Thank you very much for the response. Creating a rule for this is definitely a good idea.
However, I would prefer not to create a rule explicitly for it because, first, it expands the prompt context, and second, it would be architecturally safer and better to simply have an option instead of having to run a prompt that is not 100% reliable every time, so that it can be disabled programmatically.
Below, I will share the complete correct invocation of all tools in case anyone is interested.
Tools
Tool Architecture
A callable tool identifier has two parts:
<namespace>.<tool>
Example:
functions.ReadFile
functions is the primary built-in tool namespace.ReadFile is the underlying operation.multi_tool_use is a separate orchestration namespace.- MCP integration is available through built-in
functions tools, but individual MCP-server tools are excluded from this document as requested.
Star (*) notation
functions.* is documentation shorthand for “all tools in the functions namespace.”
It is not a callable RPC name:
Correct: functions.ReadFile
Incorrect: functions.*
Markdown stars have a separate, purely visual purpose:
* Bullet item
**Bold text**
They do not affect tool execution or tool architecture.
Compatibility notice
This document changes no runtime behavior. However, old names such as functions.read_file and functions.run_terminal_cmd are no longer exposed under those names. Any automation calling them must migrate to the exact current identifiers below.
1. Functions
1.1 File, repository, and code inspection
functions.ReadFile — read a file, optionally by line range.functions.Glob — find files by glob pattern.functions.rg — search file contents using regex.functions.ReadLints — read IDE lint diagnostics.
functions.ReadFile
Read an entire file:
{
"path": "C:\\absolute\\path\\to\\base.md"
}
Read a portion of a file:
{
"path": "C:\\absolute\\path\\to\\base.md",
"offset": 1,
"limit": 250
}
Parameters:
path: string — required absolute file path.offset?: integer — optional 1-based starting line.limit?: integer — optional number of lines to return.
Iterative read example:
{
"path": "C:\\absolute\\path\\to\\base.md",
"offset": 1,
"limit": 250
}
Then continue from the next unread line:
{
"path": "C:\\absolute\\path\\to\\base.md",
"offset": 251,
"limit": 250
}
Unlike the legacy tool, the current response does not promise a total-line-count field. Continue until no more content is returned.
functions.Glob
{
"target_directory": "C:\\absolute\\path\\to\\repository",
"glob_pattern": "**/*.go"
}
target_directory?: string — search root.glob_pattern: string — required file pattern.
functions.rg
{
"pattern": "func\\s+Main",
"path": "C:\\absolute\\path\\to\\repository",
"glob": "*.go",
"output_mode": "content"
}
Common parameters:
pattern: string — required regular expression.path?: string — search root.glob?: string — restrict matching files.output_mode?: "content" | "files_with_matches" | "count"head_limit?: numberoffset?: number-A, -B, -C — context lines.i?: boolean — case-insensitive search.multiline?: boolean — enable multiline regex matching.
functions.ReadLints
{
"paths": [
"C:\\git\\test\\eslint.config.ts"
]
}
Parameters:
paths?: string[] — optional file or directory list.- If omitted, lint diagnostics are read for the entire workspace.
- No additional filtering parameters are exposed.
1.2 File and notebook changes
functions.ApplyPatch — create or modify one text file through a patch.functions.Delete — delete one file.functions.EditNotebook — edit or create a Jupyter notebook cell.
There is no direct functions.edit_file, functions.search_replace, or functions.write_file tool in the current API.
functions.ApplyPatch
ApplyPatch uses a patch payload rather than JSON arguments:
*** Begin Patch
*** Update File: C:\absolute\path\to\file.txt
@@
-old content
+new content
*** End Patch
Create a file:
*** Begin Patch
*** Add File: C:\absolute\path\to\new-file.txt
+Initial content
*** End Patch
functions.Delete
{
"path": "C:\\absolute\\path\\to\\obsolete-file.txt"
}
functions.EditNotebook
{
"target_notebook": "C:\\absolute\\path\\to\\analysis.ipynb",
"cell_idx": 0,
"is_new_cell": false,
"cell_language": "python",
"old_string": "old cell content",
"new_string": "new cell content"
}
Required fields:
target_notebookcell_idxis_new_cellcell_languageold_stringnew_string
1.3 Shell execution and task tracking
functions.Shell — run a shell command.functions.AwaitShell — inspect or wait for a background shell process.functions.TodoWrite — update the internal task list.
functions.Shell
{
"command": "go test ./...",
"working_directory": "C:\\absolute\\path\\to\\repository",
"description": "Run Go test suite"
}
functions.AwaitShell
{
"shell_id": "shell-id",
"block_until_ms": 30000
}
functions.TodoWrite
{
"merge": true,
"todos": [
{
"id": "inspect",
"content": "Inspect the repository",
"status": "in_progress"
},
{
"id": "verify",
"content": "Run verification",
"status": "pending"
}
]
}
1.4 Web, images, and user interaction
functions.WebSearch — search the web for current information.functions.WebFetch — retrieve and read a web page.functions.GenerateImage — generate an image from a prompt.functions.AskQuestion — ask the user structured multiple-choice questions.
functions.WebSearch
{
"search_term": "Cursor Agent tools documentation",
"explanation": "Verify current public documentation."
}
functions.WebFetch
{
"url": "https://docs.cursor.com/chat/tools"
}
functions.GenerateImage
{
"description": "A minimal flat vector application icon for a note-taking app.",
"filename": "notes-icon.png",
"aspect_ratio": "1:1"
}
functions.AskQuestion
{
"title": "Deployment choice",
"questions": [
{
"id": "environment",
"prompt": "Which environment should receive the release?",
"options": [
{
"id": "staging",
"label": "Staging"
},
{
"id": "production",
"label": "Production"
}
],
"allow_multiple": false
}
]
}
1.5 Modes and delegation
functions.SwitchMode — request a mode change.functions.Subagent — launch or resume a specialized agent.
functions.SwitchMode
{
"target_mode_id": "plan",
"explanation": "The task needs an implementation design before code changes."
}
Allowed mode IDs:
A user approval is required before changing modes.
functions.Subagent
This workspace has a policy that subagents may only be started when the user explicitly requests them.
Typical fields include:
{
"description": "Review API design",
"prompt": "Inspect the API design and report the important trade-offs.",
"subagent_type": "generalPurpose",
"run_in_background": false
}
1.6 MCP framework functions
These are built-in functions tools. They support MCP integrations, but this document intentionally excludes all individual MCP-server tools.
functions.GetMcpTools — discover MCP servers and their tool schemas.functions.FetchMcpResource — read an MCP resource.functions.CallMcpTool — invoke a discovered MCP tool.
Correct architectural sequence:
functions.GetMcpTools
↓
functions.CallMcpTool
Example discovery call:
{
"server": "example-server",
"toolName": "example-tool"
}
Example MCP invocation:
{
"server": "example-server",
"toolName": "example-tool",
"description": "Perform the requested operation.",
"arguments": {}
}
2. Orchestration
multi_tool_use.parallel
Runs independent developer-tool calls concurrently.
{
"tool_uses": [
{
"recipient_name": "functions.Glob",
"parameters": {
"target_directory": "C:\\absolute\\path\\to\\repository",
"glob_pattern": "**/*.go"
}
},
{
"recipient_name": "functions.rg",
"parameters": {
"pattern": "TODO",
"path": "C:\\absolute\\path\\to\\repository"
}
}
]
}
Only independent calls should be parallelized. A call that depends on another call’s result must run afterward.