Message JSON is provider-neutral storage, not a provider wire format. Message, UserMessage, AIMessage, and SystemMessage use the embedded message codec; AIMessage and ToolResultMessage define explicit codecs so extra fields are not lost.

Wire shape

{
  "role": "assistant",
  "blocks": [
    {"Text": "answer", "type": "text"}
  ],
  "usage": {"InputTokens": 4, "OutputTokens": 2}
}

blocks is omitted when the slice is empty. AIMessage.Usage is omitted when nil but a present zero usage encodes as {}. ToolResultMessage always writes tool_use_id; is_error uses omitempty, so false is represented by absence and decodes back to false.

Round-trip example

package main

import (
	"encoding/json"
	"github.com/looprig/core/content"
)

func main() {
	original := content.ToolResultMessage{
		Message: content.Message{
			Role: content.RoleTool,
			Blocks: []content.Block{&content.TextBlock{Text: "done"}},
		},
		ToolUseID: "call-1",
		IsError:   true,
	}
	wire, err := json.Marshal(original)
	if err != nil {
		panic(err)
	}
	var restored content.ToolResultMessage
	if err := json.Unmarshal(wire, &restored); err != nil {
		panic(err)
	}
}

The codec uses MarshalBlocks and UnmarshalBlocks, so nested tool-result content remains tagged. AI usage is validated on both marshal and unmarshal, and stale blocks or usage are cleared before an AIMessage decode applies new data.

Proof

Source: core/content/message.go, core/content/block_json.go
Tests: core/content/message_json_test.go covers fixed points, usage presence, nested blocks, and preservation of tool IDs.