[Bug]: turn_off_message_logging Does Not Redact Request Messages in proxy_server_request Field When Stored to Database

[RupertoXTI]

What happened?

Bug: turn_off_message_logging Does Not Redact Request Messages in proxy_server_request Field When Stored to Database

Description

When using turn_off_message_logging: true with store_prompts_in_spend_logs: true, the response is correctly redacted in spend logs, but the request messages remain unredacted. This creates a security/privacy issue where sensitive request data is stored in the database despite redaction being enabled.

Environment

• LiteLLM Version: main-stable (latest as of 2025-11-06)
• Database: PostgreSQL
• Deployment: Docker Compose

Configuration

model_list:
  - model_name: gpt-4.1-mini
    litellm_params:
      model: gpt-4.1-mini
      api_base: https://litellm.com
      api_key: sk-xxxxx

general_settings:
  store_model_in_db: true
  store_prompts_in_spend_logs: true

litellm_settings:
  turn_off_message_logging: true

Expected Behavior

When turn_off_message_logging: true is set, both request and response should be redacted before being stored in the LiteLLM_SpendLogs table.

Actual Behavior

• ✅ Response: Correctly shows {"text": "redacted-by-litellm"}
• ❌ Request: Full unredacted request messages are stored in the proxy_server_request field in metadata column

Reproduction Steps

1. Configure LiteLLM proxy with the config above and the docker-compose.yml found in the repository
2. Run docker-compose up
3. Make a request:

curl --location 'http://0.0.0.0:4000/chat/completions' \
  --header 'Content-Type: application/json' \
  --header 'Authorization: Bearer sk-1234' \
  --data '{
    "model": "gpt-4.1-mini",
    "messages": [
      {
        "role": "user",
        "content": "what llm are you"
      }
    ]
  }'

4. Navigate to the UI and find the logs
5. Find your request
6. Find that the log for the request shows the response redacted but the request fully present without any redaction

Relevant log output

Are you a ML Ops Team?

No

What LiteLLM version are you on ?

1.79.0-stable

Twitter / LinkedIn details

No response

[gsatur]

I can confirm similar behavior. I see that responses are redacted, but not requests. I'm using v1.77.5-stable

[piernikR]

Confirming this bug + additional context

I'm experiencing the exact same issue. Adding some additional observations that may help:
Environment:

• LiteLLM Proxy deployed via official Helm chart
• general_settings.store_prompts_in_spend_logs: true

What I tried:

1. x-litellm-enable-message-redaction: true header (per-request)
   Result: Response is redacted ("text": "redacted-by-litellm") in Request Logs UI, but the Request body still shows full messages[].content with PII.

2. metadata.turn_off_message_logging: true (via extra_body)
   Result: Same behavior—response redacted, request not.

3. no-log: true
   Result: This disables logging entirely, including spend/cost tracking—not acceptable for our use case.

Suggested fix

When turn_off_message_logging: true (global or dynamic) or x-litellm-enable-message-redaction: true (header) is set, the SpendLogs payload builder should also redact/suppress:

• proxy_server_request.body.messages
• response