SFT is hard to overcome. If the model wasn't trained to use anchored edit, prompting alone won't make it do so — at least not with the models I have access to. The 2/10 on ds-v4-flash required explicitly asking it to try the tool.
Maybe this changes with models that got [upto]-like patterns in their training data, or when local inference allows the kind of forced sampling antirez demonstrated. For cloud-based agents today, the standard old/new block edit remains the practical default.
{
"model": "qwen3.7-max",
"reasoning_effort": "xhigh",
"stream": true,
"stream_options": {
"include_usage": true
},
"tools": [
{
"type": "function",
"function": {
"name": "AskUserQuestion",
"description": "Ask user a question and wait for their response.\n\nUse this tool when you need user input during execution:\n- Uncertain technical decisions\n- Conflicts with previous instructions\n- Need for requirement clarification\n- Presenting options for user to choose\n\nArgs:\n question: The question to ask the user.\n agent: The agent instance (injected automatically).\n choices: Optional list of suggested choices. User can still provide\n their own answer freely.\n\nReturns:\n User's response as a string.",
"parameters": {
"type": "object",
"properties": {
"question": {
"type": "string"
},
"choices": {
"type": "array",
"items": {
"type": "string"
}
}
},
"required": [
"question",
"choices"
]
}
}
},
{
"type": "function",
"function": {
"name": "Bash",
"description": "Execute a shell command in a specific directory.\n\nArgs:\n command: The command to execute.\n timeout: Maximum wait time in seconds.\n\nReturns:\n Command output with exit code.",
"parameters": {
"type": "object",
"properties": {
"command": {
"type": "string"
},
"timeout": {
"type": "integer",
"default": 30
},
"purpose": {
"type": "string",
"description": "简要说明本次工具调用的目的，控制在20个字以内",
"default": ""
}
},
"required": [
"command",
"purpose"
]
}
}
},
{
"type": "function",
"function": {
"name": "BetterEdit",
"description": "Edit a file using path, old_block, and new_block. The old text must match exactly once in the file; otherwise the edit fails for safety.\n\nFor large replacements, prefer anchored old_block: write the first lines, then [upto], then the final lines.\nThe tool replaces everything from the head through the tail. If the head or tail is ambiguous, the edit fails.\n\nAfter [upto], always write unique final lines before closing old_block; never close old_block immediately after [upto].\nDo not use a generic tail anchor like:\n\n some_function() {\n ...\n[upto]\n }\n\nbecause the closing brace may match many functions. Instead include final lines that are unique near that function,\nfor example its last calculation and return line before the brace.\n\nExample anchored edit:\n\n old_block: \"static int parse(void) {\n int ok = 0;\n[upto]\n return ok;\n }\"\n new_block: \"static int parse(void) {\n return parse_impl();\n }\"\n\nTo insert text, use old_block set to an exact unique anchor and new_block set to that anchor plus the added text.\n\nWithout [upto], old_block must match exactly once.\n\nArgs:\n path: Target file path.\n old_block: Exact text to find. Use [upto] marker for anchored edit.\n new_block: Replacement text.",
"parameters": {
"type": "object",
"properties": {
"path": {
"type": "string"
},
"old_block": {
"type": "string"
},
"new_block": {
"type": "string"
}
},
"required": [
"path",
"old_block",
"new_block"
]
}
}
},
{
"type": "function",
"function": {
"name": "Glob",
"description": "Find files matching glob pattern.\n\nSupports glob patterns like \"**/*.py\" or \"src/**/*.ts\".\nReturns matching file paths sorted by modification time (newest first).\n\nArgs:\n pattern: The glob pattern to match files against (e.g., \"**/*.py\").\n path: The directory to search in. Defaults to current directory.\n respect_gitignore: Whether to respect ignore rules (.gitignore, .ignore, etc).\n Only effective in git repositories. Defaults to True.\n\nReturns:\n List of matching file paths, one per line.",
"parameters": {
"type": "object",
"properties": {
"pattern": {
"type": "string"
},
"path": {
"type": "string",
"default": "."
