Forward user identity to upstream providers for per-user cost attribution

[matifali]

Context

An enterprise customer has requested the ability to see per-user cost breakdowns on the upstream provider side (e.g. Anthropic Console, OpenAI usage dashboard). Since AI Bridge already authenticates the Coder user and attaches Actor.ID + recorder.Metadata to every request context, we have the information — we just don't forward it to the upstream provider.

Problem

When enterprises use AI Bridge to proxy Claude Code (or other tools) to Anthropic/OpenAI, all requests appear to come from a single API key on the provider's console. There's no way for the customer to correlate provider-side cost reports with individual Coder users.

Proposed Solution

When AI Bridge proxies a request to a supported provider, inject the authenticated Coder user identifier into the provider's native user tracking field.

Provider field mapping

Provider	Endpoint	Field	Status	Docs
Anthropic	POST /v1/messages	metadata.user_id	Stable	API Reference
OpenAI	POST /v1/chat/completions	user	Stable	API Reference
OpenAI	POST /v1/responses	safety_identifier	Stable	API Reference

Anthropic example

{
  "model": "claude-sonnet-4-5-20250929",
  "max_tokens": 1024,
  "metadata": {
    "user_id": "<coder-actor-id>"
  },
  "messages": [...]
}

OpenAI /v1/chat/completions example

{
  "model": "gpt-4o",
  "user": "<coder-actor-id>",
  "messages": [...]
}

OpenAI /v1/responses example

{
  "model": "gpt-4.1",
  "safety_identifier": "<coder-actor-id>",
  "input": [...]
}

Implementation Notes

• The Actor is already available via context.ActorFromContext(ctx) in the interceptor — this is a matter of injecting it into the request body before forwarding upstream.
• Overwrite policy: If the client sends its own metadata.user_id / user, AI Bridge should overwrite it. The gateway is the source of truth for user identity; trusting client-supplied values would undermine attribution accuracy.
• Value to send: Actor.ID is the simplest starting point. We may want to make this configurable later (username vs. user ID vs. email). OpenAI recommends hashing to avoid sending PII — we should consider whether to hash the Coder username/email or send the opaque Coder user ID directly.
• OTEL correlation: Ensure the same identifier is emitted in OTEL spans (Basic tracing with OTel #26) so customers can correlate AI Bridge traces with provider-side reports. This is especially important for Anthropic where provider-side per-user reporting doesn't exist yet — our OTEL data becomes the single source of truth.
• All three endpoints need this: /v1/messages (Anthropic), /v1/chat/completions (OpenAI), and /v1/responses (OpenAI) are all live today — the implementation should cover all three.

Scope

Phase 1 (this issue):

• Anthropic metadata.user_id on POST /v1/messages
• OpenAI user on POST /v1/chat/completions
• OpenAI safety_identifier on POST /v1/responses

Phase 2:

• Allow modifying what identifier to send.

Customer Value

Enterprises get per-user cost attribution on the provider side with zero additional configuration — AI Bridge handles it transparently. For OpenAI, this unlocks immediate per-user dashboards and Usage API grouping. For Anthropic, it provides abuse detection and positions customers for future reporting capabilities, while AI Bridge's own OTEL + Recorder serves as the primary per-user cost attribution layer.

This is low-effort, high-value: the data is already in the request context, it just needs to be attached to the upstream request body.