I’ve been a Cursor user since the very beginning. As an engineer, this tool was a game-changer for me, boosting my productivity significantly on complex projects. I rely on advanced prompting with strict constraints and structured contexts in Composer 2. Everything was working perfectly until the most recent updates.
I am writing this because I’ve noticed a major regression that is now hindering my work. This isn’t a minor tweak; it feels like the core logic has collapsed.
Observed issues:
Context Loss: The tool seems unable to maintain the “n-1” message context. It focuses almost entirely on the latest prompt and loses the thread of the conversation immediately, even when I am well under my quota limits (under 15%).
Methodological Shift: Instead of direct code edits (which worked flawlessly before), Composer 2 has started generating unnecessary intermediary scripts to modify the programs. This adds useless complexity and introduces constant errors.
Execution Failure: There is a clear disconnect between instructions and output. The AI tends to digress or talk about the task instead of simply executing it, even with very precise framing. I ask for A, it does B.
Over-engineered Reasoning: Even for trivial tasks like changing a single word or fixing a typo, Composer 2 now triggers an exhaustive, doctoral-level “Chain of Thought” that lasts for minutes. It builds massive action plans for simple edits that should be instantaneous, wasting both time and resources.
My question to the team is simple:
You had a best-in-class “Chain of Thought” and context management system. What has changed in the recent pipeline or model integration to cause such a drop in reliability? I am now spending more time debugging the AI’s logic than I would coding manually.
As a professional user, I’m looking for technical clarity on this. We rely on Cursor for high-level engineering, and we need the tool to meet the standards of the service we subscribed to.
Hey, thanks for the detailed feedback. To turn this into an actionable report for the team and not just “Composer 2 got worse”, I need a few specific things:
Your Cursor version and OS Help > About or Cursor > About Cursor.
Request IDs for 1 to 2 recent chats where the symptoms are clearly visible, especially for “creates an intermediary script instead of a direct edit” and “loses n-1 context when under 15% quota”. You can get them like this: in the chat open the menu top right three dots > Copy Request ID. Note that if Privacy Mode is enabled, the Request ID may be unavailable or limited.
A short prompt example can be anonymized where Composer 2 went into over engineered chain of thought for a trivial edit. That makes it easier to compare against what should be happening.
On the general observations side, some complaints about Composer 2 instruction following and context loss in long sessions are being tracked internally, but without Request IDs I can’t tell if your case matches a known issue or if it’s something separate. I don’t have an ETA for fixes.
Once you send the details, we’ll look at the specific traces and I can link it to the right issue.
Root cause vs. Symptoms: The failure of your intelligence pipeline
Spare me the client support script. Requesting Request IDs or my OS version for a systemic failure of this magnitude is an insult to engineering. The problem is not the user; it is the root cause of an architecture that has lost its way.
Here is the reality that your degrees seem to be masking:
Value vs. Tokens: The core of our profession is value creation, not token generation. Multiplying intermediary scripts and 5-minute ‘Chain of Thought’ sessions for atomic edits is not a sign of power; it is an admission of inefficiency. You are confusing computational volume with relevant output.
Disregard for Occam’s Razor: Every unnecessary reasoning step for a trivial task is a mathematical disgrace. Intelligence is the art of reducing complexity to maximize action (negentropy). If your model requires a Rube Goldberg machine to fix a typo, it isn’t ‘thinking’. it is diffracting.
Violation of Information Theory (Shannon): Extracting signal from noise is the bedrock of computer science. Currently, your pipeline is an entropy generator. You have replaced the coherent sedimentation of information with algorithmic loquacity that dilutes all added value.
Failure of Coherence: Losing n-1 context while well within quota limits proves that your state persistence and logic are collapsing. Turing and Shannon would turn in their graves seeing a supposedly ‘intelligent’ system create this much disorder for so little meaning.
I do not need a patch. I need a tool that understands that engineering is the art of precision, not useless noise. As long as you remain in a logic of creating technical debt, patching the surface instead of fixing the root cause, it simply means you have understood nothing about your own field.
Got it. I’m not going to argue with you about information theory. That’s not the point right now.
The Request ID and version aren’t “client support script” stuff. Composer 2 behavior depends on the exact session, like the prompt, tool calls in the thread, context window state, client version, and the routing path the request went through. Without tracing, I can only agree or disagree with your observations in a rhetorical way, and that doesn’t help anyone. With a Request ID, I can open your real request and see what happened in the pipeline, like which tools were called, how much context was packed into the model, and where it switched into overly complex CoT instead of making a direct edit. That’s the root cause analysis you’re asking for.
We’ve noted your general feedback about Composer 2, like context handling in long sessions and overplanning on trivial edits. We’re getting similar reports, and the team understands the direction of the issue. I can’t share timelines for specific changes.
If you decide to share the Request ID from at least one representative session, we’ll take a detailed look and link it to the right internal ticket if it matches. Without that, it’ll stay at the level of general feedback.
Dean Rie, this is a textbook example of unhealthy toxicity.
It is the height of indecency: not only do you censor a critical discussion by hiding it from the forum because you are incapable of facing the core issue, but you take it upon yourself to reopen an exchange I explicitly closed, driven by nothing but pure ego and frustration.
I was clear: I stated I was closing this theater piece. Yet, your visceral need to have the last word has forced you to return with hollow, empty content, proving once again that you confuse administrative authority with technical competence.
You hide the discussion because you cannot counter a truth that inconveniences you. Reopening a closed topic just to soothe a bruised ego and misplaced frustration is a sign of absolute weakness. You refuse to debate an architecture you do not understand, retreating instead into the only thing you master: petty ticket management and enforced silence.
Your behavior is the mirror image of your system: excessive noise, useless procedures, and a complete inability to address the root of the problem. Keep your “last word” if it comforts your ego; the indecency of your conduct speaks for itself.