API translation for non-Anthropic providers (Bedrock, Vertex AI, Azure, Cohere)

[noalimoy]

What

Add request/response translation filters so that OpenAI Chat
Completions clients can transparently use AWS Bedrock, Google Vertex AI
(Gemini), Azure OpenAI, and Cohere backends through Praxis — the same
pattern already working for Anthropic Messages via anthropic_to_openai.

Each provider gets a dedicated HttpFilter under apis/src/ that:

• Rewrites the OpenAI request body to the provider's native JSON shape
• Rewrites the provider's response back to Chat Completions format
• Handles streaming: SSE payload translation for Azure/Cohere/Vertex,
  binary eventstream decoding for Bedrock
• Strips protocol-specific client headers that don't belong on the
  upstream hop (e.g. OpenAI Authorization when the gateway injects
  its own via credential_inject / gcp_adc / aws_sigv4_sign)

One filter per provider, one example config per provider, registered
in register.rs, with fixture-backed tests.

Why

Praxis already sits in the request path doing classification, routing,
credential injection, token counting, and guardrails. The Anthropic
translation path proved the pattern works. But four major providers
remain uncovered — forcing operators to either:

1. Maintain provider SDKs in every application, or
2. Add a second translation hop (LiteLLM, etc.) alongside Praxis

Adding translation filters eliminates the second hop and keeps all
AI gateway concerns in one place: one config file, one credential
boundary, one observability surface.

Prior work

This resumes praxis-proxy/ai#114 (Epic: Schema translation); a previous
proposal (praxis-proxy/ai#761) and per-provider sub-tasks (praxis-proxy/ai#778–#781)
were closed when proposals moved to praxis-proxy/enhancements, not because the work was completed.
praxis-proxy/ai#363 (Epic: API Translation) covers OpenAI/Anthropic and explicitly defers these four providers to #114.

Scope

• Per-provider request transform: OpenAI Chat Completions → native JSON
• Per-provider response transform: native JSON → Chat Completions
• Per-provider error normalization
• Streaming: SSE event payload translation (Azure, Cohere, Vertex AI)
  and binary eventstream decoding (Bedrock)
• Protocol-specific client header stripping
• Example configs and fixture-backed integration tests

Out of scope

• OpenAI Responses ↔ Chat Completions (Filter 16: responses_to_chat_completions ai#35)
• Anthropic Messages ↔ OpenAI (done, Anthropic Messages API ai#103)
• Credential lifecycle, token counting, and provider routing are handled
  by existing filters and tracked separately.

[noalimoy]

cc: @shaneutt / @franciscojavierarceo / @szedan-rh

[leseb]

Not sure, we need a new approval process for this, since we've had an epic for a while on this topic.

[noalimoy]

@leseb Makes sense. should I skip the proposal entirely and go straight to sub-task issues + implementation PRs against the epic?

[leseb]

Personally, I would go ahead and start creating sub issues to praxis-proxy/ai#114 for each providers we want to support and get to work :)

[shaneutt]

Yes, thanks for putting in the discussion, but go at praxis-proxy/ai#114 👍