fix(recipe-scan): wire cf-docuvision OCR + LLMRouter for cloud recipe scanning (kiwi#136)

Two-step pipeline: task_allocate("kiwi", "recipe_scan", service_hint="cf-docuvision")
acquires a docuvision allocation, calls /extract per image to get OCR text, then
LLMRouter structures the combined OCR output into recipe JSON via the text
extraction prompt.

Also fixes DocuvisionClient bugs:
- POST field was "image" (ignored by Pydantic) — should be "image_b64"
- Response read "text" key — docuvision returns "raw_text"
- Add hint parameter (use "text" for recipe cards, dense prose)
- Configurable timeout (default 120s; docuvision lazy-loads model on first request)

app/services/meal_plan/llm_router.py

@@ -2,17 +2,20 @@
 # BSL 1.1 — LLM feature
 """Provide a router-compatible LLM client for meal plan generation tasks.
 
-Cloud (CF_ORCH_URL set):
-  Allocates a cf-text service via cf-orch (3B-7B GGUF, ~2GB VRAM).
-  Returns an _OrchTextRouter that wraps the cf-text HTTP endpoint
-  with a .complete(system, user, **kwargs) interface.
+Cloud (CF_ORCH_URL set), tier 1 — task-based routing (preferred):
+  Calls /api/inference/task with product=kiwi, task=meal_plan.
+  The coordinator resolves the model from assignments.yaml.
+
+Cloud (CF_ORCH_URL set), tier 2 — direct allocation (fallback):
+  Allocates cf-text directly via client.allocate(). Used when the task
+  is not yet registered in the coordinator (cf-orch#61 not deployed).
 
 Local / self-hosted (no CF_ORCH_URL):
   Returns an LLMRouter instance which tries ollama, vllm, or any
   backend configured in ~/.config/circuitforge/llm.yaml.
 
-Both paths expose the same interface so llm_timing.py and llm_planner.py
-need no knowledge of the backend.
+All paths expose the same (router, ctx) interface so llm_planner.py
+needs no knowledge of the backend.
 """
 from __future__ import annotations
 
@@ -22,8 +25,7 @@ from contextlib import nullcontext
 
 logger = logging.getLogger(__name__)
 
-# cf-orch service name and VRAM budget for meal plan LLM tasks.
-# These are lighter than recipe_llm (4.0 GB) — cf-text handles them.
+# cf-orch service name and TTL for direct-allocate fallback path.
 _SERVICE_TYPE = "cf-text"
 _TTL_S = 120.0
 _CALLER = "kiwi-meal-plan"
@@ -62,35 +64,79 @@ class _OrchTextRouter:
         return resp.choices[0].message.content or ""
 
 
+# Imported at module level so tests can patch the names in this module's namespace.
+# app.services.task_inference.task_allocate — patch target for task routing tests.
+try:
+    from app.services.task_inference import TaskNotRegistered, task_allocate
+    _HAS_TASK_INFERENCE = True
+except ImportError:
+    _HAS_TASK_INFERENCE = False
+
+# circuitforge_orch.client.CFOrchClient — patch target for direct-allocate fallback tests.
+try:
+    from circuitforge_orch.client import CFOrchClient
+except ImportError:
+    CFOrchClient = None  # type: ignore[assignment,misc]
+
+# circuitforge_core.llm.router.LLMRouter — patch target for local-inference tests.
+try:
+    from circuitforge_core.llm.router import LLMRouter
+except (ImportError, FileNotFoundError):
+    LLMRouter = None  # type: ignore[assignment,misc]
+
+
 def get_meal_plan_router():
     """Return an LLM client for meal plan tasks.
 
