[bug] Provider thinking output is not shown in the playground in Claude Code

[mmabrouk]

When using Claude Code in the playground, the provider's thinking/reasoning output is not displayed anywhere in the conversation.

Observed behavior

• Start a conversation in the playground with Claude Code or Opus.
• The provider appears to produce a final response.
• No thinking/reasoning content is shown during or after the response.
• The final message is still displayed correctly.

Expected behavior

The playground should display the provider's thinking/reasoning output when it is available, while still displaying the final response.

Impact

Users cannot see whether the model is actively reasoning, which makes the interaction appear as though the provider is not thinking.

[linear-code]

AGE-3962

[dosubot]

This is a confirmed frontend parsing issue. The thinking blocks are present in the API response from Claude — the backend and LiteLLM pass them through unchanged [1] [2]. The problem is that the OSS playground's Message type defines content as a flat string [3], but Anthropic's extended thinking responses return content as an array of typed blocks:

{
  "content": [
    {"type": "thinking", "text": "reasoning here..."},
    {"type": "text", "text": "final answer here..."}
  ]
}

The createMessageFromSchema function extracts content directly without handling this array format [4], so thinking blocks are either stringified as raw JSON or silently dropped. PromptMessageConfig then renders it as a flat string [5].

Fix path:

1. Update the Message type to support content: string | ContentBlock[] where ContentBlock includes thinking, text, and tool_use types [3]
2. Update createMessageFromSchema to detect array content and preserve all block types including thinking [4]
3. Update PromptMessageConfig rendering to iterate content blocks and render thinking blocks with visual distinction (e.g., collapsible section with muted styling) [5]

Worth noting: this has already been solved in the agent playground codebase via recent PRs that add content block schema conformance [6] and explicit reasoning block handling [7]. Those implementations could serve as reference for porting the fix to the OSS playground.

To reply, just mention @dosu.

---

Share context across your team and agents. Try Dosu.