diff --git a/litellm/litellm_core_utils/prompt_templates/factory.py b/litellm/litellm_core_utils/prompt_templates/factory.py index 86cfbf70255..637a5bf117b 100644 --- a/litellm/litellm_core_utils/prompt_templates/factory.py +++ b/litellm/litellm_core_utils/prompt_templates/factory.py @@ -2292,11 +2292,23 @@ def _is_unsignable_thinking_block(block: object) -> bool: signature is null, empty, or missing (e.g. from an open-source reasoning model) is rejected with a 400 and must be dropped rather than blanked or repaired. `redacted_thinking` blocks carry no signature and are always kept. + + Anthropic also rejects a `thinking` block whose `thinking` text is empty or + whitespace-only ("each thinking block must contain thinking"), regardless of + signature. This shape reaches us when a caller replays a `thinking_blocks` + history item that originated from a non-Anthropic reasoning provider (e.g. an + OpenAI Responses-API turn with no summary text) through this Anthropic-shaped + request path (`/v1/chat/completions` -> anthropic/vertex_ai's claude models), + which is the same failure the Anthropic Responses-bridge adapter guards + against (see PR #36033) for its own separate content-block path. """ if not isinstance(block, dict) or block.get("type") != "thinking": return False signature: Final = block.get("signature") - return not (isinstance(signature, str) and len(signature) > 0) + if not (isinstance(signature, str) and len(signature) > 0): + return True + thinking_text: Final = block.get("thinking") + return not (isinstance(thinking_text, str) and len(thinking_text.strip()) > 0) def _drop_unsignable_thinking_blocks( diff --git a/tests/test_litellm/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_factory.py b/tests/test_litellm/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_factory.py index 6265779b90d..d2ee9207252 100644 --- a/tests/test_litellm/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_factory.py +++ b/tests/test_litellm/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_factory.py @@ -3578,3 +3578,186 @@ async def test_bedrock_converse_pdf_only_user_message_gets_text_block_async(): assert len(result) == 1 assert any("document" in block for block in result[0]["content"]) assert _text_blocks(result[0]) == [BEDROCK_DOCUMENT_PLACEHOLDER_TEXT] + + +def test_anthropic_messages_pt_drops_empty_but_signed_thinking_block(): + """ + Anthropic rejects a `thinking` block whose `thinking` text is empty, even + when it carries a valid-looking signature, with: + 400 messages.N.content.M.thinking: each thinking block must contain thinking + This shape is reachable via cross-provider replay of a `thinking_blocks` + history item (see PR #36033), so `_is_unsignable_thinking_block()` must + also check the thinking text, not just the signature. + """ + from litellm.litellm_core_utils.prompt_templates.factory import ( + anthropic_messages_pt, + ) + + messages = [ + {"role": "user", "content": "What's 2+2?"}, + { + "role": "assistant", + "content": "4", + "thinking_blocks": [ + { + "type": "thinking", + "thinking": "", + "signature": "sig_abc123_looks_valid", + } + ], + }, + ] + + result = anthropic_messages_pt( + messages=messages, + model="claude-sonnet-4-5-20250929", + llm_provider="anthropic", + ) + + assistant_msg = result[1] + assert isinstance(assistant_msg["content"], list) + content_types = [block.get("type") for block in assistant_msg["content"]] + assert "thinking" not in content_types, "empty-text thinking block must be dropped even though it has a signature" + + +def test_anthropic_messages_pt_keeps_non_empty_signed_thinking_block(): + """ + Regression: a real, non-empty, signed thinking block must still pass + through unchanged. + """ + from litellm.litellm_core_utils.prompt_templates.factory import ( + anthropic_messages_pt, + ) + + messages = [ + {"role": "user", "content": "What's 2+2?"