feat(inference): add llmisvc_model_provider_resolver

Cargo.toml

@@ -48,7 +48,7 @@ redis = { version = "1.5.0", default-features = false, features = ["tokio-comp"]
 regex = "1.13.1"
 reqwest = { version = "0.13.4", default-features = false, features = ["rustls", "json", "stream"] }
 rmcp = { version = "3.1.4", default-features = false, features = ["client", "transport-streamable-http-client-reqwest", "reqwest"] }
-rustls = "0.23.43"
+rustls = "0.23.45"
 schemars = "1.2.2"
 rustls-pemfile = "2.2.0"
 secrecy = { version = "0.10.3", features = ["serde"] }

docs/filters/llmisvc_model_provider_resolver.md

@@ -0,0 +1,28 @@
+<!-- Generated by: cargo xtask generate-filter-docs -->
+<!-- Do not edit manually -->
+
+# `llmisvc_model_provider_resolver`
+
+Rewrites publisher-ID body `model` values to the short model name for `LLMISvc` / `KServe` routing; the routing header is left unchanged.
+
+## Configuration Notes
+
+Reads the model name from the configured request header (default `X-Model`, typically set by an earlier `model_to_header`). When that value is a publisher ID (`publishers/.../models/<name>`), rewrite the body `"model"` field to `<name>` only. The routing header is never modified -- `KServe` routes on the publisher ID.
+
+If the header is absent/empty, or the body has no `"model"` field, this filter is a no-op (it does not invent a body `"model"`).
+
+Does **not** resolve `ExternalModel` / `ExternalProvider` CRDs, perform weighted provider selection, rewrite `Host`, or inject credentials.
+
+## Configuration
+
+| Field | Type | Required | Description |
+|-------|------|---------|-------------|
+| `header` | string | no | Request header that carries the publisher ID for `KServe` routing. Defaults to `X-Model` (same as `model_to_header`). |
+| `max_body_bytes` | integer | no | Maximum request body size to buffer before parsing. |
+
+## Example
+
+```yaml
+filter: llmisvc_model_provider_resolver
+header: X-Model   # optional, defaults to X-Model
+```

docs/filters/model_to_header.md

@@ -5,6 +5,10 @@
 
 Promotes the JSON `"model"` field from the request body to a request header.
 
+## Configuration Notes
+
+Promotion is deferred until end-of-stream so a later body-writing filter (for example `llmisvc_model_provider_resolver`) can observe the pending header in the same `StreamBuffer` pre-read pass.
+
 ## Configuration
 
 | Field | Type | Required | Description |

docs/filters/reference.md

@@ -100,6 +100,7 @@ see the [Praxis core filter reference][core-ref].
 
 | Filter | Description |
 |--------|-------------|
+| [`llmisvc_model_provider_resolver`](llmisvc_model_provider_resolver.md) | Rewrites publisher-ID body `model` values to the short model name for `LLMISvc` / `KServe` routing; the routing header is left unchanged. |
 | [`model_to_header`](model_to_header.md) | Promotes the JSON `"model"` field from the request body to a request header. |
 
 ### Metering

examples/README.md

@@ -36,6 +36,7 @@ before sending requests.
 | [json-rpc-routing.yaml](configs/json-rpc-routing.yaml) | Routes JSON-RPC 2.0 requests to different backends based on the "method" field in the JSON request body |
 | [lakera-guard.yaml](configs/lakera-guard.yaml) | Screens every request body through Lakera Guard for content moderation before forwarding to the upstream |
 | [llmd-ext-proc-routing.yaml](configs/llmd-ext-proc-routing.yaml) | A real llm-d EPP or test processor returns the trusted x-gateway-destination-endpoint header |
+| [llmisvc-model-provider-resolver.yaml](configs/llmisvc-model-provider-resolver.yaml) | Rewrites publisher-ID body `model` values to the short model name for LLMISvc / KServe routing; the routing header (default `X-Model`) is left unchanged so routing can still use the publisher ID |
 | [mcp-classifier-routing.yaml](configs/mcp-classifier-routing.yaml) | Routes MCP requests by body-derived method and tool name |
 | [mcp-stateless-broker.yaml](configs/mcp-stateless-broker.yaml) | Configurable stateless MCP broker using the final MCP 2026-07-28 stateless profile |
 | [model-to-header-routing.yaml](configs/model-to-header-routing.yaml) | Routes LLM API requests to different backends based on the "model" field in the JSON request body |

examples/configs/llmisvc-model-provider-resolver.yaml

@@ -0,0 +1,47 @@
+# LLMISvc Model Provider Resolver
+#
+# Build:
+#   cargo build -p praxis-ai-proxy
+#
+# Rewrites publisher-ID body `model` values to the short model name for
+# LLMISvc / KServe routing; the routing header (default `X-Model`) is
+# left unchanged so routing can still use the publisher ID.
+#
+# Typical chain: `model_to_header` promotes the body model to
+# `X-Model`, then `llmisvc_model_provider_resolver` strips the
+# publisher prefix from the body only.
+#
+# Example request:
+#
+#   curl -X POST http://localhost:8080/v1/chat/completions \
+#     -H "Content-Type: application/json" \
+#     -d '{"model":"publishers/rhoai/models/granite-3.1-8b","messages":[{"role":"user","content":"hi"}]}'
+#
+# Upstream body model becomes `granite-3.1-8b` while `X-Model`
+# remains `publishers/rhoai/models/granite-3.1-8b`.
+
+listeners:
+  - name: llmisvc-gateway
+    address: "0.0.0.0:8080" # dev value; binds all interfaces
+    filter_chains:
+      - rewrite-and-route
+
+filter_chains:
+  - name: rewrite-and-route
+    filters:
+      - filter: model_to_header
+        header: X-Model
+      - filter: llmisvc_model_provider_resolver
+        header: X-Model
+      - filter: router
+        routes:
+          - path_prefix: "/"
+            cluster: provider
+      - filter: load_balancer
+        clusters:
+          - name: provider
+            endpoints:
+              - "127.0.0.1:3000"
+
+insecure_options:
+  allow_private_endpoints: true # example proxies to local backends