Feature/multi provider routing policy changes

event-gateway/default-policies/openai-header-router.yaml

@@ -0,0 +1,78 @@
+name: openai-header-router
+version: v1.0.0
+description: |
+  Selects a target LLM provider for each incoming OpenAI Chat Completions
+  request by inspecting a request header, then publishes the chosen
+  provider id into SharedContext.Metadata["selected_provider"].
+
+  This policy is a "setter" — its only job is to publish the selection.
+  Downstream consumer policies (the per-provider translators such as
+  openai-to-anthropic, openai-to-azure-openai, and openai-to-mistral)
+  read that key and own both the translation and the upstream-cluster
+  selection.
+
+  Selection algorithm:
+
+  1. The configured header is read from the request (case-insensitive).
+  2. Each entry in "mappings" is checked: if its "headerValue" matches the
+     incoming header value (case-insensitive), the entry's "provider" is
+     selected.
+  3. If the header is absent, empty, or no mapping matches, the configured
+     "defaultProvider" is selected.
+
+  The defaultProvider field is required so this policy always produces a
+  selection — there is no "pass through with no provider" mode.
+
+parameters:
+  type: object
+  additionalProperties: false
+  properties:
+    headerName:
+      type: string
+      x-wso2-policy-advanced-param: false
+      description: |
+        Name of the request header to read for provider selection.
+        Comparison is case-insensitive.
+      default: x-provider
+      minLength: 1
+    defaultProvider:
+      type: string
+      x-wso2-policy-advanced-param: false
+      description: |
+        Provider id selected when the header is missing, empty, or does
+        not match any entry in "mappings". This must be a provider id that
+        a downstream translator is bound to.
+      minLength: 1
+    mappings:
+      type: array
+      x-wso2-policy-advanced-param: false
+      description: |
+        Mapping from incoming header values to provider ids. The first
+        matching entry wins.
+      items:
+        type: object
+        additionalProperties: false
+        properties:
+          headerValue:
+            type: string
+            description: |
+              Header value to match (case-insensitive, leading/trailing
+              whitespace trimmed).
+            minLength: 1
+          provider:
+            type: string
+            description: |
+              Provider id published to SharedContext.Metadata
+              ["selected_provider"] when this entry matches.
+            minLength: 1
+        required:
+          - headerValue
+          - provider
+      minItems: 1
+  required:
+    - mappings
+    - defaultProvider
+
+systemParameters:
+  type: object
+  properties: {}

event-gateway/default-policies/openai-to-anthropic.yaml

@@ -0,0 +1,84 @@
+name: openai-to-anthropic
+version: v1.0.0
+description: |
+  Translates an incoming OpenAI Chat Completions request into the Anthropic
+  Messages API format and forwards it to an Anthropic-compatible upstream.
+
+  Request transformations:
+  - system/developer messages are lifted into the top-level "system" field.
+  - messages[] is converted to Anthropic role/content blocks.
+  - assistant tool_calls become tool_use content blocks; tool-role messages
+    are grouped into a single user message of tool_result blocks.
+  - tools[] / tool_choice are converted to the Anthropic equivalents.
+  - stop is mapped to stop_sequences; image_url parts to image blocks.
+  - The request path is rewritten to /v1/messages and the anthropic-version
+    header is set.
+
+  Response transformations:
+  - The Anthropic Messages response body is rewritten into the OpenAI
+    ChatCompletion shape (id, object, created, model, choices, usage).
+  - text content blocks become choice.message.content; tool_use blocks
+    become choice.message.tool_calls; stop_reason is mapped to
+    finish_reason; cache token counts are carried in usage
+    .prompt_tokens_details.
+  - Anthropic error envelopes are rewritten into OpenAI error envelopes
+    so clients see consistent error shapes regardless of provider.
+  - Streaming (Server-Sent Events) responses are passed through
+    untouched in this policy version; clients that need true streaming
+    should request non-streaming (omit "stream": true) until a streaming
+    variant ships.
+
+  Execution modes:
+
+  1. Routed (multi-provider proxies): set the "provider" parameter to the
+     provider id this translator is bound to. The policy only runs when an
+     upstream setter (e.g. openai-header-router) publishes the same id into
+     SharedContext.Metadata["selected_provider"]. Non-matching requests pass
+     through untouched.
+
+  2. Unconditional (single-provider proxies): omit "provider". The
+     translator runs on every request.
+
+  The upstream host and Anthropic API key must be configured on the
+  upstream LlmProvider — this policy does not inject credentials.
+
+parameters:
+  type: object
+  additionalProperties: false
+  properties:
+    provider:
+      type: string
+      x-wso2-policy-advanced-param: false
+      description: |
+        Provider id this translator is bound to. When set, the policy runs
+        only if SharedContext.Metadata["selected_provider"] equals this
+        value. Omit for single-provider proxies.
+      minLength: 1
+    model:
+      type: string
+      x-wso2-policy-advanced-param: false
+      description: |
+        Anthropic model name used in the translated request (for example
+        claude-sonnet-4-20250514). Overrides the OpenAI "model" field.
+      minLength: 1
+    upstreamName:
+      type: string
+      x-wso2-policy-advanced-param: false
+      description: |
+        Name of the upstream cluster declared in the LlmProxy
+        additionalProviders list. When omitted, routing is left to whatever
+        the route's default upstream resolves to.
+      minLength: 1
+    anthropicVersion:
+      type: string
+      x-wso2-policy-advanced-param: true
+      description: |
+        Value of the anthropic-version request header sent upstream.
+      default: "2023-06-01"
+      minLength: 1
+  required:
+    - model
+
+systemParameters:
+  type: object
+  properties: {}