},
"respect_gitignore": {
"type": "boolean",
"default": true
}
},
"required": [
"pattern"
]
}
}
},
{
"type": "function",
"function": {
"name": "Grep",
"description": "Search file contents with regex pattern.\n\nArgs:\n pattern: The regular expression pattern to search for.\n path: Directory or file to search in. Defaults to current directory.\n glob: Glob pattern to filter files (e.g., \"*.py\", \"*.ts\"). Ignored if path is a file.\n output_mode: \"files_with_matches\" (default), \"content\", or \"count\".\n i: Case insensitive search.\n head_limit: Max results to return (default 100).\n respect_gitignore: Whether to respect ignore rules (.gitignore, .ignore, etc).\n Only effective in git repositories. Defaults to True.\n context: Number of lines to show before and after each match.\n Only works with output_mode=\"content\". Default is 0.\n\nReturns:\n - files_with_matches: file paths containing the pattern\n - content: file:line content for each match (with context if specified)\n - count: file path and match count",
"parameters": {
"type": "object",
"properties": {
"pattern": {
"type": "string"
},
"path": {
"type": "string",
"default": "."
},
"glob": {
"type": "string",
"default": "*"
},
"output_mode": {
"type": "string",
"default": "files_with_matches"
},
"i": {
"type": "boolean",
"default": false
},
"head_limit": {
"type": "integer",
"default": 100
},
"respect_gitignore": {
"type": "boolean",
"default": true
},
"context": {
"type": "integer",
"default": 0
}
},
"required": [
"pattern",
"i",
"context"
]
}
}
},
{
"type": "function",
"function": {
"name": "Read",
"description": "Read file segment. offset=0 is first line, negative counts from end.\n\nArgs:\n path: File to read.\n offset: Starting line (0-based). -1 = last line.\n limit: Max lines to return (capped at 1000).\n\nReturns:\n [file: PATH | lines START-END/TOTAL | ENCODING]\n Content...\n [... N more lines] # if truncated\n\nErrors: \"read_file: PATH: No such file|Is a directory|Permission denied|Binary file\"",
"parameters": {
"type": "object",
"properties": {
"path": {
"type": "string"
},
"offset": {
"type": "integer",
"default": 0
},
"limit": {
"type": "integer",
"default": 200
}
},
"required": [
"path",
"offset"
]
}
}
},
{
"type": "function",
"function": {
"name": "TodoWrite",
"description": "Update the todo list for the current session. Use proactively for complex multi-step tasks (3+ steps). Rules: keep exactly ONE task as in_progress at a time; mark completed IMMEDIATELY after finishing; remove irrelevant tasks; provide content (imperative) and optionally activeForm (present continuous).",
"parameters": {
"type": "object",
"properties": {
"todos": {
"type": "array",
"description": "The updated todo list (replaces the entire previous list)",
"items": {
"type": "object",
"properties": {
"content": {
"type": "string",
"description": "Task description in imperative form (e.g. 'Run tests')"
},
"status": {
"type": "string",
"enum": [
"pending",
"in_progress",
"completed"
],
"description": "Task state: pending = not started, in_progress = currently working, completed = done"
},
"activeForm": {
"type": "string",
"description": "Present continuous form (e.g. 'Running tests'). Optional — defaults to content if omitted."
}
},
"required": [
"content",
"status"
]
}
}
},
"required": [
"todos"
]
}
}
},
{
"type": "function",
"function": {
"name": "Write",
"description": "Write content to file (overwrite).\n\nArgs:\n path: Target file path.\n content: Content to write.\n\nReturns:\n Success with stats, or error message.",
"parameters": {
"type": "object",
"properties": {
"path": {
"type": "string"
},
"content": {
"type": "string"
}
},
"required": [
"path",
"content"
]
}
}
}
],
"enable_thinking": true
}
I followed up on @antirez's [upto] edit design. Here's what happened when I tried to bring it to my agent (powered cloud-based LLM).
Setup
Results
{ "tool_calls": { "ds-v4-flash:BetterEdit": { "times": 10, "with_upto_times": 2, "upto_error_times": 0, "error_times": 1 }, "qwen3.7-max:BetterEdit": { "times": 88, "with_upto_times": 0, "upto_error_times": 0, "error_times": 0 } } }Conclusion
SFT is hard to overcome. If the model wasn't trained to use anchored edit, prompting alone won't make it do so — at least not with the models I have access to. The 2/10 on ds-v4-flash required explicitly asking it to try the tool.
Maybe this changes with models that got [upto]-like patterns in their training data, or when local inference allows the kind of forced sampling antirez demonstrated. For cloud-based agents today, the standard old/new block edit remains the practical default.
Appendix
All Tool Call Metrics
Tool Schema