-    Tries cf-orch cf-text allocation first (cloud); falls back to LLMRouter
-    (local ollama/vllm). Returns None if no backend is available.
+    Returns (router, ctx) where ctx is a context manager the caller holds
+    open for the duration of the LLM call. Returns (None, nullcontext(None))
+    if no backend is available.
     """
     cf_orch_url = os.environ.get("CF_ORCH_URL")
-    if cf_orch_url:
-        try:
-            from circuitforge_orch.client import CFOrchClient
-            client = CFOrchClient(cf_orch_url)
-            ctx = client.allocate(
-                service=_SERVICE_TYPE,
-                ttl_s=_TTL_S,
-                caller=_CALLER,
-            )
-            alloc = ctx.__enter__()
-            if alloc is not None:
-                return _OrchTextRouter(alloc.url), ctx
-        except Exception as exc:
-            logger.debug("cf-orch cf-text allocation failed, falling back to LLMRouter: %s", exc)
 
-    # Local fallback: LLMRouter (ollama / vllm / openai-compat)
-    try:
-        from circuitforge_core.llm.router import LLMRouter
-        return LLMRouter(), nullcontext(None)
-    except FileNotFoundError:
-        logger.debug("LLMRouter: no llm.yaml and no LLM env vars — meal plan LLM disabled")
-        return None, nullcontext(None)
-    except Exception as exc:
-        logger.debug("LLMRouter init failed: %s", exc)
-        return None, nullcontext(None)
+    if cf_orch_url:
+        # Tier 1: task-based routing — coordinator owns model selection.
+        if _HAS_TASK_INFERENCE:
+            try:
+                ctx = task_allocate(
+                    "kiwi", "meal_plan",
+                    service_hint=_SERVICE_TYPE,
+                    ttl_s=_TTL_S,
+                )
+                alloc = ctx.__enter__()
+                return _OrchTextRouter(alloc.url), ctx
+            except TaskNotRegistered:
+                logger.debug(
+                    "kiwi.meal_plan not in coordinator assignments — "
+                    "falling back to direct cf-text allocation"
+                )
+            except Exception as exc:
+                logger.debug("task allocation failed, trying direct allocate: %s", exc)
+
+        # Tier 2: direct allocation — hardcoded service type.
+        if CFOrchClient is not None:
+            try:
+                client = CFOrchClient(cf_orch_url)
+                ctx = client.allocate(
+                    service=_SERVICE_TYPE,
+                    ttl_s=_TTL_S,
+                    caller=_CALLER,
+                )
+                alloc = ctx.__enter__()
+                if alloc is not None:
+                    return _OrchTextRouter(alloc.url), ctx
+                ctx.__exit__(None, None, None)  # release allocation before falling through
+            except Exception as exc:
+                logger.debug("cf-orch cf-text allocation failed, falling back to LLMRouter: %s", exc)
+
+    # Tier 3: local inference — ollama / vllm / openai-compat.
+    if LLMRouter is not None:
+        try:
+            return LLMRouter(), nullcontext(None)
+        except FileNotFoundError:
+            logger.debug("LLMRouter: no llm.yaml and no LLM env vars — meal plan LLM disabled")
+            return None, nullcontext(None)
+        except Exception as exc:
+            logger.debug("LLMRouter init failed: %s", exc)
+            return None, nullcontext(None)
+    return None, nullcontext(None)

app/services/ocr/docuvision_client.py

@@ -18,43 +18,51 @@ class DocuvisionResult:
 class DocuvisionClient:
     """Thin client for the cf-docuvision service."""
 
-    def __init__(self, base_url: str) -> None:
+    def __init__(self, base_url: str, timeout: float = 120.0) -> None:
         self._base_url = base_url.rstrip("/")
+        self._timeout = timeout
 
-    def extract_text(self, image_path: str | Path) -> DocuvisionResult:
-        """Send an image to docuvision and return extracted text."""
+    def extract_text(self, image_path: str | Path, hint: str = "text") -> DocuvisionResult:
+        """Send an image to docuvision and return extracted text.
+
+        Args:
+            image_path: Path to the image file.
+            hint: Docuvision extraction hint — "text" for dense prose (recipes),
+                  "table" for tabular data, "form" for form fields, "auto" for
+                  automatic detection.
+        """
         image_bytes = Path(image_path).read_bytes()
         b64 = base64.b64encode(image_bytes).decode()
 
