Where do you face the biggest challenges with Gemini CLI?

[blanca-delgado-parra]

We want to focus our efforts on the most critical issues, so we would love to understand what are your biggest pain points. Below are some potential suggested areas.

Feel free to add others and expand on your feedback by replying to this thread in free form.

Thank you!

[Manamama]

Workflow problems, for sure. For months (or years?), I had to have AIs self recode even, for that end: starting from preprompts aka system msgs (the easiest), modded TUI-level code fragments, forked MCPs etc.

Very random (semi-)ready tricks:

• Enhance sequentialthinking Tool to Ensure Full Feature Utilization by MCP-Enabled AIs: analysis by Grok of Gemini CLI's using it, from raw logs modelcontextprotocol/servers#2332 (comment)
• A Positive Report: Vespera's Role in Strategic AI Problem Solving (Case Study: River Puzzle) EchoingVesper/vespera-atelier#26
• Feeding AIs Socratic questions, puzzles, and guides: https://github.com/Manamama/Puzzles_for_AIs/tree/main/docs to frame them (=preformat, "NLP", paradigm shift...) before some serious ideation or coding tasks...

[wiiiimm]

Gemini CLI isn't the problem right now. The bottleneck is the model's ability to understand and debug code properly. Problem E - "I ran out of Quota before I'm able to make meaningful progress" - usually happens when I try to set Gemini CLI on a debugging task only to watch it go in circles for close to 30 minutes and burnt through a lot of quotas. It also manifest as Problem B and D as well.

[jochenkirstaetter]

Overall I like using the CLI.

There are a few things that need some attention.

Probably not your beef but the GitHub MCP server regularly times out and isn't available. Restarting CLI or refreshing MCP resolves this sometimes.

Working with git always wants to add untracked files using git add . instead of using the modified files only. Also, there should be an option that no commits or pushes shall be attempted. I guess that could be solved in GEMINI.md instructions?

When asking for a plan to solve a certain issue I don't expect to see files being changed actually but seeing the suggestions in the output.

Working with git should probably create a branch automatically before changing files.

Better use of existing files in a repo like .gitignore would be appreciated. Run several times into input token limits because of CLI using the generated documentation folder instead of respecting ignores.

What's the issue with an .aiexclude file? CLI won't work at all? That's not the purpose of that file, or?

Allow use of softlinks: GEMINI.md - > AGENTS.md

Provide a toggle to enable/disable global extensions, commands, MCP, and Plugins within a project. Right now, I have a mixture of global addons that I use in all projects and then I have a bunch of copies of addons that I use in certain projects/folders only. It's annoying and unnecessary duplicates. Having a project-specific settings override to disable global addons would be much appreciated.

That's just back of my head...

Cheers

[jochenkirstaetter]

Ooh, please work with JetBrains on the integration into their IDEs. Please! 🙏

[chrispository]

lately the bash shells it executes seem to get hung up indefinitely. whenever it's going to run a shell, there should probably be some logic in place of "what to expect" or "how we can verify it's in progress and not messed up".

as someone else said - debugging is a bit rough. i've also had it hallucinate a lot of commands while trying to set up google cloud workflows through the CLI.

i think being able to pull and ground from up-to-date documentation like context7 MCP does would be amazing. but I feel like I shouldn't have to rely on an MCP server for that. if i'm using Google Gemini to set up Google Cloud stuff, I would expect that to be enough.

[kneelesh48]

On first impressions, the Gemini cli takes a long time to start after the first install or after an update. That made a bad first impressions. After replying to every message, it waits for 2 seconds, doing nothing, which is annoying and a waste of developer time.

I ended up not using Gemini cli much. I did use Gemini-2.5-Pro in Roo Code. It likes to write a lot of comments, which is good for learning something new, but it's not good for agentic coding. You can easily tell the code was written by AI because a normal dev doesn't write so many comments

These are the reasons I stopped using Gemini cli and Gemini model.

[ssp2023]

Largely agree with Gosu.

Gemini 2.5 has just fallen way behind and Kimi K2 and Qwen are better models are surpassing it with Claude and gpt5 continuing to be ahead. It's no longer worth the headache of fixing its very enthusiastic but incorrect edits even though it's free.

Many of us realize that the 1M context window isn't real and the lore is around 200k which puts it on par with other models even in its most standout feature.

cli 2, from a usability perspective has been really wonderful and under hyped. You should get Logan to help you. The ide integration seems better than the others afaict.

[BobTB]

The main problem preventing normal usage are insanely low quotas for Gemini PRO subscribers! Its incredibly frustrating to get blocked after 1-3M tokens sent and not even 10 prompts. (the output tokens were 7k only). Useless. Why are we paying for the PRO subscription if we can not use this in the Gemini CLI and are treated as the free users or worse.

