Bedrock streaming: message_start emits input_tokens=0, breaking downstream OTel metrics

[riyas-rawther]

Summary

When using Headroom as an Anthropic API proxy with a Bedrock backend in streaming mode, the message_start SSE event always contains "usage": {"input_tokens": 0, "output_tokens": 0}. This causes downstream clients (e.g. Claude Code) that read input_tokens from message_start to report 0 for all input token metrics.

Environment

• Headroom version: 0.26.0
• Backend: litellm-bedrock (Bedrock via LiteLLM)
• Mode: streaming (stream: true)
• Client: Claude Code 2.1.181

Observed Behavior

Claude Code emits OTel metrics based on the usage fields in the Anthropic SSE stream:

• message_start → reads usage.input_tokens, usage.cache_read_input_tokens, usage.cache_creation_input_tokens
• message_delta → reads usage.output_tokens

When connecting through Headroom to Bedrock:

• input_tokens = always 0
• cache_read_input_tokens = always 0
• cache_creation_input_tokens = always 0
• output_tokens = correct (non-zero)

When connecting directly to Bedrock (without Headroom):

• All fields are correctly populated by Claude Code (e.g., input_tokens=1, cache_read_input_tokens=61711, etc.)

Evidence

Raw CloudWatch Logs comparison (same OTel pipeline, same day):

Through Headroom (my account):

04:14:30  type=cacheRead      tokens=0
04:14:30  type=cacheCreation  tokens=0
04:14:30  type=input          tokens=0
04:14:30  type=output         tokens=57

Direct to Bedrock (colleague, no Headroom):

10:00:46  type=cacheRead      tokens=61711
10:00:46  type=cacheCreation  tokens=407
10:00:46  type=input          tokens=1
10:00:46  type=output         tokens=291

Impact

• Token usage dashboards (Athena/CloudWatch) show near-zero usage for Headroom users
• Cost attribution queries produce incorrect results
• The "Top Users by Token Usage" report underreports Headroom users by ~99%
• Only output_tokens is tracked correctly

Reproduction

1. Configure Headroom with --backend bedrock
2. Send a streaming request through the proxy
3. Inspect the message_start SSE event received by the client
4. Observe usage.input_tokens is always 0

The non-streaming path does not have this issue — the full usage block from Bedrock's response is passed through correctly in JSONResponse(content=backend_response.body).