-        with httpx.Client(timeout=30.0) as client:
+        with httpx.Client(timeout=self._timeout) as client:
             resp = client.post(
                 f"{self._base_url}/extract",
-                json={"image": b64},
+                json={"image_b64": b64, "hint": hint},
             )
             resp.raise_for_status()
             data = resp.json()
 
         return DocuvisionResult(
-            text=data.get("text", ""),
-            confidence=data.get("confidence"),
+            text=data.get("raw_text", ""),
+            confidence=data.get("metadata", {}).get("confidence"),
             raw=data,
         )
 
-    async def extract_text_async(self, image_path: str | Path) -> DocuvisionResult:
+    async def extract_text_async(self, image_path: str | Path, hint: str = "text") -> DocuvisionResult:
         """Async version."""
         image_bytes = Path(image_path).read_bytes()
         b64 = base64.b64encode(image_bytes).decode()
 
-        async with httpx.AsyncClient(timeout=30.0) as client:
+        async with httpx.AsyncClient(timeout=self._timeout) as client:
             resp = await client.post(
                 f"{self._base_url}/extract",
-                json={"image": b64},
+                json={"image_b64": b64, "hint": hint},
             )
             resp.raise_for_status()
             data = resp.json()
 
         return DocuvisionResult(
-            text=data.get("text", ""),
-            confidence=data.get("confidence"),
+            text=data.get("raw_text", ""),
+            confidence=data.get("metadata", {}).get("confidence"),
             raw=data,
         )

app/services/ocr/vl_model.py

@@ -32,6 +32,29 @@ def _try_docuvision(image_path: str | Path) -> str | None:
     cf_orch_url = os.environ.get("CF_ORCH_URL")
     if not cf_orch_url:
         return None
+
+    # Tier 1: task-based routing — coordinator owns model selection.
+    try:
+        from app.services.task_inference import task_allocate, TaskNotRegistered
+        from app.services.ocr.docuvision_client import DocuvisionClient
+        try:
+            with task_allocate(
+                "kiwi", "ocr",
+                service_hint="cf-docuvision",
+                ttl_s=60.0,
+            ) as alloc:
+                doc_client = DocuvisionClient(alloc.url)
+                result = doc_client.extract_text(image_path)
+                return result.text if result.text else None
+        except TaskNotRegistered:
+            logger.debug(
+                "kiwi.ocr not in coordinator assignments — "
+                "falling back to direct cf-docuvision allocation"
+            )
+    except Exception as exc:
+        logger.debug("task allocation path failed, trying direct allocate: %s", exc)
+
+    # Tier 2: direct allocation — hardcoded service type.
     try:
         from circuitforge_orch.client import CFOrchClient
         from app.services.ocr.docuvision_client import DocuvisionClient
@@ -49,7 +72,7 @@ def _try_docuvision(image_path: str | Path) -> str | None:
             result = doc_client.extract_text(image_path)
             return result.text if result.text else None
     except Exception as exc:
-        logger.debug("cf-docuvision fast-path failed, falling back: %s", exc)
+        logger.debug("cf-docuvision fast-path failed, falling back to local VLM: %s", exc)
         return None
 
 

app/services/recipe/recipe_scanner.py

@@ -196,34 +196,63 @@ def _call_via_local_vlm(image_paths: list[Path], prompt: str) -> str:
     return output
 
 
+def _build_ocr_extraction_prompt(ocr_text: str) -> str:
+    """Build a text-LLM prompt for structuring OCR output into recipe JSON.
+
+    Swaps the image-centric preamble of _EXTRACTION_PROMPT for an OCR-centric
+    one, then appends the combined OCR text as input. The JSON schema section
+    is shared verbatim to keep the two paths in sync.
+    """
+    schema_idx = _EXTRACTION_PROMPT.find("Return a single JSON object")
+    schema_part = _EXTRACTION_PROMPT[schema_idx:] if schema_idx != -1 else _EXTRACTION_PROMPT
+    return (
+        "You are extracting a recipe from OCR text taken from a recipe card, "
+        "cookbook page, or handwritten note.\n\n"
+        "The text below was obtained via optical character recognition and may "
+        "contain minor scanning artifacts or formatting irregularities.\n\n"
+        f"{schema_part}\n\nOCR Text:\n{ocr_text}"
+    )
+
+
 def _call_vision_backend(image_paths: list[Path], prompt: str) -> str:
     """Dispatch to the best available vision backend.
 
