feat(containers): Azure routing, managed container IDs, delete response parsing

[Sameerlite]

Summary

• Add AzureContainerConfig and register it for Azure / Azure Text in ProviderConfigManager.
• Join container API paths with httpx.URL so api-version stays in the query string (fixes malformed Azure URLs).
• Managed cntr_ IDs: encode on responses/streaming output; decode in proxy container handlers and container SDK helpers so upstream sees the original provider container id.
• DeleteContainerFileResponse: accept OpenAI/Azure wire value container.file.deleted (was rejecting and surfacing 500s after successful DELETE).

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Apr 7, 2026 4:11pm

[codspeed-hq]

Merging this PR will not alter performance

✅ 16 untouched benchmarks

---

_{Comparing litellm_azure-containers-delete-response-and-routing (6843a93) with main (2bb7387)}

[greptile-apps]

Greptile Summary

This PR extends LiteLLM's container API with four improvements: Azure routing (AzureContainerConfig registered for azure/azure_text providers), managed container IDs (containers created via proxy/router are now encoded as cntr_{base64(...)} embedding provider/model affinity metadata, with decode logic stripping them before upstream calls), URL fix (join_container_api_base_path using httpx.URL ensures Azure's ?api-version=v1 always appears after path segments like /cntr_.../files), and DeleteContainerFileResponse wire-format fix (now accepts both container.file.deleted and container_file.deleted, eliminating 500s after successful DELETE).

• Style: Three inline imports in container_handler.py violate CLAUDE.md's module-level import rule (BaseLLMException in _sync_handle/_async_handle, and extract_file_data in _prepare_multipart_file_upload).
• Performance: _load_endpoints_config() in container_handler.py opens endpoints.json on every container file API call with no caching.
• Affinity edge case: encode_container_id_in_response uses target_model_names[0] as the model affinity anchor when model_info.id is absent; with multiple targets and round-robin, the encoded model may not match the deployment that actually created the container.

Confidence Score: 5/5

Safe to merge; all remaining findings are P2 style and performance suggestions

Core logic is correct and well-tested: AzureContainerConfig registration, managed ID encode/decode (None-guard fix applied), httpx.URL path joining, and DeleteContainerFileResponse wire-format fix all work correctly. The two P2 findings (inline imports, uncached JSON read) are non-blocking.

litellm/llms/custom_httpx/container_handler.py (inline imports + uncached JSON read); litellm/containers/utils.py (target_model_names affinity edge case)

Important Files Changed

Filename	Overview
litellm/llms/azure/containers/transformation.py	New AzureContainerConfig inheriting OpenAIContainerConfig; delegates auth/URL to BaseAzureLLM — clean and minimal
litellm/llms/openai/containers/utils.py	New join_container_api_base_path helper using httpx.URL to preserve query-string when appending paths — fixes Azure api-version placement bug
litellm/llms/custom_httpx/container_handler.py	New GenericContainerHandler; _load_endpoints_config() re-reads endpoints.json without caching on every request; three inline imports
litellm/containers/utils.py	decode_managed_container_id_for_request and encode_container_id_in_response added; target_model_names[0] affinity may not match actual deployment
litellm/responses/utils.py	_build_container_id/_decode_container_id helpers with correct None-guard; robust regex-based decoding with safe fallback
litellm/types/containers/main.py	DeleteContainerFileResponse now accepts container.file.deleted (OpenAI/Azure wire) and container_file.deleted (legacy)
litellm/proxy/container_endpoints/handler_factory.py	Proxy handlers decode managed container IDs; decode pattern duplicated across three functions; inline imports remain
litellm/containers/main.py	All container functions accept azure/azure_text and call decode_managed_container_id_for_request before provider config lookup
litellm/containers/endpoint_factory.py	Endpoint factory updated for azure/azure_text; decode_managed_container_id_for_request called before dispatching
tests/test_litellm/containers/test_azure_container_transformation.py	Comprehensive unit tests for AzureContainerConfig including URL construction and api-version regression tests
tests/test_litellm/containers/test_container_utils.py	Tests for decode_managed_container_id_for_request roundtrip and DeleteContainerFileResponse wire format
tests/test_litellm/containers/test_container_api.py	Container API tests covering encode/decode roundtrip at the API layer
tests/test_litellm/responses/test_responses_utils.py	ResponsesAPIRequestUtils tests including container ID encode/decode roundtrip
litellm/responses/streaming_iterator.py	Streaming iterator now calls _encode_container_id_on_output_item for output_item events so managed IDs appear incrementally in streaming
litellm/utils.py	ProviderConfigManager.get_provider_container_config now returns AzureContainerConfig for AZURE/AZURE_TEXT providers
litellm/llms/azure/containers/init.py	New empty init.py for the azure containers module
litellm/llms/openai/containers/transformation.py	All URL-building methods now use join_container_api_base_path instead of f-string concatenation

Sequence Diagram

sequenceDiagram
    participant Client
    participant Proxy as LiteLLM Proxy
    participant Handler as handler_factory.py
    participant SDK as containers/main.py
    participant Azure as Azure OpenAI

    Note over Client,Azure: Container Create (encoding)
    Client->>Proxy: POST /v1/containers {name}
    Proxy->>SDK: create_container(provider=azure)
    SDK->>Azure: POST /openai/v1/containers
    Azure-->>SDK: {id: "cntr_abc"}
    SDK->>SDK: encode_container_id_in_response()<br/>id -> cntr_{b64(litellm:azure;model_id:X;cntr_abc)}
    SDK-->>Proxy: ContainerObject(id=cntr_{b64...})
    Proxy-->>Client: {id: "cntr_{b64...}"}

    Note over Client,Azure: Follow-up (decoding)
    Client->>Proxy: GET /v1/containers/cntr_{b64...}/files
    Proxy->>Handler: _process_request(container_id=cntr_{b64...})
    Handler->>Handler: _decode_container_id()<br/>-> original=cntr_abc, provider=azure
    Handler->>SDK: list_container_files(cntr_abc, azure)
    SDK->>Azure: GET /openai/v1/containers/cntr_abc/files?api-version=v1
    Azure-->>SDK: ContainerFileListResponse
    SDK-->>Proxy: ContainerFileListResponse
    Proxy-->>Client: file list

Loading

_{Reviews (3): Last reviewed commit: "Fix p1 concern" | Re-trigger Greptile}

[Sameerlite]

@greptile-apps will raise another PR for P1 as this is not an issue introduced by this feat but an already existing issue

[greptile-apps]

That makes sense. Addressing pre-existing issues in a separate PR keeps this one focused and makes the changes easier to review. Just make sure to link the P1 PR here when you open it so we can track the fix.

_{Tip: You can customize Greptile's behavior for this repo with .greptile/rules and .greptile/config.}