feat(guardrails): optional skip system message in unified guardrail inputs

[Sameerlite]

Summary

Adds optional exclusion of role: system messages from unified guardrail evaluation inputs (flattened texts and structured_messages) on OpenAI Chat Completions and Anthropic Messages, while leaving the request payload to the LLM unchanged.

Fixes LIT-2382

Configuration

• Global: litellm_settings.skip_system_message_in_guardrail (also sets litellm.skip_system_message_in_guardrail).
• Per guardrail: litellm_params.skip_system_message_in_guardrail (true / false / omit to inherit). Registry sets the attribute on the callback after init.

Implementation

• litellm/llms/base_llm/guardrail_translation/utils.py — resolution helper and OpenAI-shaped message filter.
• OpenAI and Anthropic chat guardrail translation handlers apply the flag when building inputs.
• Docs: guardrails quick start + litellm_settings table in config reference.

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Apr 10, 2026 7:00pm

[codspeed-hq]

Merging this PR will not alter performance

✅ 16 untouched benchmarks

---

_{Comparing litellm_skip-system-message-unified-guardrails (2abfaa5) with main (d0e347a)}

[greptile-apps]

Greptile Summary

This PR adds an optional skip_system_message_in_guardrail flag that excludes role: system messages from unified guardrail evaluation inputs (texts and structured_messages) while leaving the LLM payload intact. The feature supports both global (litellm_settings.skip_system_message_in_guardrail) and per-guardrail override semantics, and is surfaced in the dashboard UI as a tri-state selector (Yes / No / Use global default).

Confidence Score: 5/5

Safe to merge — feature is opt-in, backward compatible, and no P0/P1 issues found.

All remaining findings from prior rounds are P2 or lower (inline import comment, Anthropic test coverage gap). Core logic is sound: task_mappings continue to reference original message indices so write-back is correct, tri-state semantics are consistently implemented across both handlers and all UI forms, and update_in_memory_litellm_params propagates the new field correctly via Pydantic dict.

No files require special attention.

Important Files Changed

Filename	Overview
litellm/llms/base_llm/guardrail_translation/utils.py	Adds resolution helper and filter function; inline import litellm is justified to avoid circular import but warrants a comment (flagged in prior review).
litellm/llms/openai/chat/guardrail_translation/handler.py	Correctly applies skip flag to both texts_to_check (via _extract_inputs) and structured_messages; task_mappings remain consistent since msg_idx still refers to the original list.
litellm/llms/anthropic/chat/guardrail_translation/handler.py	Skip flag applied correctly to both _extract_input_text_and_images and structured_messages; Anthropic system is a top-level field (not a message), so the flag primarily filters the OpenAI-converted structured_messages.
litellm/proxy/guardrails/guardrail_registry.py	Sets skip_system_message_in_guardrail on the callback via setattr after init; update_in_memory_litellm_params (via vars() iteration) also covers the update path since the field is declared in LitellmParams.
litellm/types/guardrails.py	Adds skip_system_message_in_guardrail: Optional[bool] = None to BaseLitellmParams with clear docstring; tri-state semantics (None = inherit) are well-defined.
tests/test_litellm/proxy/guardrails/guardrail_hooks/unified_guardrails/test_unified_guardrail.py	Unit tests cover utility functions, per-guardrail override semantics, and the OpenAI handler integration; Anthropic handler path is not directly tested (flagged in prior review).
ui/litellm-dashboard/src/components/guardrails/guardrail_info_helpers.tsx	Adds SkipSystemMessageChoice type and two mapping helpers; logic is correct and properly handles the tri-state.
ui/litellm-dashboard/src/components/guardrails/add_guardrail_form.tsx	Adds skip_system_message_choice selector; omits key for inherit on create, which correctly results in None default on backend.
ui/litellm-dashboard/src/components/guardrails/edit_guardrail_form.tsx	Adds skip_system_message_choice to edit form; deletes key for inherit while guardrail_info.tsx sends null — both lead to None on backend for full-replace semantics.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["Guardrail pre-call hook"] --> B["effective_skip_system_message_for_guardrail(guardrail)"]
    B --> C{Per-guardrail\nskip set?}
    C -->|"skip_system_message_in_guardrail = True/False"| D["Use per-guardrail value"]
    C -->|"None (inherit)"| E["Read litellm.skip_system_message_in_guardrail"]
    D --> F{skip = True?}
    E --> F
    F -->|Yes| G["Filter system msgs from texts_to_check"]
    F -->|Yes| H["openai_messages_without_system(structured_messages)"]
    F -->|No| I["Include all messages in texts_to_check"]
    F -->|No| J["Pass full messages as structured_messages"]
    G --> K["GenericGuardrailAPIInputs\n(filtered texts + structured_messages)"]
    H --> K
    I --> K
    J --> K
    K --> L["guardrail.apply_guardrail(inputs)"]
    L --> M["Apply responses back via task_mappings\n(original message indices preserved)"]
    M --> N["LLM receives unchanged full messages"]

Loading

_{Reviews (3): Last reviewed commit: "fix(guardrails): type structured_message..." | Re-trigger Greptile}