event-gateway/default-policies/openai-to-azure-openai.yaml

@@ -0,0 +1,78 @@
+name: openai-to-azure-openai
+version: v1.0.0
+description: |
+  Translates an incoming OpenAI Chat Completions request into the Azure
+  OpenAI request shape and forwards it to an Azure-OpenAI-compatible
+  upstream.
+
+  Transformation:
+  - The request path is rewritten to
+    /openai/deployments/{deployment}/{pathSuffix}?api-version=<apiVersion>.
+  - The body passes through unchanged — Azure OpenAI accepts the OpenAI
+    request shape directly.
+  - The deployment id is taken from the "model" parameter when set,
+    otherwise from the request body's "model" field.
+
+  Execution modes:
+
+  1. Routed (multi-provider proxies): set the "provider" parameter to the
+     provider id this translator is bound to. The policy only runs when an
+     upstream setter (e.g. openai-header-router) publishes the same id into
+     SharedContext.Metadata["selected_provider"]. Non-matching requests
+     pass through untouched.
+
+  2. Unconditional (single-provider proxies): omit "provider". The
+     translator runs on every request.
+
+  The upstream host and Azure api-key must be configured on the upstream
+  LlmProvider — this policy does not inject credentials.
+
+parameters:
+  type: object
+  additionalProperties: false
+  properties:
+    provider:
+      type: string
+      x-wso2-policy-advanced-param: false
+      description: |
+        Provider id this translator is bound to. When set, the policy runs
+        only if SharedContext.Metadata["selected_provider"] equals this
+        value. Omit for single-provider proxies.
+      minLength: 1
+    apiVersion:
+      type: string
+      x-wso2-policy-advanced-param: false
+      description: |
+        Azure OpenAI api-version query-string value (for example
+        2024-02-15-preview). Required because Azure rejects requests
+        without it.
+      minLength: 1
+    model:
+      type: string
+      x-wso2-policy-advanced-param: false
+      description: |
+        Azure deployment id used in the rewritten path. When omitted, the
+        request body's "model" field is used as the deployment id.
+      minLength: 1
+    upstreamName:
+      type: string
+      x-wso2-policy-advanced-param: false
+      description: |
+        Name of the upstream cluster declared in the LlmProxy
+        additionalProviders list. When omitted, routing is left to whatever
+        the route's default upstream resolves to.
+      minLength: 1
+    pathSuffix:
+      type: string
+      x-wso2-policy-advanced-param: true
+      description: |
+        Endpoint suffix appended after the deployment segment in the
+        rewritten path. Defaults to /chat/completions.
+      default: /chat/completions
+      minLength: 1
+  required:
+    - apiVersion
+
+systemParameters:
+  type: object
+  properties: {}

