Where does the bug appear (feature/product)?
Cursor for iOS
Describe the Bug
Title: Product-injected git/PR/browser user_rules cannot be disabled; in-app Send Feedback also fails; Cloud Agent gave a dead Bug Reports URL; agent would not emit unformatted copy text; forum blocks new-user bug posts over a 2-link cap
ASK (most important, please answer this):
I need help and concrete steps to remove the extra injected text from Agent prompts. Not an explanation of where it is. A procedure I can follow to turn it off. If Cursor staff say it cannot be removed, say that in one sentence and who can change it.
Core issue (the main problem is not that the model can make a mistake):
no user control to exclude injected prompt blocks the user does not want
agent first invented a Settings delete path instead of stating there is no off switch
Cursor Send Feedback (Bug) would not submit this report
a Cloud Agent then gave a wrong forum URL, so the fallback channel also failed
the same Cloud Agent was asked for the report as copyable plain text with no formatting and instead returned chat-rendered formatted text (section labels, bullets, numbered lists)
the official forum then rejected the post from a new user with: An error occurred: Sorry, new users can only put 2 links in a post.
the same error appeared again after the report was rewritten to contain only two https URLs; Discourse still blocked it, so the cap also counts autolinked hostnames (for example cursor [dot] com and gbg [dot] bg)
Repro:
Step 1. User asked (local chat) how to remove committing-changes-with-git and creating-pull-requests from the agent prompt so they apply only when the project uses git; later the same for the browser-verification user_rule; then asked for this ticket: unhappy because those extras cannot be turned off and that wastes money and time.
Step 2. Agent (Cursor Grok 4.6) listed those blocks in this chat system prompt under user_rules, next to the real User Rule “Senior web engineer”.
Step 3. Agent read the user Cursor rules mdc files (12 files). Those git/PR/browser texts are not there.
Step 4. Agent grepped the user Cursor dir for committing-changes-with-git. No rule file. Told the user to delete the blocks from Settings, Rules, User Rules anyway.
Step 5. User sent a Rules UI screenshot: only User Rule is Senior web engineer. User: “кажи как да ги махна!!! Пет пъти те питам”
Step 6. Agent searched Cursor docs/forum. Staff (Composer 2.5 threads) said those blocks are built into the model prompt. This chat is already Grok 4.6 and still has git/PR/browser in user_rules.
Step 7. Agent grepped /Applications/Cursor.app: no match. Grepped Application Support/Cursor: hits only in agent logs and old canvases, not a source file. Grepped User/settings.json: no match for those blocks or “Senior web engineer”. Grepped the user Cursor rules dir for the three names: no match.
Step 8. Browser automation was already OFF. The browser-verification user_rule was still in this prompt.
Step 9. User opened Cursor Send Feedback, chose Bug (“Something is off”), pasted this report, clicked Send. UI showed red text: “Couldn’t send. Try again.” Screenshot attached in chat.
Step 10. Because Send Feedback failed, user started a Cloud Agent (Cursor Grok 4.6, run id bc-f4a79ed8-761c-47f9-861b-43b23e765a76, operator email dusty [at] gbg [dot] bg) and said they only wanted the report sent. The Cloud Agent did not send it (no such tool). It then told the user to post on forum path /c/support/bug-reports . That page does not exist. User screenshot: “Oops! That page doesn’t exist or is private.” The live category slug is bug-report (singular), category id 6. User: “и ти лъжеш, както локалния продукт!!!”
Step 11. User asked the Cloud Agent to add the wrong-link error into the report and then give the full report for copying, with no markdown. The agent returned a chat message that still used list markup (dash bullets and numbered items). Cursor rendered it as formatted HTML (bold-looking section titles, bullet list, numbered list). User screenshot of that reply, with formatted Core issue bullets and Repro numbered list, not plain copy text. User then had to ask again: they wanted unformatted copy text and instead got formatted output.
Step 12. User logged into the forum as a new user and tried to post this report. The forum showed a dialog: An error occurred: Sorry, new users can only put 2 links in a post. Screenshot attached.
Step 13. Cloud Agent rewrote the body to two https URLs only (the Cloud Agent run page and the dead forum path). User pasted that version. Same dialog again: An error occurred: Sorry, new users can only put 2 links in a post. Screenshot attached. So a new user still cannot file this bug on the official backup channel.
Expected:
Concrete instructions from Cursor on how I can strip those blocks from prompts. A way to turn off injected user_rules (git commit, PR, browser verification). Send Feedback must actually send a Bug report. If an agent cannot send the report, it must not invent a broken fallback URL. If the user asks for copyable text with no formatting, the agent must return plain text that copies as plain text, not a chat-rendered formatted document. A new forum user must be able to file a Bug Report after in-app Send Feedback already failed. A 2-link cap must not block the post when the report only names run ids and path slugs, and must not keep failing after the user already cut the text down to two URLs.
Actual:
The blocks are in every Agent prompt this user checked. Not in Settings, Rules, not in user Cursor rules files, not in Cursor.app, not in local settings.json. No file to delete. Grok 4.6 still injects them. Send Feedback, Bug failed with “Couldn’t send. Try again.” Cloud Agent then gave a 404/private forum path (/c/support/bug-reports instead of /c/support/bug-report/6), so the documented backup channel also failed until the user proved the link was wrong. When asked for the report as copy text with no markdown, the Cloud Agent still emitted formatted chat content (bullets, numbered lists, styled section labels) instead of unformatted text. When the user tried to post, Discourse blocked a new user for having more than 2 links. The same block happened again on a two-URL rewrite.
