feat(bedrock): skip dummy user continue for assistant prefix prefill

[Sameerlite]

Summary

Bedrock Converse message setup used to append (or prepend) the default user continue message (Please continue.) whenever modify_params was true and an assistant message sat at the start or end of the transcript. That breaks OpenAI-style assistant prefill (prefix: true), where the partial assistant turn must remain last.

Fixes #17990
Fixes #24158
Fixes LIT-2310

Change

• In BedrockConverseMessagesProcessor._initial_message_setup, skip injecting the dummy user turn when the boundary assistant message has prefix: true.
• Covers both leading and trailing assistant turns.

Tests

• test_bedrock_converse_trailing_prefix_assistant_skips_user_continue
• test_bedrock_converse_leading_prefix_assistant_skips_user_continue

# LiteLLM settings
litellm_settings:
  set_verbose: false
  drop_params: false
  modify_params: true

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Apr 9, 2026 11:23am

[codspeed-hq]

Merging this PR will not alter performance

✅ 16 untouched benchmarks

---

_{Comparing litellm_bedrock-prefix-skip-dummy-user (86c1b88) with main (97f722f)}

[greptile-apps]

Greptile Summary

This PR fixes BedrockConverseMessagesProcessor._initial_message_setup to skip injecting a dummy "Please continue." user turn when the boundary assistant message carries prefix: true (OpenAI-style assistant prefill). Both the leading and trailing assistant cases are covered, and two new focused unit tests confirm the expected behavior.

Confidence Score: 5/5

This PR is safe to merge — it makes a targeted, well-tested fix to the Bedrock message preprocessing path.

The change is a two-line guard added to each of two symmetric code paths, both properly covered by new mock-only unit tests. No regressions are introduced to the default (non-prefix) behavior, and no custom rules are violated.

No files require special attention.

Vulnerabilities

No security concerns identified.

Important Files Changed

Filename	Overview
litellm/litellm_core_utils/prompt_templates/factory.py	Added prefix guard to both leading and trailing assistant-message injection paths in _initial_message_setup; logic is correct and minimal.
tests/test_litellm/llms/bedrock/chat/test_converse_transformation.py	Adds two new mock-only unit tests verifying that trailing and leading prefix: true assistant messages do not trigger dummy-user injection; tests are correct and do not weaken existing coverage.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[_initial_message_setup] --> B{messages empty?}
    B -->|yes| C[Inject continue message]
    B -->|no| D{first msg is assistant?}
    D -->|no| F{last msg is assistant?}
    D -->|yes| E{prefix == true?}
    E -->|yes| F
    E -->|no| G[Prepend user continue message]
    G --> F
    F -->|no| Z[Return messages]
    F -->|yes| H{prefix == true?}
    H -->|yes| Z
    H -->|no| I[Append user continue message]
    I --> Z

Loading

_{Reviews (1): Last reviewed commit: "feat(bedrock): skip dummy user continue ..." | Re-trigger Greptile}