feat(codec): add OCI Generative AI provider codec in Rust core

[fede-kamel]

Overview

Adds Oracle Cloud Infrastructure (OCI) Generative AI as a built-in provider codec in the Rust core, with exposure through the Python and Node bindings — the Rust rework requested in #549 review.

• I confirm this contribution is my own work, or I have the right to submit it under this project's license.
• I searched existing issues and open pull requests, and this does not duplicate existing work.

Details

What: OCIGenAIChatCodec (LlmCodec + LlmResponseCodec) and OCIGenAIStreamingCodec in crates/core/src/codec/oci_genai.rs, covering both OCI chat wire formats — GENERIC (Meta Llama, Google Gemini, xAI Grok, OpenAI, and imported open-weights models such as NVIDIA Nemotron on dedicated AI clusters) and COHERE — with typed ApiSpecificRequest::OCIGenAI / ApiSpecificResponse::OCIGenAI variants, registration in the provider-surface resolver (strongest-signal detection on the chatRequest/servingMode envelope, placed first; oci/oci.genai hints honored), PyOCIGenAIChatCodec (pyo3) and OCIGenAIChatCodec (napi) bindings, and full support in the guardrails, PII-redaction overlay, and adaptive components that match exhaustively on provider surfaces.

Why: OCI GenAI traffic is currently opaque to Relay middleware and observability. With a core codec, request intercepts (PII redaction, guardrails, policy), streaming, cost accounting, and LLMEnd annotations work on OCI payloads in every language binding, exactly as they do for the OpenAI and Anthropic built-ins.

How: Follows the established provider pattern (anthropic.rs as the closest template): unit-struct codec, PROVIDER_SURFACE descriptor, in-module tests, multi-convention key handling (SDK camelCase, CLI kebab-case, snake_case). encode() uses merge-not-replace baseline-compare-and-patch per the LlmCodec contract: unchanged messages pass through from the raw payload verbatim, so encode(decode(original), original) == original at the JSON level and unmodeled provider fields survive edits (fully for GENERIC; COHERE holds identity for unedited requests, while message edits rebuild the modeled message/chatHistory/preambleOverride fields). Model edits are rejected with InvalidArgument; tool/tool-choice edits encode via Function mapping with ProviderNative passthrough.

Testing:

• 32 new codec unit tests (fixtures ported from real OCI wire traffic): decode/encode round-trips for both formats, identity invariants with unmodeled fields at envelope/request/message level, redaction-style edits preserving untouched messages verbatim, tool-call conversion, param patch-only-changed, response decoding across all three key conventions, request/response detection incl. non-shadowing against the other three surfaces, and streaming.
• 7 new cross-provider parity cases; 2 PII-redaction overlay tests (GENERIC choices + COHERE text); resolver, guardrails-component, and Python test_builtin_codecs.py coverage for the new surface.
• cargo build --workspace clean; cargo test --workspace fully green (core lib 982+ tests, integration suite passing); cargo clippy --workspace --all-targets zero warnings; cargo fmt clean; missing_docs satisfied.
• End-to-end validation: the Rust codec (via the Python binding) decoded, identity-encoded, and round-tripped a request whose encoded payload was posted verbatim to the signed OCI REST chat endpoint and answered by a dedicated AI cluster serving an imported NVIDIA Nemotron 3 model; the response codec normalized the reply including token usage.

Breaking changes: None intended. from_provider_surface in the adaptive ACG request-surface mapping now returns Option (OCI has no ACG applier) — internal API, call sites updated.

Where should the reviewer start?

crates/core/src/codec/oci_genai.rs — detection rules and the baseline-compare-and-patch encode() are the key decisions — then the identity-invariant cases in crates/core/tests/unit/codec/oci_genai_tests.rs and the new cases in parity_tests.rs.

Related Issues: (use one of the action keywords Closes / Fixes / Resolves / Relates to)

• Closes [Enhancement]: Support Oracle Cloud Infrastructure (OCI) Generative AI as a provider #548
• Relates to Add OCI Generative AI provider codecs and LangChain integration guide #549 (Python approach, closed per review — this PR is the Rust-core rework)

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Enterprise

Run ID: b5ed41f6-3f7e-4551-a27b-527155f9f1f3

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

_{Comment @coderabbitai help to get the list of available commands.}

[fede-kamel]

Splitting this into a stacked series of smaller PRs (≤~1k lines each) in foundation order for reviewability: (1) types + response codec, (2) request codec with merge-not-replace encode, (3) resolver registration + guardrails/PII/adaptive surface support, (4) streaming codec + Python/Node bindings. Same commit content, same test coverage, same end-to-end validation; series tracked in #548. First PR of the series follows shortly.