}, + { + "role": "assistant", + "content": "4", + "thinking_blocks": [ + { + "type": "thinking", + "thinking": "Let me add these numbers together.", + "signature": "sig_abc123_looks_valid", + } + ], + }, + ] + + result = anthropic_messages_pt( + messages=messages, + model="claude-sonnet-4-5-20250929", + llm_provider="anthropic", + ) + + assistant_msg = result[1] + assert isinstance(assistant_msg["content"], list) + thinking_block = next((b for b in assistant_msg["content"] if b.get("type") == "thinking"), None) + assert thinking_block is not None, "non-empty signed thinking block must be kept" + assert thinking_block["thinking"] == "Let me add these numbers together." + assert thinking_block["signature"] == "sig_abc123_looks_valid" + + +def test_anthropic_messages_pt_keeps_redacted_thinking_block(): + """ + Regression: `redacted_thinking` blocks carry no signature and no plaintext + `thinking` field by design, and must always be kept regardless of the new + emptiness check (which only applies to `type == "thinking"` blocks). + """ + from litellm.litellm_core_utils.prompt_templates.factory import ( + anthropic_messages_pt, + ) + + messages = [ + {"role": "user", "content": "What's 2+2?"}, + { + "role": "assistant", + "content": "4", + "thinking_blocks": [ + { + "type": "redacted_thinking", + "data": "encrypted_opaque_blob", + } + ], + }, + ] + + result = anthropic_messages_pt( + messages=messages, + model="claude-sonnet-4-5-20250929", + llm_provider="anthropic", + ) + + assistant_msg = result[1] + assert isinstance(assistant_msg["content"], list) + content_types = [block.get("type") for block in assistant_msg["content"]] + assert "redacted_thinking" in content_types, "redacted_thinking blocks must always be kept" + + +def test_anthropic_messages_pt_drops_unsigned_thinking_block(): + """ + Regression (pre-existing behaviour): a thinking block with no signature + (or an empty/null one) must still be dropped, independent of whether the + thinking text is populated. + """ + from litellm.litellm_core_utils.prompt_templates.factory import ( + anthropic_messages_pt, + ) + + messages = [ + {"role": "user", "content": "What's 2+2?"}, + { + "role": "assistant", + "content": "4", + "thinking_blocks": [ + { + "type": "thinking", + "thinking": "Let me add these numbers together.", + "signature": "", + } + ], + }, + ] + + result = anthropic_messages_pt( + messages=messages, + model="claude-sonnet-4-5-20250929", + llm_provider="anthropic", + ) + + assistant_msg = result[1] + assert isinstance(assistant_msg["content"], list) + content_types = [block.get("type") for block in assistant_msg["content"]] + assert "thinking" not in content_types, "unsigned thinking block must still be dropped" + + +def test_is_unsignable_thinking_block_treats_whitespace_only_as_empty(): + """ + Edge case: a `thinking` field that is present but whitespace-only (e.g. + a single trailing newline forwarded from another provider's empty + reasoning summary) is functionally empty and Anthropic's API will still + reject it with "each thinking block must contain thinking". We treat it + the same as a fully empty string and drop the block. + + Note the sibling check ~30 lines below this function's definition + (the "don't pass empty text blocks" comment) uses a bare + `len(thinking_block) > 0`, which by itself would treat whitespace-only + text as non-empty. That sibling check is always combined with + `not _is_unsignable_thinking_block(m)` in an `and`, so this function's + stricter whitespace-aware check is still the one that decides whether a + whitespace-only block survives — the two checks don't disagree in + practice, but this function is intentionally the stricter of the two + since it is also called standalone (via `_drop_unsignable_thinking_blocks`) + without that extra `len(...) > 0` guard. + """ + from litellm.litellm_core_utils.prompt_templates.factory import ( + _is_unsignable_thinking_block, + ) + + whitespace_only_block = { + "type": "thinking", + "thinking": " \n\t ", + "signature": "sig_abc123_looks_valid", + } + + assert _is_unsignable_thinking_block(whitespace_only_block) is True