sgl-project/sglang · error · ValueError

Anthropic redacted_thinking history is not supported

Error message

Anthropic redacted_thinking history is not supported

What it means

Raised while converting an Anthropic request to the internal chat-completion format: an assistant message in the conversation history contains a redacted_thinking block (Anthropic's encrypted, opaque thinking blocks), which SGLang cannot re-encode because it lacks Anthropic's encryption keys/context. The whole request is rejected rather than silently corrupting history.

Source

Thrown at python/sglang/srt/entrypoints/anthropic/serving.py:386

        def _convert_assistant_thinking_blocks(
            blocks: list[AnthropicContentBlock],
        ) -> tuple[Optional[str], Optional[str]]:
            """Reconstruct prior-turn thinking as ``(reasoning_content, text)``.

            At most one is set: encoders that frame the reasoning channel take
            it as ``reasoning_content``, everything else gets it re-wrapped and
            spliced into content.

            ``redacted_thinking`` carries encrypted bytes that no local
            parser can interpret, so we raise rather than silently drop it.
            On non-reasoning models (no detector configured) the rewrap is
            best-effort: we log a warning and drop the thinking text so a
            history echo doesn't 400 the whole request — the prior thinking
            is opaque context the model didn't need anyway.
            """
            if any(block.type == "redacted_thinking" for block in blocks):
                raise ValueError("Anthropic redacted_thinking history is not supported")

            thinking_parts = [
                block.thinking
                for block in blocks
                if block.type == "thinking" and block.thinking
            ]
            if not thinking_parts:
                return None, None

            reasoning_text = "\n".join(thinking_parts)
            if self.openai_serving_chat.supports_native_reasoning_history():
                return reasoning_text, None

            try:
                return None, self.openai_serving_chat.wrap_reasoning_history(
                    reasoning_text
                )
            except ValueError as e:

View on GitHub (pinned to 0132848349)

Solutions

1. Strip redacted_thinking blocks from assistant history before sending (the code already drops plain 'thinking' text best-effort; redacted blocks must be removed client-side)
2. Regenerate the history from plain text/transcript instead of raw Anthropic response objects
3. If using an Anthropic SDK passthrough of prior responses, filter content blocks to types text/tool_use only

Example fix

# before
messages = [{"role": "assistant", "content": prior_anthropic_response_content}]
# after
messages = [{
  "role": "assistant",
  "content": [b for b in prior_anthropic_response_content
               if b.get("type") not in ("redacted_thinking", "thinking")]
}]

SUPPORTED = {"text", "tool_use", "image", "document"}
for msg in messages:
    if msg.get("role") == "assistant" and isinstance(msg.get("content"), list):
        msg["content"] = [b for b in msg["content"] if b.get("type") in SUPPORTED]

def has_redacted_thinking(messages) -> bool:
    return any(b.get("type") == "redacted_thinking"
               for m in messages if isinstance(m.get("content"), list)
               for b in m["content"])

try: resp = client.messages.create(...)
except ValueError as e:
    if 'redacted_thinking' in str(e): strip_history_and_retry()

When it happens

Trigger: POST /v1/messages where messages[] contains an assistant turn whose content includes {"type": "redacted_thinking", "data": "..."} — typically captured from a previous real Anthropic API response that was replayed as history.

Common situations: Replaying or continuing conversations originally served by Anthropic's API (which emits redacted_thinking for some safety-flagged thinking); logging request/response pairs and feeding them back; agent frameworks that store raw Anthropic responses as history.

Related errors

• tool_choice references tool {tool_name!r} but it is not in t
• tool_choice={tc_type!r} requires at least one custom tool; a
• thinking.budget_tokens must be >= 1024 (got {})
• thinking.budget_tokens is not allowed when thinking.type is
• thinking.display is not allowed when thinking.type is 'disab

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/4e1f5cd469e2c521. Report an issue: GitHub.