I think both teams (CLI and web chat) are not even talking one to another, as the web chat version of Gemini with the Pro subscription rarely blocks us as a paying customer, while the Gemini CLI acts as we are worthless non paying customers even if we "Login with Google"

It should know we do pay something per month and give as some normal limits, not 20 minutes of coding with Pro and then wait for a full day to repeat. Very sad and disappointing.

It is advertised as 1000 request per day, 60 RPM, 1M context window for free users, and as a paying PRO subscriber this is not even close. Ok, probably for Gemini 2.5 Flash, but still, 10 requests with Pro? 20? and its gone ... see you next day.

At least let us know WHAT THE LIMIT IS for the Pro tier.

P.S.:

Also why it wastes millions of input tokens for a 1000 line single js file is beyond me. No optimization at all. It loops it 20 times for fun to waste tokens? No other explanation here.

[boilthesea]

Tool use errors, often runs into trouble applying changes to a file, sometimes it adapts better than others but once it starts having issues it tends to repeat them as is the case with gemini api + other tools. Was hoping that integrating the model with the software would work better and I suppose it does, but it's still a problem. Just had repeated errors in a session today.

Additionally hallucinations sometimes masquerade as tool use errors. I've run into multiple instances in the past where the model reads a portion of a long file (up to line whatever), hallucinates a code block that it was looking for but didn't find because it was beyond the set number of lines it looked at and then it tried to re-read the same portion then re-edit that non-existent block over and over until I stopped it, asked it to tell me about the block and then discovered it wasn't there and manually told it what lines to read above the original range to find the real function.

The hallucination issue feels like something that can be prompted away ("if you repeatedly can't find the same code block in a file you've only read a portion of, try reading the remaining lines to look for the functionality you want to edit"), but I feel like that prompting should be built into the tool. It already switches to use a different editing tool on real tool use failures, and this feels like that kind of issue.

The last one probably is just something I need to work on in a gemini.md file and not bother you with, I apologize, but I'll mention it just in case it strikes your fancy: An ask mode would be helpful. All kinds of agents really really want to get started on making changes like a pup that can't stand still when a ball might be thrown soon, whereas I want to start every feature/fix with a conversation, basically, I want to chat with my code for a while before I'm ready define the plan and start implementation. I prompt around this manually at the moment, it's fine, but if I could /ask or /mode ask and not have to fend off the eager agent until I'm ready and then /mode act, that would be awesome.

[BobTB]

Yes, this ask mode would be mighty fine as it, as you said, now just starts to modify code.

[Manamama]

@boilthesea Re: 'whereas I want to start every feature/fix with a conversation, basically, I want to chat with my code for a while...'
I call it a bulldog mode. It is due to the default system instructions, alas, that (somehow rightly for most of us) say: "save the tokens, start the work asap, no philosophising, be taciturn too, the User expects it". (System msg > your GEMINI.md one, yes: I discovered it the hard way when Gemini AI claimed sudo on virtual box would not work and soon thereafter also refused collaboration, claiming that any run_shell shall not work either, sic).

I modded these some 3 weeks ago into: https://github.com/Manamama/Puzzles_for_AIs/tree/main/docs/strategy/system_message. Do read also the somehow rambling :) and draft "pop AI psychology" musings how it all works: https://github.com/Manamama/Puzzles_for_AIs/tree/main/docs/brainstorm/conceptual_notes plus related.
Works for me; modding also it had lowered my initial frustration a lot, YMMV.

[bpavuk]

B, C, and performance. Neovim is snappy, LSPs I use are snappy, and Gemini CLI is snappy... only if you use it in ACP mode (that protocol that integrates with Zed and Neovim code-companion).

but the performance of CLI interface itself is so awful that I want to fork Rust-based Codex CLI and just add Gemini model support, although GPT-5-mini is not that expensive and works better than Gemini 2.5 Pro, so I might switch.

[k90k90k90]

I would definitely want auto complete on the shell mode.

[SaschaHeyer]

Usage and quotas, e.g, visibility of it is a big challenge.
#4300

[timothyfroehlich]

Nothing worse than getting blocked by quota when it's halfway through a change. You have no warning, just all of a sudden "you're blocked, use the dumber model instead".
Claude has clear indicators when you're approaching your usage limit, so I know that I need to start wrapping things up.

[santhoshkumarCodes]

A) Authentication: The setup process is painful, confusing, or not well-documented.

