christopher David: So we built a little file explorer.

[00:32] Christopher David: This is not going to be the final UX.

[00:34] Christopher David: We still got a reason about what the actual UX should be.

[00:39] Christopher David: But it was able to drive this out via Cloud Artifacts using the GitHub FS library from WANIX.

[00:53] Christopher David: I'll give you a tour of that in a second.

[00:55] Christopher David: The first let's kind of revisit what exactly we're building here.

[01:02] Christopher David: Okay, so there's the building blocks.

[01:05] Christopher David: We did index and query code bases in the last video.

[01:08] Christopher David: We did read GitHub repo in this video.

[01:12] Christopher David: So the building blocks are to be enabled, the agentic loop.

[01:16] Christopher David: The agentic loop is the user asks, do something.

[01:20] Christopher David: What are the defines the relevant context via Grettele for code based indexing or for

[01:27] Christopher David: when we want to have it search also through PDFs for any other files we use Ragee for that.

[01:32] Christopher David: Based on that context, we want to make a plan.

[01:35] Christopher David: That plan will be basically having it edit files in a certain way.

[01:38] Christopher David: It'll go through and find the files to edit them.

[01:43] Christopher David: Some probably some sort of confirmation from the user are you cool with this.

[01:48] Christopher David: It's about to be pushed to your GitHub and then the changes are merged and then the user can repeat that.

[01:53] Christopher David: Building blocks needed for that indexing query and code bases that's done with Grettele reading the GitHub repo.

[02:00] Christopher David: That's just a combination of the GitHub file API.

[02:04] Christopher David: The GitHub API plus this.

[02:09] Christopher David: WANIX has this.

[02:12] Christopher David: Go language, GitHub file system, and basically.

[02:20] Christopher David: Given a GitHub repository and access token, this file system will use the GitHub API to expose a read right file system of repository contents.

[02:27] Christopher David: It's roots will contain all branches as directories.

[02:30] Christopher David: I just generally like their approach because they've got other different file systems.

[02:34] Christopher David: We're doing all sorts of different things with enabling the agent to manipulate files in this file system that's built in the browser of the OSU.

[02:42] Christopher David: More on that stuff later, but I was happy to be able to pull that library into our code base and do things like on the URL, bring this down.

[02:58] Christopher David: This is the WANIX repo.

[03:02] Christopher David: If you load any repo here, let's do Open Agents Ink Slash.

[03:10] Christopher David: Open Agents.

[03:13] Christopher David: It'll show you all of the branches loaded there.

[03:17] Christopher David: It'll show you the name is the branches.

[03:19] Christopher David: How many files in each one?

[03:21] Christopher David: Got a little file explorer here without proper nesting and slow janky.

[03:26] Christopher David: If you click on a file, it'll pop up and this will preview.

[03:31] Christopher David: I've also got a larger view of this over here.

[03:35] Christopher David: Show the branches.

[03:38] Christopher David: Just was getting the basic UI showing this VHMX, query in the GitHub API.

[03:47] Christopher David: We have all this content now and we're going to have some visualization of the changes that the agent is making and showing it to you kind of like

[03:57] Christopher David: Cloud is.

[03:58] Christopher David: I think I'm now super motivated to do this because there's so many limitations on this cloud thing where I'll be working happily for an hour and then it'll say you run out of messages for the next four hours.

[04:12] Christopher David: I'm excited to get this built our own version of this so we don't have that limitation.

[04:17] Christopher David: I'll use it in their same API.

[04:21] Christopher David: We have a file explorer, cool and then checkboxes.

[04:28] Christopher David: You can envision maybe if you want to manually send some files to a query or to an LLM call.

[04:36] Christopher David: This would be good for that.

[04:39] Christopher David: Grab tile handles a lot of that querying for us.

[04:41] Christopher David: Okay, let's step back to this.

[04:43] Christopher David: So other building blocks we need.

[04:46] Christopher David: A planner, a given a prompt and the context come over the plan and then turn that into code at it.

[04:54] Christopher David: And then right. Fortunately, very basic versions of this entire flow accepting the artifacts are

[05:01] Christopher David: in a hackathon project that some team did like a year plus ago showing how you can do a

[05:13] Christopher David: comment and PR via GitHub.

[05:21] Christopher David: They use like GPT 3.5 maybe on this like one of the earliest versions.

[05:28] Christopher David: I know they use DaVinci like the old models but they were able to like have it successfully generate a

[05:34] Christopher David: PR.

[05:36] Christopher David: This used to be open source and then they closed source it but I had backed it up and so here it is.

[05:41] Christopher David: And then they've got things like, you know, they built this little file system and we're able to kind of like take a bunch of the algorithms that they use.

[05:50] Christopher David: And so that's a just some time.

[05:54] Christopher David: And then that sort of peer-butt algorithm will help us have basic versions of the planner

[06:05] Christopher David: and maybe the code at it's definitely the writing of the GitHub repo or the showing how to interact with the API.

[06:10] Christopher David: So think we have to extend our GitHub FS a little bit to handle branching.

[06:15] Christopher David: And then the code edits, the peer bot repo has a very basic version of that but we

[06:24] Christopher David: will probably use A-dark because I think they have got the latest state of the art for that.

[06:29] Christopher David: Although A-dark might be super based on GPT 4 where obviously you can be using some of it for a lot of this.

[06:35] Christopher David: And then artifacts, this is not necessary for like this flow of here but if we want to add artifacts,

[06:44] Christopher David: thank you to Pliny for finding the artifacts section of the SONIT system prompt.

[06:56] Christopher David: Okay, additive plus artifacts, thinking with a good GitHub code base, all of the source plus plugin based

[07:03] Christopher David: extensibility.

[07:05] Christopher David: All right, next, maybe we'll do the plan. See you soon.