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

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Mar 21, 2026 6:03am

@greptileai review

Merging this PR will not alter performance

✅ 16 untouched benchmarks

_{Comparing acebot712:add-promptguard-guardrail (bfe69af) with main (d8e4fc4)}

Greptile Summary

This PR adds PromptGuard as a new first-class guardrail vendor in LiteLLM, following the established pattern used by the ~25 other guardrail integrations in the codebase. The implementation covers the full stack: Python CustomGuardrail subclass, Pydantic config model, auto-discovery registry, UI Guardrail Garden card with eval scores, and 40 mocked unit tests.

Previously flagged issues (HTTPX client allocation before credential validation, dict[str, Any] Python 3.8 incompatibility, missing request timeout, unhandled JSONDecodeError, and the redact path injecting a texts key when none was present) have all been addressed in follow-up commits.

Key changes:

• SupportedGuardrailIntegrations.PROMPTGUARD enum entry and PromptGuardConfigModel added to the LitellmParams union
• PromptGuardGuardrail.apply_guardrail correctly handles allow (pass-through), block (raise GuardrailRaisedException), and redact (update structured_messages and/or texts — guarded to avoid creating a texts key that didn't exist before)
• block_on_error defaults to True (fail-closed) and can be overridden via the PROMPTGUARD_BLOCK_ON_ERROR env var
• 10-second request timeout on the outbound POST /api/v1/guard call
• All 40 tests use mocked HTTP responses with no real network calls, satisfying the CI/CD constraint

Confidence Score: 4/5

• This PR is safe to merge — it adds a self-contained new integration with no changes to shared code paths beyond enum and type registration.
• All previously identified issues have been addressed in follow-up commits. The integration follows established patterns, tests are comprehensive and fully mocked, and the change surface is well-isolated. One point withheld because self.api_key retains an Optional[str] type annotation after the non-None guard in __init__, meaning downstream use in the HTTP header dict is technically unsound to the type system even though it's safe at runtime.
• No files require special attention — the implementation is straightforward and well-tested.

Sequence Diagram

sequenceDiagram
    participant Client
    participant LiteLLMProxy as LiteLLM Proxy
    participant PG as PromptGuardGuardrail
    participant PGAPI as PromptGuard API
    participant LLM as LLM Provider

    Client->>LiteLLMProxy: POST /v1/chat/completions
    LiteLLMProxy->>PG: apply_guardrail(inputs, "request")
    PG->>PG: Build payload (messages, direction="input", model?, images?)
    PG->>PGAPI: POST /api/v1/guard (X-API-Key, timeout=10s)
    alt API error
        PGAPI-->>PG: Exception
        alt block_on_error=True
            PG-->>LiteLLMProxy: raise exception
            LiteLLMProxy-->>Client: 500 error
        else block_on_error=False
            PG-->>LiteLLMProxy: return inputs unchanged
        end
    else decision = "allow"
        PGAPI-->>PG: {"decision": "allow"}
        PG-->>LiteLLMProxy: return inputs unchanged
        LiteLLMProxy->>LLM: Forward request
        LLM-->>LiteLLMProxy: LLM response
        LiteLLMProxy->>PG: apply_guardrail(inputs, "response")
        PG->>PGAPI: POST /api/v1/guard (direction="output")
        PGAPI-->>PG: {"decision": "allow"}
        PG-->>LiteLLMProxy: return inputs unchanged
        LiteLLMProxy-->>Client: 200 response
    else decision = "block"
        PGAPI-->>PG: {"decision": "block", "threat_type": "...", "confidence": ...}
        PG-->>LiteLLMProxy: raise GuardrailRaisedException
        LiteLLMProxy-->>Client: 400 Blocked by PromptGuard
    else decision = "redact"
        PGAPI-->>PG: {"decision": "redact", "redacted_messages": [...]}
        PG->>PG: Update structured_messages and/or texts in inputs
        PG-->>LiteLLMProxy: return modified inputs
        LiteLLMProxy->>LLM: Forward redacted request
        LLM-->>LiteLLMProxy: LLM response
        LiteLLMProxy-->>Client: 200 response
    end

_{Last reviewed commit: "Fix CI lint: black f..."}

All @greptileai review comments have been addressed:

@ishaan-jaff - would appreciate your review when you get a chance. This adds PromptGuard as a first-class guardrail provider (backend hook, UI garden card, docs page, 41 tests). Happy to address any further feedback.