Split api handlers

[michalkulakowski]

🛠 Summary

JIRA/Issue if applicable.
Describe the changes.

🧪 Checklist

• Unit tests added.
• The documentation updated.
• Change follows security best practices.
  ``

[Copilot]

Pull request overview

This PR splits OpenAI REST API handling into a shared base (OpenAIApiHandler) plus endpoint-specific handlers, moving /v3/responses logic into a dedicated OpenAIResponsesHandler and updating serving code, build targets, tests, and documentation accordingly.

Changes:

• Introduces OpenAIApiHandler base class and refactors chat/completions handler to inherit from it.
• Adds dedicated OpenAIResponsesHandler with Responses-specific parsing + (streaming/unary) serialization and wires it into servables/calculator.
• Updates Bazel targets, unit tests, and adds REST API documentation for /v3/responses.

Reviewed changes

Copilot reviewed 14 out of 15 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
src/test/http_openai_handler_test.cpp	Updates tests to construct correct handler type per endpoint and validates new streaming event sequence APIs.
src/llm/visual_language_model/legacy/servable.cpp	Instantiates OpenAIResponsesHandler for RESPONSES and adjusts streaming chunk emission behavior for empty chunks.
src/llm/language_model/legacy/servable.cpp	Same as above for legacy language model servable.
src/llm/servable.hpp	Switches execution context to store a polymorphic std::shared_ptr<OpenAIApiHandler>.
src/llm/servable.cpp	Creates responses vs chat handler based on endpoint; adjusts streaming chunk emission comments/behavior.
src/llm/http_llm_calculator.cc	Emits response.created lifecycle event for streaming via handler hook; uses generic serializeFailedEvent.
src/llm/apis/openai_api_handler.hpp / src/llm/apis/openai_api_handler.cpp	New shared base class with common parsing, tool parsing, usage, and shared utilities.
src/llm/apis/openai_responses.hpp / src/llm/apis/openai_responses.cpp	New Responses endpoint handler (request parsing + unary/streaming serialization).
src/llm/apis/openai_completions.hpp / src/llm/apis/openai_completions.cpp	Refactors chat/completions handler to inherit shared logic and removes embedded Responses logic.
src/llm/BUILD	Adds Bazel targets for the new handler libraries and wires dependencies.
src/BUILD	Links the new Responses handler library into the main build.
docs/model_server_rest_api_responses.md	Adds user documentation for /v3/responses, including streaming event semantics.

[Copilot]

Local file path allowlist check uses std::mismatch prefix matching, which can be bypassed by paths that share the same prefix (e.g., allowed "/tmp/allowed" would also allow "/tmp/allowed2/..."), and doesn’t canonicalize symlinks/relative segments. Consider using std::filesystem::weakly_canonical/lexically_normal and then verifying the canonical path is within allowedLocalMediaPath (including a path-separator boundary check).

[Copilot]

Response IDs are derived only from seconds since epoch (e.g., "resp-"). Under concurrency, multiple requests created within the same second can collide and produce identical response IDs across different streams/unary calls. Consider using microseconds (as earlier streaming implementation did) or a random/UUID-based suffix to guarantee uniqueness while keeping the existing prefix format.

[Copilot]

When input[i].content is an array, multiple input_text parts will overwrite contentText (only the last text part is preserved). The OpenAI Responses schema allows multiple content parts; if more than one input_text is provided, consider concatenating (or otherwise preserving ordering) instead of replacing, so the full prompt reaches the chat template/model.