Litellm oss staging 04 08 2026

[krrish-berri-2]

Relevant issues

Pre-Submission checklist

Please complete all items before asking a LiteLLM maintainer to review your PR

• 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

Delays in PR merge?

If you're seeing a delay in your PR being merged, ping the LiteLLM Team on Slack (#pr-review).

CI (LiteLLM team)

CI status guideline:

• 50-55 passing tests: main is stable with minor issues.
• 45-49 passing tests: acceptable but needs attention
• <= 40 passing tests: unstable; be careful with your merges and assess the risk.

• Branch creation CI run
  Link:

• CI run for the last commit
  Link:

• Merge / cherry-pick CI run
  Links:

Type

🆕 New Feature
🐛 Bug Fix
🧹 Refactoring
📖 Documentation
🚄 Infrastructure
✅ Test

Changes

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Apr 13, 2026 3:19am

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
6 out of 8 committers have signed the CLA.

✅ michelligabriele
✅ avarga1
✅ silencedoctor
✅ milan-berri
✅ Sameerlite
✅ acebot712
❌ abhyudayareddy
❌ krrish-berri-2
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[codspeed-hq]

Merging this PR will not alter performance

✅ 16 untouched benchmarks

---

_{Comparing litellm_oss_staging_04_08_2026 (fa605d8) with main (5544803)}

[greptile-apps]

Greptile Summary

This PR adds a new PromptGuard guardrail integration (prompt injection detection, PII redaction, topic filtering) along with a Vertex AI streaming fix (normalising proto finish_reason names through map_finish_reason()), a DashScope cache_control preservation override, a public_key.pem leading-whitespace fix, and two new model-info capability flags (supports_none_reasoning_effort, supports_xhigh_reasoning_effort).

• Redact silent pass-through (promptguard.py lines 186–196): when PromptGuard returns decision: \"redact\" but omits redacted_messages, the original content is returned to the LLM with no log and no error — the intended redaction is silently skipped regardless of block_on_error.

Confidence Score: 4/5

Safe to merge after addressing the redact silent pass-through bug in PromptGuard

One P1 bug: when PromptGuard returns decision='redact' without a redacted_messages payload, the original (potentially PII-containing) content is forwarded to the LLM unmodified and no warning is emitted. All other changes — streaming fix, DashScope override, PEM fix, model data updates — are correct and well-tested.

litellm/proxy/guardrails/guardrail_hooks/promptguard/promptguard.py (redact decision silent pass-through)

Important Files Changed

Filename	Overview
litellm/proxy/guardrails/guardrail_hooks/promptguard/promptguard.py	New PromptGuard guardrail integration; silent pass-through bug when redact decision carries no redacted_messages
litellm/proxy/guardrails/guardrail_hooks/promptguard/init.py	Standard guardrail initializer and registry wiring for PromptGuard; looks correct
litellm/litellm_core_utils/streaming_handler.py	Bug fix: wraps raw Vertex AI proto enum finish_reason name through map_finish_reason() so tool_calls override works correctly
litellm/llms/dashscope/chat/transformation.py	Overrides remove_cache_control_flag_from_messages_and_tools to preserve cache_control for DashScope provider
litellm/proxy/auth/public_key.pem	Fixes leading whitespace on -----BEGIN PUBLIC KEY----- header that caused PEM load failure
litellm/types/proxy/guardrails/guardrail_hooks/promptguard.py	Defines Pydantic config model for PromptGuard with api_key, api_base, and block_on_error fields
tests/test_litellm/proxy/guardrails/guardrail_hooks/test_promptguard.py	Comprehensive mock-only test suite for PromptGuard guardrail; no real network calls

Sequence Diagram

sequenceDiagram
    participant Proxy
    participant PromptGuardGuardrail
    participant PromptGuardAPI
    participant LLM

    Proxy->>PromptGuardGuardrail: apply_guardrail(inputs, "request")
    PromptGuardGuardrail->>PromptGuardAPI: POST /api/v1/guard
    alt API error
        PromptGuardAPI-->>PromptGuardGuardrail: Exception
        alt block_on_error=True
            PromptGuardGuardrail-->>Proxy: GuardrailRaisedException
        else block_on_error=False
            PromptGuardGuardrail-->>Proxy: original inputs
        end
    else decision=allow
        PromptGuardAPI-->>PromptGuardGuardrail: decision:allow
        PromptGuardGuardrail-->>Proxy: inputs unchanged
        Proxy->>LLM: original messages
    else decision=block
        PromptGuardAPI-->>PromptGuardGuardrail: decision:block
        PromptGuardGuardrail-->>Proxy: GuardrailRaisedException
    else decision=redact with redacted_messages
        PromptGuardAPI-->>PromptGuardGuardrail: decision:redact + redacted_messages
        PromptGuardGuardrail-->>Proxy: inputs with redacted content
        Proxy->>LLM: redacted messages
    else decision=redact WITHOUT redacted_messages
        PromptGuardAPI-->>PromptGuardGuardrail: decision:redact only
        Note over PromptGuardGuardrail: Bug: original inputs returned silently
        PromptGuardGuardrail-->>Proxy: original inputs unredacted
        Proxy->>LLM: unredacted messages
    end

Loading

_{Reviews (4): Last reviewed commit: "Merge pull request #25616 from BerriAI/m..." | Re-trigger Greptile}

[greptile-apps]

Duplicate import from the same module

ChatCompletionToolParam and AllMessageValues are both imported from litellm.types.llms.openai in two separate statements. Consolidate them into one import per the project's style conventions.

Suggested change

	from litellm.types.llms.openai import ChatCompletionToolParam
	from litellm.secret_managers.main import get_secret_str
	from litellm.types.llms.openai import AllMessageValues
	from litellm.types.llms.openai import AllMessageValues, ChatCompletionToolParam

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]

Silent redaction bypass when redacted_messages is absent

When PromptGuard returns decision: "redact" but redacted_messages is null or missing, the if redacted: guard silently falls through and the original unredacted inputs are returned unchanged — the PII/sensitive content passes to the LLM with no log warning and no error. This undermines the security guarantee of the redact path regardless of block_on_error.

Suggested change

	if decision == "redact":
	redacted = result.get("redacted_messages")
	if redacted:
	if structured_messages:
	inputs["structured_messages"] = redacted
	if "texts" in inputs:
	extracted = self._extract_texts_from_messages(
	redacted,
	)
	if extracted:
	inputs["texts"] = extracted
	return inputs
	if decision == "redact":
	redacted = result.get("redacted_messages")
	if redacted:
	if structured_messages:
	inputs["structured_messages"] = redacted
	if "texts" in inputs:
	extracted = self._extract_texts_from_messages(
	redacted,
	)
	if extracted:
	inputs["texts"] = extracted
	else:
	verbose_proxy_logger.warning(
	"PromptGuard returned decision='redact' but no "
	"redacted_messages; original content will be used."
	)