Tools with array/object parameters fail when LLM sends stringified JSON

[crottolo]

Bug Description

Tools with typed parameters (array, object, number, boolean) fail with Zod validation error when the LLM sends the parameter as a string instead of the proper type.

Error Examples

Case 1: Array as string (Todowrite)

⚙ Todowrite [todos=[{"id": "1", "content": "Database: Migration", "status": "completed", "priority": "high"}, ...]]

Error: The Todowrite tool was called with invalid arguments: [
  {
    "expected": "array",
    "code": "invalid_type",
    "path": ["todos"],
    "message": "Invalid input: expected array, received string"
  }
]

Case 2: Number as string (Bash timeout)

⚙ Bash [command=pnpm build 2>&1 | head -100, description=Build project to verify i18n fixes, timeout=180000]

Error: The Bash tool was called with invalid arguments: [
  {
    "expected": "number",
    "code": "invalid_type",
    "path": ["timeout"],
    "message": "Invalid input: expected number, received string"
  }
]

Root Cause

LLMs sometimes serialize typed parameters as strings:

• { "todos": "[{...}]" } instead of { "todos": [{...}] }
• { "timeout": "180000" } instead of { "timeout": 180000 }
• { "enabled": "true" } instead of { "enabled": true }

Affected Tools

Tool	Parameter	Type
Todowrite	todos	array
Question	questions	array
Bash	timeout	number
Webfetch	timeout	number
Websearch	numResults	number
LSP tools	line, character	number
Edit	replaceAll	boolean
Multiedit	replaceAll	boolean

Proposed Fix

Add automatic coercion in Tool.define():

1. If validation fails with type mismatch, attempt to coerce string values
2. For strings starting with [ or {: try JSON.parse()
3. For numeric strings: convert with Number()
4. For "true"/"false": convert to boolean
5. Retry validation with coerced values
6. Backward compatible: if coercion fails, throw original error

Environment

• OpenCode version: 1.1.8
• All providers affected (Anthropic, OpenAI, Google, etc.)

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• AI_InvalidToolInputError: Invalid input for tool todowrite #1373: AI_InvalidToolInputError: Invalid input for tool todowrite - Shows the exact same problem where LLM sends stringified JSON array instead of proper array format for the todos parameter
• [Bug] Question tool fails to parse JSON with French accented characters (UTF-8) #7484: Question tool fails to parse JSON with French accented characters - Related validation issue where array parameters are incorrectly parsed as strings

Feel free to ignore if your specific case is different from these.