event-gateway/default-policies/openai-to-gemini.yaml

@@ -0,0 +1,86 @@
+name: openai-to-gemini
+version: v1.0.0
+description: |
+  Translates an incoming OpenAI Chat Completions request into the Gemini
+  generateContent API format and forwards it to a Gemini-compatible
+  upstream.
+
+  Request transformations:
+  - system/developer messages are lifted into the top-level
+    "systemInstruction" field.
+  - messages[] is converted to Gemini contents[] with role/parts blocks;
+    assistant tool_calls become functionCall parts and tool-role messages
+    become functionResponse parts.
+  - tools[] are converted to functionDeclarations and tool_choice to a
+    functionCallingConfig.
+  - sampling fields (temperature, top_p, max_tokens, stop, etc.) are mapped
+    into generationConfig; image_url parts become inlineData/fileData.
+  - The request path is rewritten to
+    /{apiVersion}/models/{model}:generateContent.
+
+  Response transformations:
+  - The Gemini generateContent response body is rewritten into the OpenAI
+    ChatCompletion shape (id, object, created, model, choices, usage).
+  - text parts become choice.message.content; functionCall parts become
+    choice.message.tool_calls; finishReason is mapped to finish_reason and
+    usageMetadata token counts are carried in usage.
+  - Streaming (Server-Sent Events) responses are passed through untouched
+    in this policy version; clients that need true streaming should request
+    non-streaming (omit "stream": true) until a streaming variant ships.
+
+  Execution modes:
+
+  1. Routed (multi-provider proxies): set the "provider" parameter to the
+     provider id this translator is bound to. The policy only runs when an
+     upstream setter (e.g. openai-header-router) publishes the same id into
+     SharedContext.Metadata["selected_provider"]. Non-matching requests
+     pass through untouched.
+
+  2. Unconditional (single-provider proxies): omit "provider". The
+     translator runs on every request.
+
+  The upstream host and Gemini API key must be configured on the upstream
+  LlmProvider — this policy does not inject credentials.
+
+parameters:
+  type: object
+  additionalProperties: false
+  properties:
+    provider:
+      type: string
+      x-wso2-policy-advanced-param: false
+      description: |
+        Provider id this translator is bound to. When set, the policy runs
+        only if SharedContext.Metadata["selected_provider"] equals this
+        value. Omit for single-provider proxies.
+      minLength: 1
+    model:
+      type: string
+      x-wso2-policy-advanced-param: false
+      description: |
+        Gemini model name used in the translated request (for example
+        gemini-2.5-pro). Overrides the OpenAI "model" field and is used in
+        the rewritten path.
+      minLength: 1
+    upstreamName:
+      type: string
+      x-wso2-policy-advanced-param: false
+      description: |
+        Name of the upstream cluster declared in the LlmProxy
+        additionalProviders list. When omitted, routing is left to whatever
+        the route's default upstream resolves to.
+      minLength: 1
+    apiVersion:
+      type: string
+      x-wso2-policy-advanced-param: true
+      description: |
+        Gemini API version segment used in the rewritten path
+        (/{apiVersion}/models/{model}:generateContent).
+      default: "v1beta"
+      minLength: 1
+  required:
+    - model
+
+systemParameters:
+  type: object
+  properties: {}

event-gateway/default-policies/openai-to-mistral.yaml

@@ -0,0 +1,61 @@
+name: openai-to-mistral
+version: v1.0.0
+description: |
+  Translates an incoming OpenAI Chat Completions request into the format
+  expected by Mistral AI's chat completions endpoint and forwards it to a
+  Mistral-compatible upstream.
+
+  Mistral's chat completions API is largely OpenAI-compatible, so the
+  translation is small in scope:
+  - The "model" field is overridden with the configured Mistral model.
+  - OpenAI request fields that Mistral does not accept (logprobs, n,
+    service_tier, store, metadata, user, etc.) are stripped.
+  - The request path is rewritten to /v1/chat/completions.
+
+  Execution modes:
+
+  1. Routed (multi-provider proxies): set the "provider" parameter to the
+     provider id this translator is bound to. The policy only runs when an
+     upstream setter (e.g. openai-header-router) publishes the same id into
+     SharedContext.Metadata["selected_provider"]. Non-matching requests
+     pass through untouched.
+
+  2. Unconditional (single-provider proxies): omit "provider". The
+     translator runs on every request.
+
+  The upstream host and Mistral API key must be configured on the upstream
+  LlmProvider — this policy does not inject credentials.
+
+parameters:
+  type: object
+  additionalProperties: false
+  properties:
+    provider:
+      type: string
+      x-wso2-policy-advanced-param: false
+      description: |
+        Provider id this translator is bound to. When set, the policy runs
+        only if SharedContext.Metadata["selected_provider"] equals this
+        value. Omit for single-provider proxies.
+      minLength: 1
+    model:
+      type: string
+      x-wso2-policy-advanced-param: false
+      description: |
+        Mistral model name used in the translated request (for example
+        mistral-large-latest). Overrides the OpenAI "model" field.
+      minLength: 1
+    upstreamName:
+      type: string
+      x-wso2-policy-advanced-param: false
+      description: |
+        Name of the upstream cluster declared in the LlmProxy
+        additionalProviders list. When omitted, routing is left to whatever
+        the route's default upstream resolves to.
+      minLength: 1
+  required:
+    - model
+
+systemParameters:
+  type: object
+  properties: {}