ToolUseChunk is a partial tool call. Providers may split the ID, name, and JSON arguments over several chunks and may interleave several calls.

API surface

type ToolUseChunk struct {
	Index     int
	ID        string
	Name      string
	InputJSON string
}

Use streamaccumulator.ToolUses rather than indexing a slice with Index:

var calls streamaccumulator.ToolUses
calls.Add(&content.ToolUseChunk{Index: 0, ID: "call-1", Name: "weather", InputJSON: `{"city":`})
calls.Add(&content.ToolUseChunk{Index: 0, InputJSON: `"Boston"}`})
for _, call := range calls.Blocks() {
	// Parse and validate call.Input only after the complete stream.
	fmt.Println(call.ID, call.Name, string(call.Input))
}

The accumulator uses a map keyed by Index, so negative or very large provider values cannot panic or allocate an unbounded slice. It preserves the last non-empty ID/name and emits complete blocks sorted in ascending index order. The raw concatenated JSON is not validated by core.

Tool loop boundary

Render fragments if desired, but execute only after the stream’s representation is complete and the caller has validated the tool name and argument schema. Then append a correlated ToolResultMessage.

Proof

Source: core/content/chunk.go, core/content/streamaccumulator/streamaccumulator.go
Tests: core/content/streamaccumulator/streamaccumulator_test.go covers interleaving, late IDs, empty fragments, and hostile indexes.