PR #24368: fix(bedrock): sort assistant content blocks so text precedes toolUse

• Repository: BerriAI/litellm
• Author: Chesars
• State: closed | merged: True
• Link: https://github.com/BerriAI/litellm/pull/24368

Description (problem / solution / changelog)

Relevant issues

Fixes https://github.com/berriAI/litellm/issues/24361

Pre-Submission checklist

• I have Added testing in the tests/test_litellm/ directory, Adding at least 1 test is a hard requirement - see details
• My PR passes all unit tests on make test-unit
• My PR's scope is as isolated as possible, it only solves 1 specific problem
• I have requested a Greptile review by commenting @greptileai and received a Confidence Score of at least 4/5 before requesting a maintainer review

Type

🐛 Bug Fix

Changes

When the Responses API (/v1/responses) routes to Bedrock Converse, multi-turn tool calling fails on Claude Sonnet 4 and Haiku 3.5+ because toolUse content blocks appear before text blocks in assistant messages.

The Responses API converts function_call and message output items into separate consecutive assistant messages. The Bedrock Converse transformer merges these into one message, but preserves input order. If function_call came before message, the merged result is [toolUse, text] — Claude rejects this with:

"tool_use ids were found without tool_result blocks immediately after"

Add _sort_bedrock_assistant_content_blocks() that reorders content blocks within assistant messages: reasoningContent → text → toolUse. Applied in both sync and async Bedrock Converse transformation paths, after the existing deduplication step.

Tests added

• test_sort_bedrock_assistant_content_blocks_text_before_tooluse — direct unit test
• test_sort_bedrock_assistant_content_blocks_reasoning_first — reasoning ordering
• test_sort_bedrock_assistant_content_blocks_preserves_order_when_correct — no-op when correct
• test_bedrock_converse_sorts_text_before_tooluse_sync — integration test through sync path

Changed files

• litellm/litellm_core_utils/prompt_templates/factory.py (modified, +36/-0)
• tests/litellm_core_utils/test_bedrock_converse_dedup_factory.py (modified, +131/-0)

---

PR #25853: fix(bedrock): handle orphaned tool blocks caused by conversation compaction

• Repository: BerriAI/litellm
• Author: AlmogBaku
• State: open | merged: False
• Link: https://github.com/BerriAI/litellm/pull/25853

Description (problem / solution / changelog)

Summary

Fixes Bedrock Converse API failures when clients like OpenAI Codex CLI compact long conversations. Compaction truncates history, leaving behind orphaned toolUse/toolResult blocks that Bedrock rejects.

Errors fixed:

• "tool_use ids were found without tool_result blocks"
• "The toolConfig field must be defined when using toolUse and toolResult content blocks."
• "tool_choice.type: Field required" (follow-on from injected toolConfig missing toolChoice)

Related issues: #25669, #24361

Changes

All fixes are gated on modify_params=True and follow existing LiteLLM patterns.

1. litellm/utils.py — has_tool_call_blocks()

Extended to detect Anthropic-format content blocks (type="tool_use" / type="tool_result") and role="tool" messages, in addition to the existing OpenAI tool_calls array check. Previously, when a Responses API request arrived with Anthropic-format tool content, the function returned False and _transform_request_helper skipped injecting toolConfig.

2. litellm/litellm_core_utils/prompt_templates/factory.py

Applied to both _bedrock_converse_messages_pt (sync) and _bedrock_converse_messages_pt_async (async):

a. Sort assistant content blocks (inside loop, after deduplication): Ensures order reasoningContent → text → toolUse within each assistant message. Bedrock rejects requests where text blocks appear after toolUse blocks, as it interprets them as breaking the toolUse→toolResult pairing.

b. Tool pairing sanitization (post-loop):

• Pass 1: For each assistant message with orphaned toolUse blocks (no matching toolResult), collect all orphaned IDs in order, then inject dummy toolResult blocks as a single batch into the next user message (or a newly inserted user message). Batch injection is important — injecting one at a time via prepend reverses the order, and Bedrock requires toolResult order to match toolUse order.
• Pass 2: Remove toolResult blocks whose toolUseId has no matching toolUse in any assistant message. If a user message becomes empty, replace its content with [BedrockContentBlock(text=" ")] (single space — Bedrock rejects empty strings).

3. litellm/llms/bedrock/chat/converse_transformation.py

Applied to both _transform_request (sync) and _async_transform_request (async):

After building bedrock_messages, if tool blocks are now present (injected by fix #2) but toolConfig is absent (because the original OpenAI messages had no tool blocks, so _transform_request_helper never added it), inject a dummy toolConfig with:

• A single ToolBlock(toolSpec=...) with empty input schema
• toolChoice=ToolChoiceValuesBlock(auto={}) — required by Bedrock/Claude; omitting it causes "tool_choice.type: Field required"

Test plan

• Verify no regression on normal tool-calling requests (tools present in original messages)
• Verify compacted conversations with orphaned toolUse no longer hit "tool_use ids were found without tool_result blocks"
• Verify compacted conversations with orphaned toolResult have them silently removed
• Verify toolConfig is injected when tool blocks exist after compaction sanitization
• Verify assistant messages with multiple parallel tool calls inject toolResult blocks in correct order

🤖 Generated with Claude Code

Changed files

• litellm/litellm_core_utils/prompt_templates/factory.py (modified, +457/-814)
• litellm/llms/bedrock/chat/converse_transformation.py (modified, +60/-0)
• litellm/utils.py (modified, +15/-0)
• tests/test_litellm/llms/bedrock/chat/test_converse_transformation.py (modified, +260/-0)