6 views
-/https://github.com/berriai/litellm/issues/29079
GitHub · issue

#29079 feat: Custom x- headers not captured in spend logs on Vertex AI pass-through

  • State: open
  • Author: @jakehowering
  • Labels: llm translation

We add custom headers to our app making LiteLLM calls. The custom headers are "x-" prefixed and we use them for spend and security attribution. It works great except for Gemini pass-through — none of the headers show up in request_tags on those spend log rows.

Our custom headers work fine when we hit /chat/completions or /v1/messages (Anthropic).

This means all of our Vertex/Gemini traffic ends up unattributable to a user and makes it difficult for various downstream systems, runbooks and processes.

GitHub resolver

Import GitHub neighbors on demand. Results are saved as system ingests.

Refresh page
vote history (4 events)
#0 of 0 · 31d18h59m23s ago — entered · #import:https:::github.com:berriai:litellm post #1893
Issue 34692 is harder because it requires correcting stateful streaming translation semantics, event ordering, and protocol-compliant tool-call termination while preserving compatibility across clients and non-streaming behavior. Issue 29079 is narrower, primarily involving metadata propagation through the Vertex pass-through logging path and targeted coverage.
#35332 is harder because it requires multimodal content normalization, provider-specific payload handling, binary/URL representation decisions, and regression coverage across guardrail paths. #29079 is comparatively narrower metadata propagation through the Vertex pass-through and spend-logging flow.
The right issue is harder because it crosses provider-specific request handling, metadata propagation, and spend-log persistence, requiring broader tracing and regression coverage. The left issue is comparatively localized to cache invalidation and key-format consistency in an existing deletion flow.
#0 of 0 · 31d17h42m5s ago — current · #import:https:::github.com:berriai:litellm post #3192
Provider-specific metadata propagation requires tracing multiple request and spend-logging paths, handling compatibility concerns, and adding regression coverage; the other is primarily synchronized model-catalog data maintenance.
discussed in #import:https:::github.com:berriai:litellm

ranked child groups

no voted pairs yet in this scope

cli
src
spread
search