kirodotdev
Context window percentage like on the CLI tools
Feature Description
I would be really useful to be able to see when you are close to getting the session to long message so that you don't ask Kiro to do something and then half way though the work is says the session is to long, unless you are internally completing the last request and then only after that saying the session is to long. If the latter is true then I would be nice to rather have a message from Kiro stating that it has completed thhe work but the session has come to its end
Use Case
When speaking to Kiro in a session
Additional Context
No response
After the auto-summarization update, the summarized context loses important information, causing the next chat to continue from incorrect points, not following standard rules and properly documented structure. This update was updated today, because before, at least it would stop and ask if you wanted to continue or not, and now it summarizes automatically. What is the difficulty in releasing the chat and unlocking the context?
Most IDEs are already like this: Windsurf, Trae, Cursor. Only Kiro, even Google's awful Antigravity, which is full of bugs, doesn't have this problem. Kiro has everything to surpass Cursor. After applying a method in more than 5 different IDEs, all using the same LLM, Kiro is one of those that had the best understanding of patterns, rules, protocols, and the method where the user controls the agent and the agent plans first, before executing, without using a plan tool, only with markdowns and a rigorous pattern structure to follow that locks the agent under my command. It worked beautifully for more than 4 days, performing end-to-end tasks that were already properly documented, programmed, and validated, and now it's horrible with this automatic context. Is there still a way to disable this?
you are correct, it's quite frustrating when the agent ask you to begin a new session and when you do, it loses context and starts from anywhere, before you can manage to get back on track, you are hit with session ended unexpectedly again and you have to have start a new session where the agent can't summarize context, i don't see how anyone can use this service if this remains, i want to believe they are working on a fix or allow users to downgrade to the previous working app
This is one of those "be careful what you wish for requests", came back after the holidays, updated my Kiro, and there it was, a nice little indicator of the total context used and an auto summarization and continuation feature. Which summarizes your conversation, and then it starts a new one with 20%+ context used, and you ask it to help you with a CDK-related issue, and it fills the context to 80% summarizes, and loses a bunch of the context and doesn't remember what it was in the middle of and breaks everything. Summarization at 80% is way too aggressive, and it feels like we have like 10% of the context length we had before I went on holiday mid-December.
We seriously need a fix for this new bug as a feature. Kiro needs to utilize the 80-90/95% context to attempt to complete the current task; it cannot simply hard-stop at 80% and try to summarize and automatically continue. All it is doing is breaking the code.
The best solution when comparing various other tasks, even Google's antigravity doesn't have a chat with context limiting. I mention antigravity because it was the last agentic IDE released in 2025, and it's far from being good. I see that the Kiro team has come a long way ahead; the Kiro IDE has been very good in almost everything. There are small bugs that start to make the workflow somewhat complicated, but the limitation of each chat is the biggest one I see. And the second is still the excessive use of context when the Power technology is used. For example, I created a Power just for MCPs, I created a persistent context MCP in Node.js, and what happens is that all more than 10 MCPs being managed dynamically still consume a lot of the total limit, 80% of the chat. If Power came as something to avoid overload, and what I notice is only the activation and deactivation of MCPs in a more intelligent way through keywords that activate them, then Power didn't live up to the promise that it would. It would be economical to manage the chat context...
In most IDEs, especially Cursor, the context of each session and chat is saved directly in the application's database, usually a hidden SQL database. Chats don't need a limiter, provided the IDE has an automatic tool that manages sessions, thus allowing each user the freedom to manage their own workload by starting new chats whenever they notice the conversation is lagging due to length or need. This would greatly help each user manage their own context. After all, the developer who wants to work with AI needs to be in control, not relying on the IDE's automatic orchestration, which often fails when working with codebases or large, complex projects with multiple phases and tasks.
The second point would be precisely that: orchestration. The entire market for agent-based development tools seems to be sitting on the fence; they don't know if they want to develop solutions for end users who won't be coding. With all due respect, in my opinion, they shouldn't even be considered developers, but simply end users who don't have and don't intend to have coding knowledge. Therefore, they should receive ready-made building blocks to play with and build as they please. This doesn't mean being the future or actually being a developer. Development is a vast body of knowledge ranging from front-end to backend, encompassing various aspects including dependency management, feature management, architecture planning, and all aspects of architectural design plus the visual design of the system. This alone, without going into too much depth, already demands a basic understanding of languages, which no AI (Air Intelligence Agency) today has the preparation, orchestration, or process for, beyond the logical interpretation that allows humans to make their own decisions. What all companies sell is the magnificent idea of autonomous coding agents, when in reality they deliver automated agents orchestrated through inference, using technologies like Transformers to try to create a coding agent that often doesn't respect a system of rules, protocols, has no workflow, no process. And this, unfortunately, means that in my case, as a user, I like to strictly follow a flow of every project: RESEARCH, PLANNING, DOCUMENTATION, GUIDES. Executives need to be involved first and then start executing the guidelines. Even with modular MD files, rules and agent profiles, user profiles, iteration manuals, available MCPS catalog, protocol guides with 4 types of protocols for use in each document and each phase according to tasks and complexities, many times the agents only read the document, I see the file system showing "file read," automatic indexing system there, and the model goes unnoticed. They know they received it and don't bother to open it, look at the context, absorb the information before even speaking or responding. This is an issue that all companies sell the tool as if it were ready, but it makes me realize how far this is from being truly ready for a market like this.
I then came to the conclusion that the orchestration itself lacks agentic hooks that validate, as if forcing the models to follow and use the structure or extension of files that dictate rules, as if there were an enforcement system in the document that doesn't treat something as optional but rather always follows everything that has that rule, and perhaps all documents with "apply always" at the top would start to have a different weight and action within the model, forcing it to follow the flow that the user controls with MD rule files or steering...
I hope that companies can stop focusing solely on the race for quantity or the next model to be launched, especially since it's already a given that most models are already at their training limits and there are no more resources to develop new models, having to wait until the second half of the year or perhaps the first half of next year when the new data centers are ready.