feat(azure): add azure_to_openai translation filter

[noalimoy]

Summary

Add an azure_to_openai HttpFilter under apis/src/azure/ that lets standard Chat Completions clients talk to Azure OpenAI deployments through Praxis. Azure's wire format is nearly identical to Chat Completions — the filter handles the small but operationally critical differences:

• Request: injects api-version query parameter, strips the model field (Azure uses the deployment name in the URL path), removes Accept-Encoding to prevent double-compressed bodies.
• Response (JSON): strips prompt_filter_results and per-choice content_filter_results/content_filter_offsets.
• Response (SSE): reuses the shared SseFrameParser rather than hand-rolling line splitting — the parser already handles cross-chunk buffering, bare \r and split \r\n line endings, data: without space after colon, multi-line data: joining, and buffer-overflow protection. The filter calls parse_chunk, strips Azure fields from each frame's data payload via the same strip_azure_fields used for non-streaming JSON, and re-emits clean SSE.
• Errors: fills type: null from the HTTP status code and strips innererror, normalizing Azure errors to the standard Chat Completions error envelope.

SSE handled inline, not as a separate filter

Unlike anthropic_stream_events (which is a standalone filter doing full format translation between OpenAI and Anthropic event schemas), Azure SSE is already Chat Completions format — the only work is stripping the same content_filter_results fields that strip_azure_fields already removes from non-streaming JSON. A separate filter would add config surface and registration overhead for no new logic, so the SSE path lives alongside the JSON and error paths in on_response_body.

Follows the same {provider}/to_openai module structure established by anthropic/to_openai. This is the first provider filter for Epic #114 (Schema translation); the remaining providers (Bedrock, Vertex AI, Cohere) follow separately.

Related issue

Part of #114
Closes #886

Validation

• Unit tests — 34 Azure-specific tests in praxis-ai-apis (config parsing, request/response stripping, SSE frame rebuild, partial-line buffering via SseFrameParser, multi-byte UTF-8 boundary, error normalization)
• Integration or functional tests — 5 tests via azure_translation.rs (request forwarding with api-version, content-filter stripping, SSE event stripping, error normalization with type: null+innererror, pass-through of already-valid errors)
• make lint — clippy (default + all-features), nightly fmt, rustdoc all pass

Checklist

• I reviewed every changed line and can explain the change.
• New capabilities include an example config and functional example test.
• User-facing behavior and generated documentation are updated.
• Performance-sensitive changes include appropriate benchmark or load-test evidence.
• Commits are signed and include a Signed-off-by trailer.

Breaking changes

None. New filter only; no existing behavior or API is changed.

[praxis-bot]

Review: feat(azure): add azure_to_openai translation filter

Well-structured filter following established {provider}/to_openai module patterns. Clean separation of request, response, SSE, and error normalization. Good reuse of the shared SseFrameParser for streaming. Unit test coverage (34 tests) and integration tests (5 tests) are thorough.

Findings

#	Severity	File	Description
1	Medium	apis/src/azure/wire.rs	normalize_error_response does not handle entirely-missing type field
2	Medium	apis/src/azure/to_openai/mod.rs	rebuild_sse_frames silently drops event_type from SSE frames

[praxis-bot]

PR Review

Azure-to-OpenAI translation filter — re-review focusing on new findings only (2 prior comments already posted).

1 new finding below.

[leseb]

@noalimoy do not force push - it's hard to determine which commits addressed reviews, PTAL https://github.com/praxis-proxy/ai/blob/main/CONTRIBUTING.md#pr-review-process

[leseb]

A few finding from AI which look legit:

[P1] api-version is lost. Both the Azure filter and subsequent path_rewrite overwrite ctx.rewritten_path from the original URI. The documented ordering drops the required query parameter; reversing it drops the deployment path. Existing api-version values are also duplicated instead of replaced. The integration test never inspects the forwarded URI. Code, test, Azure requirement.

[P1] Asynchronous moderation evidence is deleted after content was streamed. The response transformation strips content_filter_results and content_filter_offsets. Azure asynchronous filtering emits text before moderation annotations; those offsets identify which emitted range must be redacted. Preserve/translate these annotations or reject unsupported asynchronous mode. Code, Azure streaming semantics.

Additionally, could we add an integration test using the official OpenAI Python SDK? It should cover non-streaming, streaming, and errors through the proxy, with a strict local fake Azure backend asserting the rewritten path, api-version, and request body. No live Azure endpoint is needed. Since client compatibility is the core contract of this filter, I consider this required before merging.

[leseb]

@noalimoy please read https://github.com/praxis-proxy/ai/blob/main/CONTRIBUTING.md#pr-review-process on how to resolve praxis bot discussions. Thanks

[leseb]

please resolve all the conversations when ready

[noalimoy]

@noalimoy please read https://github.com/praxis-proxy/ai/blob/main/CONTRIBUTING.md#pr-review-process on how to resolve praxis bot discussions. Thanks

@leseb sorry, that last force-push was a mistake. review fixes are in 6b192bb; :)

[leseb]

@noalimoy lint fails

[noalimoy]

@noalimoy lint fails

fixed