Litellm apikey trace

[Harshit28j]

Claude code trace issues

[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.
1 out of 2 committers have signed the CLA.

✅ Harshit28j
❌ github-actions[bot]
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[Harshit28j]

@greptile can you please review this PR?

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Feb 27, 2026 4:42pm

[greptile-apps]

Greptile Summary

This PR threads the kwargs dict (containing litellm_params with API key metadata, passthrough_logging_payload, call_type, and litellm_call_id) from pass_through_request through the streaming handler chain to the provider-specific logging handlers. It also updates async_success_handler in litellm_logging.py to merge response_cost and model from kwargs into model_call_details. This fixes API key attribution in Langfuse traces for streaming pass-through requests.

• kwargs propagation: pass_through_endpoints.py → streaming_handler.py → Anthropic/OpenAI/Vertex handlers now carry the full kwargs containing API key metadata
• Cost merging in litellm_logging.py: async_success_handler now picks up response_cost and model from kwargs when model_call_details doesn't already have them
• Bug: Anthropic handler overwrites proxy_server_request: The .update() call in _create_anthropic_response_logging_payload (line 144) replaces the entire proxy_server_request key with just {"body": {"user": user}}, losing URL, method, body, and headers from the incoming kwargs. The OpenAI handler uses .setdefault() which is correct.
• Bug: Missing asyncio import in test: test_unit_test_streaming.py uses await asyncio.sleep(0.5) but never imports asyncio, causing a NameError at runtime

Confidence Score: 2/5

• PR has a data-loss bug in the Anthropic handler and a broken test that needs fixing before merge
• The core plumbing of kwargs through the streaming pipeline is well-structured, but the Anthropic handler's .update() call now overwrites the incoming proxy_server_request metadata (a regression introduced by this PR), and the new streaming test will fail at runtime due to a missing asyncio import.
• Pay close attention to litellm/proxy/pass_through_endpoints/llm_provider_handlers/anthropic_passthrough_logging_handler.py (proxy_server_request overwrite) and tests/pass_through_unit_tests/test_unit_test_streaming.py (missing import)

Important Files Changed

Filename	Overview
litellm/litellm_core_utils/litellm_logging.py	Correctly merges response_cost and model from kwargs into model_call_details for pass-through endpoints, enabling proper cost tracking in standard logging payloads.
litellm/proxy/pass_through_endpoints/llm_provider_handlers/anthropic_passthrough_logging_handler.py	Accepts kwargs parameter to thread API key metadata, but the .update() call on litellm_params overwrites the incoming proxy_server_request with just the user field, losing URL/method/body/headers.
litellm/proxy/pass_through_endpoints/llm_provider_handlers/openai_passthrough_logging_handler.py	Correctly preserves incoming litellm_params from kwargs and uses safe .setdefault() pattern for proxy_server_request. Adds call_type and litellm_call_id propagation.
litellm/proxy/pass_through_endpoints/llm_provider_handlers/vertex_passthrough_logging_handler.py	Cleanly threads kwargs through to vertex logging payload builder by renaming local variable to _kwargs to avoid shadowing the parameter.
litellm/proxy/pass_through_endpoints/pass_through_endpoints.py	Passes the kwargs dict (containing litellm_params, passthrough_logging_payload, etc.) to chunk_processor for both streaming code paths.
litellm/proxy/pass_through_endpoints/streaming_handler.py	Propagates kwargs through chunk_processor to _route_streaming_logging_to_handler and renames local variable to handler_kwargs to avoid shadowing. Clean refactor.
tests/pass_through_unit_tests/test_unit_test_streaming.py	New test for kwargs propagation is missing import asyncio, causing a NameError at runtime. The test will always fail as-is.
tests/test_litellm/litellm_core_utils/test_litellm_logging.py	Good test verifying that response_cost from kwargs is merged into model_call_details and that API key metadata appears in the standard logging object.
poetry.lock	Bumps litellm-proxy-extras from 0.4.48 to 0.4.49. Routine dependency update.

Sequence Diagram

sequenceDiagram
    participant PT as pass_through_request
    participant CP as chunk_processor
    participant RH as _route_streaming_logging_to_handler
    participant AH as AnthropicHandler
    participant OH as OpenAIHandler
    participant VH as VertexHandler
    participant ASH as async_success_handler

    PT->>CP: kwargs (litellm_params, passthrough_logging_payload, call_type, litellm_call_id)
    CP->>CP: Yield raw bytes to client
    CP->>RH: kwargs forwarded via asyncio.create_task
    alt Anthropic endpoint
        RH->>AH: _handle_logging_anthropic_collected_chunks(kwargs)
        AH->>AH: _create_anthropic_response_logging_payload(kwargs)
        AH-->>RH: handler_kwargs (response_cost, model, litellm_params)
    else OpenAI endpoint
        RH->>OH: _handle_logging_openai_collected_chunks(kwargs)
        OH-->>RH: handler_kwargs (response_cost, model, litellm_params)
    else Vertex endpoint
        RH->>VH: _handle_logging_vertex_collected_chunks(kwargs)
        VH-->>RH: handler_kwargs (response_cost, model)
    end
    RH->>ASH: result + **handler_kwargs
    ASH->>ASH: Merge response_cost & model from kwargs into model_call_details
    ASH->>ASH: Build standard_logging_object with API key metadata

Loading

_{Last reviewed commit: f67cb1a}

[greptile-apps]

_{9 files reviewed, 4 comments}

_{Edit Code Review Agent Settings | Greptile}

[Harshit28j]

@greptile can you please review this PR?

[greptile-apps]

Missing asyncio import causes NameError

asyncio is used here (await asyncio.sleep(0.5)) but is never imported in this file. This will cause a NameError at runtime, meaning this test will always fail. Add import asyncio at the top of the file with the other module-level imports.

Suggested change

	await asyncio.sleep(0.5)
	await asyncio.sleep(0.5)

The import needs to be added at the top of the file:

import asyncio
import json
import os
...

Context Used: Context from dashboard - CLAUDE.md (source)

[greptile-apps]

Additional Comments (1)

litellm/proxy/pass_through_endpoints/llm_provider_handlers/anthropic_passthrough_logging_handler.py
.update() overwrites incoming proxy_server_request

Now that kwargs carries the full litellm_params from _init_kwargs_for_pass_through_endpoint (which includes proxy_server_request with url, method, body, and headers), this .update() call replaces the entire proxy_server_request with just {"body": {"user": user}}, losing all the original request metadata.

Before this PR, kwargs was always {} so there was nothing to overwrite. With the new plumbing, this becomes a data-loss bug.

The OpenAI handler correctly uses .setdefault() at its equivalent code path (line 570 of openai_passthrough_logging_handler.py). The same approach should be used here:

                    kwargs["litellm_params"].setdefault("proxy_server_request", {}).setdefault("body", {})["user"] = user