-    Priority: cf-orch vision -> local Qwen2.5-VL -> Anthropic API.
+    Priority: cf-orch docuvision (OCR + text LLM) -> local Qwen2.5-VL -> Anthropic API.
     Raises RuntimeError with a clear message when no backend is available.
     """
     errors: list[str] = []
 
-    # 1. Try cf-orch vision allocation
+    # 1. Try cf-orch task allocation → cf-docuvision OCR, then text LLM structuring.
+    #    Two-step: docuvision extracts text from the image(s), then LLMRouter
+    #    converts the OCR text to structured recipe JSON using the extraction prompt.
     cf_orch_url = os.environ.get("CF_ORCH_URL")
     if cf_orch_url:
         try:
-            from circuitforge_orch.client import CFOrchClient
+            from app.services.task_inference import TaskNotRegistered, task_allocate
             from app.services.ocr.docuvision_client import DocuvisionClient
+            from circuitforge_core.llm.router import LLMRouter
 
-            client = CFOrchClient(cf_orch_url)
-            with client.allocate(
-                service="cf-vision",
-                model_candidates=["qwen2.5-vl-7b", "cf-docuvision"],
-                ttl_s=90.0,
-                caller="kiwi-recipe-scan",
-            ) as alloc:
-                if alloc is not None:
+            try:
+                with task_allocate("kiwi", "recipe_scan", service_hint="cf-docuvision", ttl_s=120.0) as alloc:
+                    # Step 1: OCR each image via cf-docuvision
                     doc_client = DocuvisionClient(alloc.url)
-                    # docuvision takes a single image -- use first image only for now
-                    result = doc_client.extract_text(image_paths[0])
-                    if result.text:
-                        return result.text
+                    ocr_parts: list[str] = []
+                    for i, path in enumerate(image_paths):
+                        result = doc_client.extract_text(path, hint="text")
+                        prefix = f"(Page {i + 1} of the same recipe)\n" if len(image_paths) > 1 else ""
+                        ocr_parts.append(f"{prefix}{result.text}")
+                    combined_ocr = "\n\n".join(ocr_parts)
+
+                    if not combined_ocr.strip():
+                        raise ValueError("Docuvision returned no text — image may not be a recipe")
+
+                    # Step 2: Text LLM structures OCR output into recipe JSON
+                    text = LLMRouter().complete(_build_ocr_extraction_prompt(combined_ocr))
+                    if text:
+                        return text
+
+            except TaskNotRegistered:
+                logger.debug("kiwi.recipe_scan not yet registered in cf-orch assignments")
         except Exception as exc:
             logger.debug("cf-orch vision failed for recipe scan: %s", exc)
             errors.append(f"cf-orch: {exc}")

