I've been waiting for this feature | full codebase indexing in RooCode

Augment indexing is really good

Can you add Claude Code and AugmentCode as agents in your benchmark? I wonder how these two would compare to RooCode running in agent mode.

Dope

Great content as always! You mentioned that you're looking for perhaps new subjects and things to cover. I would really like to see a head-to-head comparison of the state-of-the-art codeing with AI. Claude Code (using Opus and Sonnet interchangeably) vs Roo Code with the best models possible vs Augment Code (my favorite) trying to see which one, beyond benchmarks, would actually get the job done the fastest and in the most comprehensive way possible. It would be interesting, preferably, to have this work on legacy code—for example, like the real-life stuff, not building a game or building a to-do list app or whatnot that's overdone.

But actually working on some legacy code and seeing which one would come out on top would be really interesting and really down to earth and to the point.

You need to re-visit the MCP documentation. I states that it is very sensitive to USE and DO NOT USE when it comes to tools.

Please compare it with Augment Code's agent.

I've been playing with code-indexing today. So far it is another toy to play with. The need to explicitly prompt to use the tool is really annoying and hopefully this gets better in the future.
But when it invokes, it feels like magic and it always pointed to correct place in the codebase. I will need to adjust my prompting to utilize code-index correctly.

I am exited already :D

I hear CAT but I do not see CAT. WHERE CAT?!?!?!?!?!?!?!

It should save the vectorized codebase in a local database in the root of the project.

Hi. I'm a follower since many months now. I don't know how to code and I have started coding since AI has enabled commoners like me. You were the only person who helped me better understand how to customise system prompts in roo. It was a game changer for me. Because of this one change, I have been able to create an agentic system which is helping me develop full stack apps for the first time. I have even created two really functional as well as good looking android apps. If you could cover any of these topics again and share your workflow or best practices from end to end development, completely code illiterate users like me could be helped greatly. I even created an mcp tool which helps me do these things:

Project management
Task management
Context management
Memory management

I created it borrowing from various ideas. Would you mind giving me feedback if the code I'm generating is good or bad?

These are some recommendations. I never miss any video from you. I don't like vibe coding even though I can't tell the difference in code. I do it purposefully and with planning and documentation. You are among the few who inspired me to build my workflow.

took em long enough to catch up to augment code.

i will say this, i have unofficially incorporated semantic search into my augment code extension and it has just wiped the board in terms of productivity and focus with claude 4. with the added web search and scraping, i have utilized concepts from some white papers and public repos like agentseek, roo code, cline, etc, along with augments built in features to even further enhance productivity.

The current race for these agent coders,
is the ability to breakdown complex tasks to subtasks, boomerangs, run by independent agents who SHARE context with the orchestrator, all running in multitude simultaneously while the main orchestrator agent handles the call backs.
I have the libraries, connections, the entire framework to run 6 agents in such capacity. i can even get it to run multiple ide's instances and push tasks to those agents from the orchestrator, but it cuts off on handling the individual task completion summaries back to the orchestrator. By using augment codes memory features through development with a focus on general workflow productivity stats, you can then utilize the resulting data summaries, create a user guideline prompt from it, then that enhanced methodology is built into future sessions.

Unfortunately i cant get into the dev program for augment code but that wont stop me. I need My agent office to build my vibes and my business.

Draw arrows like a pro just use ShareX to do screenshots.

I don't know. Seems like a distraction.

I am using Devstral 26b q8 together with Qwen3 30b a3b q4. The Qwen model is lightning fast because it is effectively 1.5 GB memory access per forward pass compared to 26 GB for Devstral. So for searching, documenting and other non coding tasks, I can launch that.

I have a 128 GB MacBook Pro, so I can simply load both models at the same time. However if you don't have that luxury, there is a project called swap LLM that can save state of the current model and in a few seconds swap it for the alternative.

Wait a minute, don't Claude Code, Windsurf, Cursor (et al) already do codebase rag/indexing??????/

Code indexing was introduced to me with Augment Code which was magic to me. I’m glad I can see this in other tools.

I agree that Augment chat is strange. It is also strange that it cannot maintain context when you switch from chat to agent. Who else does it that way? BUT, Augment Code with Claude 4 is now an actual tool I use. In fact, I pair Augment Code with Claude Code now. Both flat-rate, which makes a huge, huge, huge difference in costs compared to API costs of the Cline/Roos of the world (at the amount I use it anyway).

Augment forced to lower its costs 😂😂😂 admittedly, it became more valuable when they moved to Claude 4 at the same cost but I’m glad there’s options now!

I haven't been able to make qdrant work with a remote deployment I did, it gives me `Error - Failed during initial scan: fetch failed`, which I find a bit interesting as I haven't been able to solve for

I found augments chat to be clearly inferior to their agent. I never use their chat even when I'm just asking a question about the codebase or something.

deepseek r1 0528 just dropped today you need to test it sir

What deep research system do you use?

I also noticed that the augment code's chat is not using their context engine. The strange thing is that I started to noticed this new pattern after they moved to Claude 4 so my assumption is that they tried to rush things to be the one of the first companies that are using Claude 4 and didn't test the new model (with their existing system prompts) properly.
I really would like to have an option where I can pick the "old" or "new" model for a short period of time while they do some proper testing with the new model instead of just forcing everyone to use the new model on day release.

The context condensing seems to work GREAT!

For the thumbnail of this video, did you intentionally try to look like the fat guy from Jurassic Park? :)

you are AMAZING as usual!

Need google's new OP embeddings

Kind of surprised it wasn't part of it yet.

didn’t work for me got an error

Is this different from the way cursor indexes your codebase?

Installed the codebase index today and started using it. Tested a new workflow. I have a detailed markdown plan broken into phases for implementing a new feature in my codebase. I use RooCode with VS Code LM API access to Sonnet 4, and have it implement a phase of the plan. Then I use Augment Code to perform a comprehensive implementation audit to see if anything detailed in the phase was missed. For the most part Roo Code completed each phase correctly except for a few where it might have missed about 25% of the requirements. So pretty good overall in comparison to Augment Code. This is a 50,000 line codebase BTW.

finally indexing , I can unsubscribe cursor soon🎉 ty for all the videos❤

Please keep us posted on the development of your codebase crawler mode.

Lmstudio and cline/ roo does not show token usage locally, i reported this as a bug months back, still does not work. Would be nice to see how much tokens used locally to calculate price for online tools, but these crazy guys don't get it

Love the video. I’m curious to know how Roo did with the same prompts without code_search though.

The cat at minute ~19 is a special effect of roo code ? or it was your pet hungry ? 🙂