🔎 Seeking Testers: The New Codebase Investigator Agent in gemini-cli

[silviojr]

We've been working on a new way to help you understand complex codebases, and we're looking for your help to test an experimental new feature: the Codebase Investigator Agent.

🗞️What's New?

Simple code searches are great for finding specific lines, but they often fail when you need to build a complete picture of how a feature works across multiple files. The Codebase Investigator is an autonomous agent that tackles these complex, multi-step investigations.

Instead of just searching, it explores your codebase based on a high-level objective. It's perfect for ambiguous questions like "How does our caching layer work?" where the answer isn't in a single place. When it's done, it provides a comprehensive report with a summary, a full exploration trace, and an analysis of all the code it deemed relevant. That report can be used by the main agent to accomplish complex tasks or just provide you with rich insights.

🧪How to Test It

This feature is enabled as a tool by default in our latest preview release (v0.10.0-preview). The CLI agent can choose when to invoke it, but you can always nudge the agent to invoke this new feature, like: “Use codebase investigator to understand how the authentication layer is implemented and write unit tests for all its modules”.

⚙️How to configure it?

Can you disable it? Of course! Can you change its settings? Yes! You can increase/decrease the maximum number of turns that the agent can take in its investigation as well as the amount of thinking it puts into the work.

{
  "experimental": {
    "codebaseInvestigatorSettings": {
      "enabled": true,
      "maxNumTurns": 15,
      "model": "gemini-2.5-pro",
      "thinkingBudget": -1
    }
  }
}

💬We need your feedback!

This is an early version of the agent, and your feedback is crucial. Please reply to this thread with your thoughts, especially regarding accuracy, agent behavior, performance and use cases!

Thanks for helping us build the best version of Gemini CLI! ❤️

--with @abhipatel12

[diegomazuco]

Bom dia, acabei de atualizar a versão "Release v0.10.0-preview.2" e achei muito interessante e funcional essa novidade, já estou testando em um dos meus projetos e já me ajudou muito!

[bobmcallan]

The Codebase Investigator, needs to output it's findings to a markdown, and allow the user to work with and refine the findings. Then create a stepped implementation of the result. The current output format is to difficult to read, and current one-shot approach is not usable for refactoring across multiple files.

[kpe]

looks like the codebase investigator is triggering the quota limit very fast (or the quota limit in my ai pro subscription is just broken, after waiting for 24 hours, it gets triggered after less than a minute with the codebase investigator), not sure how to look for the cause, or monitor how much quota is remaining.... hope someone could give me a hint
(otherwise, a really neat feature the codebase investigator, yesterday i could get some good results)

[diegomazuco]

Também tive essa impressão hoje, no primeiro comando de análises que fiz, em 30 minutos já acusou que o limite do Gemini PRO havia se esgotado.

[Whamp]

same thing happens to me too, although i think the problem is the undefined unguaranteed limits in the ai pro plan.

[jthomazini]

This issue is also happening to me. Or they have severely reduced the size of the quota, or something is broken, or the agent is using all the quota intentionally, and they haven't disclosed this quirk.

[gil-ch-palo]

The current user experience is suboptimal when the maxTurns limit is reached, as the process terminates without providing any partial results.

I suggest one of two solutions:

1. Return Partial Results: Upon hitting the limit, provide the user with all findings gathered up to that point.
2. Offer Continuation: Prompt the user with an option to continue the investigation beyond the set limit.
3. Provide Both 😁

[timrichardson]

I tested this by manually invoking it in the current stable release. I have model routing turned on, and it routed some requests to flash-lite while being quite intensive with pro, and nothing went to 2.5-flash.

I don't miss aider as much now.

[metheai]

Hi community and developers,

I am using Gemini CLI for all my work and really enjoying it. It helped me complete perfect deliverables so far.

However, about the Codebase Investigator Agent, I am not in a position to give any positive feedback yet because:

• I literally have never seen it being completed with a success message.
• The core challenge is maxTurns, preventing a complete analysis most of the time, especially in huge codebases.
• Even though maxTurns is not reached, the system throws an error (sorry, I'm unable to send an instance right but I will edit my message once I encounter with any); especially in huge codebases. By huge, I mean probably over 5k LOC.
• There are no detailed explanations actively mentioning which file or code piece is being analyzed currently, or what it discovers during the process; which leads to an impression that the tool is stuck or frozen.

I'm a big fan of Gemini CLI and really support Codebase Investigator Agent. Please regard these as constructive criticism, not unbacked swagger.

I really hope this helps.

The edit is below:

Error types:

• Model stream ended without a finish reason

[Jkith-dev]

I've only gotten this with the codebase investigator agent so far. Hasn't been useful so far.