app/services/task_inference.py

@@ -0,0 +1,124 @@
+# app/services/task_inference.py
+# BSL 1.1 — LLM feature
+"""Task-based service allocation via the cf-orch coordinator.
+
+Calls POST /api/inference/task instead of a hardcoded service type.
+The coordinator resolves model_id and service_type from assignments.yaml.
+
+Fallback contract (for callers):
+  - 404 → TaskNotRegistered  (fall back to direct client.allocate())
+  - other error → RuntimeError
+  - CF_ORCH_URL unset → RuntimeError (guard with os.environ.get first)
+"""
+from __future__ import annotations
+
+import logging
+import os
+from collections.abc import Generator
+from contextlib import contextmanager
+from dataclasses import dataclass
+
+import httpx
+
+logger = logging.getLogger(__name__)
+
+
+class TaskNotRegistered(Exception):
+    """Coordinator returned 404 for a product/task pair.
+
+    Means the task is not yet in assignments.yaml. Callers should fall
+    back to direct service allocation (client.allocate()).
+    """
+
+
+@dataclass(frozen=True)
+class Allocation:
+    url: str
+    allocation_id: str
+    service: str
+
+
+def _orch_url() -> str:
+    return os.environ.get("CF_ORCH_URL", "").rstrip("/")
+
+
+@contextmanager
+def task_allocate(
+    product: str,
+    task: str,
+    *,
+    service_hint: str,
+    ttl_s: float = 120.0,
+) -> Generator[Allocation, None, None]:
+    """Context manager: allocate a service via task-based routing.
+
+    Calls POST /api/inference/task, yields Allocation, releases on exit.
+    Supports both `with task_allocate(...) as alloc:` and manual
+    `ctx = task_allocate(...); alloc = ctx.__enter__()` patterns.
+
+    **Sync-only**: uses the synchronous httpx API. Do not call from an
+    ``async def`` handler without wrapping in ``asyncio.to_thread``. Current
+    call sites (``llm_router.py``, ``vl_model.py``) are synchronous.
+
+    Args:
+        product: CF product name (e.g. "kiwi")
+        task: Task identifier (e.g. "meal_plan", "ocr")
+        service_hint: Service type for the release DELETE call. The
+            coordinator response does not include service_type, so the
+            caller provides it. When the coordinator is updated to return
+            service in the response (cf-orch#63), this becomes unused.
+        ttl_s: Allocation TTL in seconds.
+
+    Raises:
+        TaskNotRegistered: Coordinator returned 404.
+        RuntimeError: Coordinator unreachable, returned non-404 error, or
+            returned a malformed (non-JSON / missing fields) response.
+        RuntimeError: CF_ORCH_URL is not set.
+    """
+    base = _orch_url()
+    if not base:
+        raise RuntimeError("CF_ORCH_URL is not set")
+
+    try:
+        resp = httpx.post(
+            f"{base}/api/inference/task",
+            json={"product": product, "task": task, "payload": {}},
+            timeout=30.0,
+        )
+    except httpx.RequestError as exc:
+        raise RuntimeError(f"cf-orch unreachable: {exc}") from exc
+
+    if resp.status_code == 404:
+        raise TaskNotRegistered(
+            f"No assignment for product={product!r} task={task!r} — "
+            "ensure cf-orch#61/62 are deployed and coordinator reloaded"
+        )
+    if not resp.is_success:
+        raise RuntimeError(
+            f"cf-orch /api/inference/task failed: "
+            f"HTTP {resp.status_code} — {resp.text[:200]}"
+        )
+
+    try:
+        data = resp.json()
+        alloc = Allocation(
+            url=data["url"],
+            allocation_id=data["allocation_id"],
+            service=data.get("service") or service_hint,
+        )
+    except (KeyError, ValueError) as exc:
+        raise RuntimeError(
+            f"cf-orch /api/inference/task returned malformed response: {exc} — "
+            f"body: {resp.text[:200]}"
+        ) from exc
+
+    try:
+        yield alloc
+    finally:
+        try:
+            httpx.delete(
+                f"{base}/api/services/{alloc.service}/allocations/{alloc.allocation_id}",
+                timeout=10.0,
+            )
+        except Exception as exc:
+            logger.debug("cf-orch task allocation release failed (non-fatal): %s", exc)