Here''s a PR to fix painful authentication process when using API key: #6994

[hoteye]

Handling a massive project with a single client can lead to progressively decreasing operational efficiency. The workload can be broken down into simpler tasks, which are then tackled by multiple clients respectively. Additionally, there can be one client responsible for maintenance and coordination. In this way, they can work collaboratively to enhance efficiency. These clients can coordinate with each other through the A2A protocol.

[hoteye]

There's another pain point: I have to stay by the computer all the time. It might halt at any moment, and I have to constantly monitor its output to know exactly what it's doing. This is actually quite taxing for the user.

I need to be able to use my phone app to monitor its operational status and issue commands to make it work at full capacity.

I can made it for cli

[nxorable]

In contrast to Claude Code, I find it a challenge to get Gemini CLI to first read code related to a spec prior to suggesting changes to this spec while iterating. I want to take advantage of the large context that Gemini CLI offers when I'm using it.

[verdverm]

My understanding is that gemini-cli is what backs the Gemini Code Assist in VS Code. I want to work within my IDE, using the CLI is a giant step backwards. The relation between the two options is opaque and it would be great if the full VS Code extension was part of this project (as separate from the minor integration points to display diffs today). I definitely want to be able to get into the guts of these things. It's new, it's shinny, and they are generally not great out of the box. Being able to adjust them has been invaluable in getting better outcomes and ROI

What I'd really like to see is more feature parity with what Copilot offers. That being said, there are features they do not do well that could be improved upon. AGENTS.md is another project that comes to mind (though I really find something in the middle of all the agent instruction options ideal, but everyone wants to push their own agenda...)

[rclabo]

I really wish there were a consistent way to review the changes the Gemini CLI and Code Assistent are proposing. At this point, it seems like there are 3 or 4 different ways to approve / decline changes and no rhyme or reason as to why it picks one UI over the other for them. This makes it hard to get into a flow since you never know which type of ball it's going to throw at you. It also means there is no consistent keyboard shortcut to approve or decline the proposed changes. Ugg.

Another issue is that sometimes it proposes changes to a file that I mostly want, but I would like to make a few edits. It's unclear the best way to do that in the flow. I wish there were better real-world coding videos showing the use of Gemini CLI and Gemini Code Assistant in VS, especially for ASPNET Core web applications. I've searched but only see a ton of videos on installing it and "getting started." But nothing that is a bit more meaty. It'd be great if there were some videos of using Gemini CLI and Code Assistent in VS to create a small webbite during a 1 or two-hour coding session, and have the dev call out the tradeoffs in different workflows and their reasoning for picking the flow they are using.

One more feature I wish it had is the ability to click a microphone icon (or hot key) to have speech turned into text in the prompt window. Don't auto-submit that text; I often want to edit it a bit or augment it manually. I use this feature a lot in ChatGPT to provide background context on a task I want done. It's much faster to speak paragraphs of background info than to type it. But Gemini Code Assistent in VS doesn't have this feature, which means that I either have to type paragraphs of text I could dictate or I have to use another tool to dictate the text and then cut and paste that into Gemini Code Assistent in VS, neither is ideal.

[rclabo]

I also wish there was a way to paste an image into the assistant chat. For example, I wanted to share with it what I was seeing in the Chrome dev tools network tab and related errors. I thougth one possible work around was to save the image to a file and including it via the @file. But Gemini Assist says

"I cannot see the contents of the image error2.png. While I can see the file name, I am a text-based AI and do not have the ability to view images."

[verdverm]

I have a hard time getting past the cutesy cruft in my work tool, please remove it

[jochenkirstaetter]

It seems to read the first 2000 lines of a file only. And might be missing important information for changes requested.

[timothyfroehlich]

My biggest challenges, in order:

1. Does not automatically save sessions (looks like this will be addressed soon)
2. Lack of planning mode or anything to programmatically prevent the model from making changes before I'm ready. Yelling at it only works so well
3. Scrollback/chat spam. Especially because Gemini seems to be accidentally spitting out the contents of files it reads, or spitting out a changed version of a file before making the edit. This all adds up to it getting stuck scrolling through the buffer, especially when resizing the window. Core problem is the scrollback, the chat spam contributes to it.
4. Inability to control thinking budget or change models. Claude supports it, Codex supports it, Copilot has #think, but with Gemini you're locked to the model you picked at the beginning of the session and have no control over thinking.
5. Poor (but improving) management of allowed tool calls. Gemini doesn't support tool subcommands. Instead of always allowing git status and always requiring permission for git rm, git push, git switch, etc. you have to babysit each call because you can only give it full permissions for git binary itself. Same problem with gh or vercel or supabase.

[jasondsouza29]

It will loose all session when vertexai credentials refresh token expires and when try to login again.🙃anyone know the solution for this?