Fix: Handle array-style content blocks in messagesToPrompt

[superdav42]

Summary

• OpenAI API supports both string and array content formats for message content. Some clients (e.g. OpenClaw, multi-modal clients) send content as an array of typed blocks: [{"type": "text", "text": "Hello"}]
• When content is an array, JavaScript's template literal interpolation converts it to [object Object] instead of the actual text, causing garbled prompts sent to Claude CLI
• Adds an extractContent() helper that handles both formats: returns strings as-is, extracts and joins text from array-style content blocks, and falls back to String() for any unexpected type
• Updates all three msg.content references in messagesToPrompt (system, user, and assistant cases) to use extractContent(msg.content)

Details

The OpenAI chat completions API defines content as either a string or an array of content parts:

// String format (already handled)
{ "role": "user", "content": "Hello" }

// Array format (was broken, now fixed)
{ "role": "user", "content": [{"type": "text", "text": "Hello"}] }

The previous code used msg.content directly in template literals and parts.push(), which works fine for strings but produces [object Object] for arrays.

The new extractContent() function:

function extractContent(content: string | Array<{type: string; text?: string}>): string {
  if (typeof content === "string") return content;
  if (Array.isArray(content)) {
    return content
      .filter((block) => block.type === "text")
      .map((block) => block.text ?? "")
      .join("\n");
  }
  return String(content ?? "");
}

Test plan

• Verify string content still works as before (no regression)
• Verify array content [{"type": "text", "text": "..."}] is correctly extracted
• Verify mixed content arrays with non-text blocks are filtered properly
• Verify empty/null content is handled gracefully

🤖 Generated with Claude Code