feat: strata engine — proxy models to a remote OpenAI-compatible server

docs/configure.md

@@ -6,7 +6,7 @@ OpenArc uses a YAML based configuration system! Models are added to `config.yaml
 There are a few different blocks in a models config that accept different parameters. Some will be common across implementations while others are specific. `openarc add` writes to these blocks; `openarc add --help` shows one help panel per key.
 
 ### load_config
-- engine: ovgenai, openvino, optimum
+- engine: ovgenai, openvino, optimum, strata
 - model_type: llm, vlm, whisper, kokoro, emb, rerank, qwen3_tts_*
 - device: device for this model
 - tool_call_parser: gemma4, qwen35, hermes
@@ -163,6 +163,25 @@ openarc add \
   --runtime-config '{"PERFORMANCE_HINT": "LATENCY"}'
 ```
 
+### Strata (remote proxy)
+
+`engine=strata` does not load a local model: it proxies requests to a running [Strata](https://github.com/Niko1221/Strata) server's OpenAI-compatible API, which lets OpenArc front models OpenVINO cannot run (GGUF MoE models, for example). `--model-path` carries the endpoint base URL instead of an IR directory, and `--device` is a placeholder. The served model id is discovered from the server's `GET /v1/models` at load time. Alternatively, set the URL as `endpoint` in `--runtime-config` and use any placeholder for `--model-path`.
+
+The proxy deliberately speaks only the stable OpenAI wire protocol (`GET /v1/models`, `POST /v1/chat/completions`, SSE deltas, `usage` events) and nothing Strata-specific, so it is decoupled from Strata's release cadence — and works with any other OpenAI-compatible server that supports those endpoints. All proxy logic lives in the single module `src/engine/strata_proxy.py`.
+
+Metrics for strata models (`ttft`, throughputs, token counts) are measured client-side from the upstream usage event; they carry a `proxy: true` marker.
+
+```
+openarc add \
+  --model-name flash-next \
+  --model-path http://localhost:8080 \
+  --engine strata \
+  --model-type llm \
+  --device strata \
+  --temperature 0.7 \
+  --max-tokens 4096
+```
+
 ### Qwen3-TTS
 
 All 3 flavors of qwen3-tts are supported, with a rich set of configuration options. PRs to improve performance or other enhancements are welcome

src/cli/groups/add.py

@@ -64,11 +64,11 @@ def _fail(extra: str = "") -> NoReturn:
     help='Public facing name of the model.')
 @click.option('--model-path', '--m',
     required=True,
-    help='Path to OpenVINO IR converted model.')
+    help="Path to OpenVINO IR converted model. For engine=strata this is the endpoint base URL instead (e.g. http://localhost:8080).")
 @click.option('--engine', '--en',
-    type=click.Choice(['ovgenai', 'openvino', 'optimum']),
+    type=click.Choice(['ovgenai', 'openvino', 'optimum', 'strata']),
     required=True,
-    help='Engine used to load the model (ovgenai, openvino, optimum)')
+    help='Engine used to load the model (ovgenai, openvino, optimum, strata)')
 @click.option('--model-type', '--mt',
     type=click.Choice([
         'llm', 'vlm', 'whisper', 'qwen3_asr', 'kokoro',
@@ -135,8 +135,13 @@ def add(ctx, model_path, model_name, engine, model_type, device, runtime_config,
     keeps using the contract's own default.
     """
 
-    # Validate model path
-    if not validate_model_path(model_path):
+    # Validate model path. engine=strata carries an endpoint URL in model_path,
+    # not an on-disk IR directory, so the file check does not apply.
+    if engine == "strata":
+        if "://" not in model_path:
+            console.print(f"[red]engine=strata expects an endpoint URL in --model-path (e.g. http://localhost:8080), got: {model_path}[/red]")
+            ctx.exit(1)
+    elif not validate_model_path(model_path):
         console.print(f"[red]Model file check failed! {model_path} does not contain openvino model files OR your chosen path is malformed. Verify chosen path is correct and acquired model files match source on the hub, or the destination of converted model.[/red]")
         ctx.exit(1)
 
@@ -157,7 +162,8 @@ def add(ctx, model_path, model_name, engine, model_type, device, runtime_config,
     # An explicit CACHE_DIR in --runtime-config wins. Resolved to an absolute
     # path here because runtime_config values are passed to OpenVINO verbatim,
     # unlike model_path they are not resolved against the config file.
-    if model_type in ("llm", "vlm"):
+    # engine=strata has no IR (model_path is a URL), so no cache applies.
+    if model_type in ("llm", "vlm") and engine != "strata":
         parsed_runtime_config.setdefault(
             "CACHE_DIR", str((Path(model_path) / "model_cache").resolve())
         )

src/cli/modules/server_config.py

@@ -263,11 +263,13 @@ def get_all_models(self) -> Dict[str, Dict[str, Any]]:
 
     def _resolve_model_paths(self, model_config: Dict[str, Any]) -> Dict[str, Any]:
         """Return a copy of model_config with relative model_path, draft_model_path,
-        and cache_dir made absolute by joining them onto the config file's directory."""
+        and cache_dir made absolute by joining them onto the config file's directory.
+
+        URL values (e.g. a strata engine's endpoint in model_path) are left alone."""
         resolved = dict(model_config)
 
         path = resolved.get("model_path")
-        if path and not Path(path).is_absolute():
+        if path and "://" not in path and not Path(path).is_absolute():
             resolved["model_path"] = str((self.config_file.parent / path).resolve())
 
         draft_model_path = resolved.get("draft_model_path")