Bug: Gemma 4 tool parser breaks on nested arguments (arrays/objects)

The Gemma 4 tool call parser in mlx_vlm/tool_parsers/gemma4.py has two bugs that cause it to produce malformed output whenever a tool's arguments contain nested structures (arrays of objects, nested objects, etc.). This makes tools with complex schemas unusable.

Root Cause 1: Non-greedy regex truncates arguments

The regex used to extract tool call arguments:

_tool_call_regex = re.compile(r"call:(\w+)\{(.*?)\}", re.DOTALL)

(.*?) is non-greedy, so it matches up to the first } it finds. When arguments contain nested objects (e.g. [{key:val}]), the inner } terminates the match early, silently discarding everything after it.

Example — given this model output:

call:edit{path:<|"|>test.txt<|"|>,edits:[{newText:<|"|>orange<|"|>,oldText:<|"|>apple<|"|>}]}

The regex captures only:

path:<|"|>test.txt<|"|>,edits:[{newText:<|"|>orange<|"|>,oldText:<|"|>apple<|"|>

The trailing ]} (closing the array and outer object) is lost.

Root Cause 2: Flat key-value parser can't handle nested structures

Even if the regex captured the full string, the parser splits on : to find keys and on , to find value boundaries. It has no concept of nesting depth, so an array-of-objects value like [{newText:<|"|>orange<|"|>,oldText:<|"|>apple<|"|>}] gets split on the inner comma, producing two separate top-level keys instead of one nested value.

Observed behavior

For a tool call like:

call:edit{edits:[{newText:<|"|>orange<|"|>,oldText:<|"|>apple<|"|>}],path:<|"|>test.txt<|"|>}

The parser produces:

{
  "edits": "[{newText:<|\"|>orange<|\"|>",
  "oldText": "apple"
}

• edits gets a garbage string fragment instead of an array
• oldText is promoted to a top-level key (it should be nested inside edits[0])
• path is lost entirely (truncated by the regex)

Expected behavior

{
  "edits": [{"newText": "orange", "oldText": "apple"}],
  "path": "test.txt"
}

Reproduction

from mlx_vlm.tool_parsers.gemma4 import parse_tool_call

# Any tool call with nested objects/arrays will fail
text = 'call:edit{path:<|"|>test.txt<|"|>,edits:[{newText:<|"|>orange<|"|>,oldText:<|"|>apple<|"|>}]}'
result = parse_tool_call(text)
print(result)
# Produces: {'name': 'edit', 'arguments': {'edits': '[{newText:<|"|>orange<|"|>', 'oldText': 'apple'}}
# Missing: path, edits is malformed

Fix needed

The parser needs to be replaced with a recursive descent parser that:

1. Tracks brace/bracket nesting depth to find the correct outer } (or uses balanced-group matching)
2. Recursively parses nested {...} objects and [...] arrays instead of flat comma-splitting

Environment

• mlx-vlm version: 0.4.4 (commit 0ea9088)
• Parser file: mlx_vlm/tool_parsers/gemma4.py (introduced in commit 43b9b20)
• Affected: any tool with array or object parameters (very common in agentic/MCP use cases)