## Summary Add **Amazon Bedrock Mantle** as a supported provider. Mantle is AWS's OpenAI-compatible Bedrock inference surface (prefix `bedrock_mantle/`, `custom_llm_provider: bedrock_mantle`), distinct from the Converse API we integrate today, and the only Bedrock surface exposing newer models (GPT-5.5/5.4, GPT-OSS, `anthropic.claude-mythos-preview`). Investigation: gitlab-org/modelops/applied-ml/code-suggestions/ai-assist#2524. LiteLLM docs: https://docs.litellm.ai/docs/providers/bedrock_mantle ## Key findings - **No LiteLLM bump** — pinned `1.84.5` already routes `bedrock_mantle/<model>` (verified via `get_llm_provider`); the AI Gateway's generic `partition("/")` parser needs no change. - **Endpoint is user-supplied** — set via the existing `API` self-hosted platform (Endpoint field) and passed through to LiteLLM as `api_base`. - **API key is optional** — env vars (`BEDROCK_MANTLE_API_KEY` / `AWS_*`) can supply auth. ## Decisions - **No new gitlab provider/platform.** Bedrock Mantle is OpenAI-compatible, so it is configured through the **existing `API` self-hosted platform**: Endpoint `https://bedrock-mantle.{region}.api.aws/v1`, Model identifier `bedrock_mantle/<model>`, optional API key. All enabling work lives in the AI Gateway. (Originally scoped as a new dedicated platform; dropped once we confirmed the `API` platform already supports it.) - **GitLab-hosted managed models deferred** (filed as follow-up, not in this epic's first round). ## First-round children 1. [AIGW] Enable `bedrock_mantle` provider (allow-lists + auth/config) — !6240 (merged) 2. [AIGW] Support self-hosted (BYO) Bedrock Mantle models + diagnostics — !6241 3. [gitlab] Docs: self-hosted Bedrock Mantle setup guide (via the `API` platform) Closed as not needed (Mantle works through the existing `API` platform): [gitlab] backend + GraphQL platform, [gitlab] frontend platform.