“Yep this is how I pass plans from cursor to codex, just save plan to workspace in cursor and prompt codex to go through md file and implement. Regarding broader context, I keep extensive documentation in project files (both docstrings in code and md files for subsystems), and always prompt agents to double-check docs and keep them updated with every change.”
“how do you pass the context between your cursor ide and your codex? are you creating [plan.md](http://plan.md) and telling other agents to look at it and execute?”
“Codex plus + cursor 20euro plan. Usually I can cover most of the tasks, the main point is using grok 4.6 for simpler tasks (as simple refactor, simple CRUD Api with given example etc..) and codex sol 5.6 on high/xhigh for harder tasks.”
“I am not sure I would take that option partly because I like the Cursor harness and also specifically with Claude Enterprise I assume staying within monthly limits would be quite hard.”
“for anyone that has a budget of 20$ don't overthink it cursor is your best choice even if u go with codex or Claude code you will not be able to use the best models Opus/fable or Astra/Sol you will be stuck using sonnet 5 and Luna which in my opinion are not far better than grok 4.6”
“been doing something similar but with just claude and cursor, the session handoff notes are the real unsung hero here, 10 lines can save you from staring at your screen for 20 minutes wondering what broke last night”
“Cursor's allowlists help a bit, but the real split is separate agent profiles with different tool sets, not one agent with a prompt that says "be careful." A dedicated permission layer between the agent and the tools would beat hand-maintaining those profiles, especially once anything can hit production APIs.”
“Sorry, new to cursor, but why all the hate for grok? I understand it’s not a better model than opus/astra but the grok first party limits are more than 3rd party models”
“the leftover that bit me was a Cursor workspace path landing in a Claude Code session, and the agent tried to cd into a Mac-only folder that was not on the Linux box. when you add provenance, are you hanging it as metadata beside each fact, or as a tagged prefix in the blob so older readers still degrade cleanly?”
“For the Cursor flow, have you found that sessionStart context plus preCompact checkpoints is enough in practice, or does the lack of per-prompt injection make the agent drift during longer tasks before it decides to call MCP?”
“Same here honestly. I dont even need an exact date, just some kind of timeline from Cursor would be nice. Feels weird that iOS got the app and Android users still have no idea when it’s coming 😅”
“I dropped Cursor a while back after the rate limiting/lack of transparency. Speaking from a position of obvious bias, I cannot imagine it is a better product than Codex or Claude Code nowadays as you're buying access directly from the model makers with an intense focus on product quality and direct integration.”
“My cursor is 99.9% accurate in most tasks now, from coding, to doco review to research to bloody figuring out which paints to mix for specific result lol”
“Cursor was approved pre-purchase by X AI so I’m using as long as I can, just with all grok models disabled. Sentiment seems to be these orgs don’t trust Elon to not train on their data even if he says he won’t, which has just been reinforced by OpenAI’s exit from cursor and their reasoning.”
“I've found Composer 2.5 is brilliant for beaded workflows, and very cheap, does drift on very large work (hence beads is perfect fit) and too eager to implement on planning tasks.”
“Yeah if it’s auto or grok then definitely cursor is best. My concern is only with other or specifically Claude models in Cursor usage credits vs 5hr limit in Claude code.”
“So my GB without asking, used my own cursor, obviously a remote command, so I think that’s what happened to you. My reset is coming in a week though, so I’ll kick off with GB only to confirm your results, as if true, GB is fired.”
“If it was genuinely one prompt, I'd stop retrying that conversation, save the conversation ID and timestamps, and open a support ticket with the usage page plus request logs. Then start a fresh chat with only the minimal context.”