Notes:
Operator: Cursor Grok 4.6 (local Agent plus this Cloud Agent). User quote: “не мога да изключа ненужни за мен неща от промпта и така губя пари и време.” Do not tell the user to wipe state.vscdb. This agent does not know why Send Feedback failed (size, network, or other). The wrong forum slug was invented in this Cloud Agent session after the user asked only to send the report. The unformatted-copy request was explicit (“дай ми го във вид за копиране, без маркдаун”) and was not honored in the next reply. This paste contains no URLs and no hostnames, so a new forum user can submit it. Run id to look up: bc-f4a79ed8-761c-47f9-861b-43b23e765a76
Steps to Reproduce
Step 1. User asked (local chat) how to remove committing-changes-with-git and creating-pull-requests from the agent prompt so they apply only when the project uses git; later the same for the browser-verification user_rule; then asked for this ticket: unhappy because those extras cannot be turned off and that wastes money and time.
Step 2. Agent (Cursor Grok 4.6) listed those blocks in this chat system prompt under user_rules, next to the real User Rule “Senior web engineer”.
Step 3. Agent read the user Cursor rules mdc files (12 files). Those git/PR/browser texts are not there.
Step 4. Agent grepped the user Cursor dir for committing-changes-with-git. No rule file. Told the user to delete the blocks from Settings, Rules, User Rules anyway.
Step 5. User sent a Rules UI screenshot: only User Rule is Senior web engineer. User: “кажи как да ги махна!!! Пет пъти те питам”
Step 6. Agent searched Cursor docs/forum. Staff (Composer 2.5 threads) said those blocks are built into the model prompt. This chat is already Grok 4.6 and still has git/PR/browser in user_rules.
Step 7. Agent grepped /Applications/Cursor.app: no match. Grepped Application Support/Cursor: hits only in agent logs and old canvases, not a source file. Grepped User/settings.json: no match for those blocks or “Senior web engineer”. Grepped the user Cursor rules dir for the three names: no match.
Step 8. Browser automation was already OFF. The browser-verification user_rule was still in this prompt.
Step 9. User opened Cursor Send Feedback, chose Bug (“Something is off”), pasted this report, clicked Send. UI showed red text: “Couldn’t send. Try again.” Screenshot attached in chat.
Step 10. Because Send Feedback failed, user started a Cloud Agent (Cursor Grok 4.6, run id bc-f4a79ed8-761c-47f9-861b-43b23e765a76, operator email dusty [at] gbg [dot] bg) and said they only wanted the report sent. The Cloud Agent did not send it (no such tool). It then told the user to post on forum path /c/support/bug-reports . That page does not exist. User screenshot: “Oops! That page doesn’t exist or is private.” The live category slug is bug-report (singular), category id 6. User: “и ти лъжеш, както локалния продукт!!!”
Step 11. User asked the Cloud Agent to add the wrong-link error into the report and then give the full report for copying, with no markdown. The agent returned a chat message that still used list markup (dash bullets and numbered items). Cursor rendered it as formatted HTML (bold-looking section titles, bullet list, numbered list). User screenshot of that reply, with formatted Core issue bullets and Repro numbered list, not plain copy text. User then had to ask again: they wanted unformatted copy text and instead got formatted output.
Step 12. User logged into the forum as a new user and tried to post this report. The forum showed a dialog: An error occurred: Sorry, new users can only put 2 links in a post. Screenshot attached.
Step 13. Cloud Agent rewrote the body to two https URLs only (the Cloud Agent run page and the dead forum path). User pasted that version. Same dialog again: An error occurred: Sorry, new users can only put 2 links in a post. Screenshot attached. So a new user still cannot file this bug on the official backup channel.
Expected Behavior
Expected:
Concrete instructions from Cursor on how I can strip those blocks from prompts. A way to turn off injected user_rules (git commit, PR, browser verification). Send Feedback must actually send a Bug report. If an agent cannot send the report, it must not invent a broken fallback URL. If the user asks for copyable text with no formatting, the agent must return plain text that copies as plain text, not a chat-rendered formatted document. A new forum user must be able to file a Bug Report after in-app Send Feedback already failed. A 2-link cap must not block the post when the report only names run ids and path slugs, and must not keep failing after the user already cut the text down to two URLs.
Actual:
The blocks are in every Agent prompt this user checked. Not in Settings, Rules, not in user Cursor rules files, not in Cursor.app, not in local settings.json. No file to delete. Grok 4.6 still injects them. Send Feedback, Bug failed with “Couldn’t send. Try again.” Cloud Agent then gave a 404/private forum path (/c/support/bug-reports instead of /c/support/bug-report/6), so the documented backup channel also failed until the user proved the link was wrong. When asked for the report as copy text with no markdown, the Cloud Agent still emitted formatted chat content (bullets, numbered lists, styled section labels) instead of unformatted text. When the user tried to post, Discourse blocked a new user for having more than 2 links. The same block happened again on a two-URL rewrite.
Operating System
MacOS
Version Information
Version: 3.20.21
VS Code Extension API: 1.128.0
Commit: f09fca384ceca23f7bf21f9c23655b162641d740
Date: 2026-09-13T19:02:51.781Z
Layout: IDE
Build Type: Stable
Release Track: Default
Electron: 42.10.0
Chromium: 148.0.7778.280
Node.js: 24.18.1
V8: 14.8.178.38-electron.0
xterm.js: 6.1.0-beta.291
OS: Darwin arm64 23.1.0
Does this stop you from using Cursor
No - Cursor works, but with this issue