fix(responses-bridge): extract list-format system content into instructions

[jo-nike]

Summary

When system message content is a list of content blocks (e.g. [{"type": "text", "text": "..."}]) instead of a plain string, the responses API bridge in convert_chat_completion_messages_to_responses_api() was passing it through as a role: system message in the input items. APIs like ChatGPT Codex reject this with "System messages are not allowed".

This happens when requests arrive via the Anthropic /v1/messages adapter, which converts system prompts into list-format content blocks in the OpenAI chat completions format before the responses bridge processes them.

Fix

Extract text from list content blocks and concatenate into the instructions parameter, matching the existing behavior for string system content. The else branch that added list system content as input items is replaced with an elif isinstance(content, list) branch that joins text parts.

How to reproduce

1. Configure a ChatGPT Codex model (e.g. chatgpt/gpt-5.2-codex) with mode: "responses"
2. Send a request via the Anthropic /v1/messages endpoint with a system message
3. The Anthropic adapter converts it to chat completions format with list content blocks
4. The responses bridge passes the list content as role: system in input items
5. ChatGPT Codex rejects with "System messages are not allowed"

Changes

• litellm/completion_extras/litellm_responses_transformation/transformation.py: Replace the else branch (which appended list system content as input items) with an elif isinstance(content, list) branch that extracts text from content blocks into instructions

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
litellm	Error		Feb 14, 2026 6:47am

[CLAassistant]

All committers have signed the CLA.

[greptile-apps]

Greptile Overview

Greptile Summary

Fixes a bug where system messages with list-format content blocks (e.g. [{"type": "text", "text": "..."}]) were incorrectly passed through as role: system input items in the responses API bridge, causing "System messages are not allowed" errors on APIs like ChatGPT Codex. The fix extracts text from these content blocks and concatenates them into the instructions parameter, matching the existing behavior for string system content.

• Correctly handles the Anthropic /v1/messages adapter flow, which converts system prompts into list-format content blocks before the responses bridge processes them
• The extraction logic handles both {"type": "text", "text": "..."} dicts and plain strings within the list
• Missing tests: No unit tests are included for the new list-content system message handling path, and the existing test suite has no coverage for system message extraction at all
• Minor: The change removes the else fallback that previously handled unexpected content types, which are now silently ignored

Confidence Score: 4/5

• This PR is safe to merge — the logic change is correct and well-scoped, though it would benefit from test coverage before merging.
• The fix is a straightforward and correct change to a single method. The new elif isinstance(content, list) branch properly extracts text from content blocks into the instructions parameter, matching the existing string handling. The only concern is the lack of unit tests, which is important for a regression fix. The change is minimal and unlikely to cause side effects.
• No files require special attention beyond the missing test coverage for litellm/completion_extras/litellm_responses_transformation/transformation.py.

Important Files Changed

Filename	Overview
litellm/completion_extras/litellm_responses_transformation/transformation.py	Replaces the catch-all else branch for system messages with an explicit elif isinstance(content, list) that extracts text from content blocks into instructions. The logic is correct but lacks test coverage and silently drops unrecognized content types.

Flowchart

flowchart TD
    A[Anthropic /v1/messages request<br/>with system prompt] --> B[Anthropic Adapter<br/>_add_system_message_to_messages]
    B -->|string system| C[ChatCompletionSystemMessage<br/>content = string]
    B -->|list system blocks| D["ChatCompletionSystemMessage<br/>content = [{type: text, text: ...}]"]
    C --> E[Responses Bridge<br/>convert_chat_completion_messages_to_responses_api]
    D --> E
    E -->|isinstance str| F[instructions = content]
    E -->|isinstance list NEW| G["Extract text from blocks<br/>join into instructions"]
    E -->|else REMOVED| H["❌ Old: passed as role:system input item<br/>→ 'System messages not allowed' error"]
    F --> I[Responses API Request<br/>instructions parameter set]
    G --> I
    I --> J[API call succeeds]

Loading

_{Last reviewed commit: 19ef616}

[greptile-apps]

_{1 file reviewed, 2 comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

Missing test coverage for new branch

This PR adds handling for list-format system content, but there are no corresponding unit tests. The existing test file (tests/test_litellm/completion_extras/litellm_responses_transformation/test_completion_extras_litellm_responses_transformation_transformation.py) has no tests for system message handling at all — neither for string content nor for this new list-content path.

Please add a test case that verifies:

1. A system message with content: [{"type": "text", "text": "Hello"}, {"type": "text", "text": "World"}] produces instructions = "Hello World" and no system input items.
2. A system message with list content combined with another string system message properly concatenates into instructions.

This would help prevent regressions and satisfy the PR template requirement of "Add at least 1 test in tests/litellm/".

Context Used: Rule from dashboard - What: Ensure that any PR claiming to fix an issue includes evidence that the issue is resolved, such... (source)

_{Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!}

[greptile-apps]

Silently drops non-str/list system content

The old else branch handled any non-string content type by passing it through as an input item. The new code only handles str and list — any other content type (e.g., a dict or unexpected type) is now silently ignored. While this is unlikely to occur given the AllMessageValues typing, consider adding a final else clause with a warning log for defensive robustness:

Suggested change

	elif isinstance(content, list):
	# Extract text from content blocks (e.g. [{"type": "text", "text": "..."}])
	text_parts = []
	for block in content:
	if isinstance(block, dict) and block.get("type") == "text":
	text_parts.append(block.get("text", ""))
	elif isinstance(block, str):
	text_parts.append(block)
	extracted = " ".join(text_parts)
	if instructions:
	instructions = f"{instructions} {extracted}"
	else:
	instructions = extracted
	elif isinstance(content, list):
	# Extract text from content blocks (e.g. [{"type": "text", "text": "..."}])
	text_parts = []
	for block in content:
	if isinstance(block, dict) and block.get("type") == "text":
	text_parts.append(block.get("text", ""))
	elif isinstance(block, str):
	text_parts.append(block)
	extracted = " ".join(text_parts)
	if instructions:
	instructions = f"{instructions} {extracted}"
	else:
	instructions = extracted
	else:
	verbose_logger.warning(
	f"Unexpected system message content type: {type(content)}"
	)

_{Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!}

[jo-nike]

Note: All CI failures are pre-existing on main and unrelated to this change:

• lint: proxy/litellm_pre_call_utils.py (too many statements) and proxy/policy_engine/pipeline_executor.py (unused variable) — files not touched by this PR
• test (all): poetry.lock out of sync with pyproject.toml

The unit-test and validate-model-prices-json checks both pass.