tests/services/test_llm_router_task.py

@@ -0,0 +1,127 @@
+"""Tests for task-based routing added to get_meal_plan_router()."""
+from __future__ import annotations
+
+from unittest.mock import MagicMock
+
+import pytest
+
+
+def _make_task_ctx(url: str = "http://node:8080") -> MagicMock:
+    """Mock context manager returned by task_allocate()."""
+    alloc = MagicMock()
+    alloc.url = url
+    alloc.allocation_id = "alloc-task-1"
+    alloc.service = "cf-text"
+    ctx = MagicMock()
+    ctx.__enter__ = MagicMock(return_value=alloc)
+    ctx.__exit__ = MagicMock(return_value=False)
+    return ctx
+
+
+def _make_task_ctx_not_registered() -> MagicMock:
+    """Mock context manager that raises TaskNotRegistered on enter."""
+    from app.services.task_inference import TaskNotRegistered
+    ctx = MagicMock()
+    ctx.__enter__ = MagicMock(side_effect=TaskNotRegistered("not registered"))
+    ctx.__exit__ = MagicMock(return_value=False)
+    return ctx
+
+
+def _make_direct_alloc_ctx(url: str = "http://node:8080") -> MagicMock:
+    """Mock context manager returned by CFOrchClient.allocate()."""
+    alloc = MagicMock()
+    alloc.url = url
+    ctx = MagicMock()
+    ctx.__enter__ = MagicMock(return_value=alloc)
+    ctx.__exit__ = MagicMock(return_value=False)
+    return ctx
+
+
+def test_task_path_returns_orch_router_on_success(monkeypatch):
+    """get_meal_plan_router() returns _OrchTextRouter when task allocation succeeds."""
+    monkeypatch.setenv("CF_ORCH_URL", "http://coord:7700")
+    import unittest.mock as um
+    # Patch the name as it exists in llm_router's own namespace (module-level import).
+    with um.patch("app.services.meal_plan.llm_router.task_allocate",
+                  return_value=_make_task_ctx(url="http://node:9001")):
+        from app.services.meal_plan.llm_router import get_meal_plan_router, _OrchTextRouter
+        router, ctx = get_meal_plan_router()
+
+    assert isinstance(router, _OrchTextRouter)
+    assert router._base_url == "http://node:9001"
+
+
+def test_task_not_registered_falls_back_to_direct_allocate(monkeypatch):
+    """get_meal_plan_router() falls back to direct cf-text allocation on TaskNotRegistered."""
+    monkeypatch.setenv("CF_ORCH_URL", "http://coord:7700")
+    direct_ctx = _make_direct_alloc_ctx(url="http://node:9002")
+
+    import unittest.mock as um
+    # Patch task_allocate in llm_router's namespace so TaskNotRegistered is raised.
+    with um.patch("app.services.meal_plan.llm_router.task_allocate",
+                  return_value=_make_task_ctx_not_registered()), \
+         um.patch("app.services.meal_plan.llm_router.CFOrchClient") as MockClient:
+        MockClient.return_value.allocate.return_value = direct_ctx
+        from app.services.meal_plan.llm_router import get_meal_plan_router, _OrchTextRouter
+        router, ctx = get_meal_plan_router()
+
+    assert isinstance(router, _OrchTextRouter)
+    assert router._base_url == "http://node:9002"
+
+
+def test_no_cf_orch_url_returns_llm_router(monkeypatch):
+    """get_meal_plan_router() returns LLMRouter when CF_ORCH_URL is not set."""
+    monkeypatch.delenv("CF_ORCH_URL", raising=False)
+
+    import unittest.mock as um
+    mock_lr = MagicMock()
+    with um.patch("app.services.meal_plan.llm_router.LLMRouter", return_value=mock_lr):
+        from app.services.meal_plan.llm_router import get_meal_plan_router
+        router, ctx = get_meal_plan_router()
+
+    assert router is mock_lr
+
+
+def test_tier1_general_exception_falls_back_to_direct_allocate(monkeypatch):
+    """get_meal_plan_router() falls back to direct allocation when task_allocate raises RuntimeError."""
+    monkeypatch.setenv("CF_ORCH_URL", "http://coord:7700")
+    direct_ctx = _make_direct_alloc_ctx(url="http://node:9003")
+
+    import unittest.mock as um
+    failing_ctx = MagicMock()
+    failing_ctx.__enter__ = MagicMock(side_effect=RuntimeError("coordinator down"))
+    failing_ctx.__exit__ = MagicMock(return_value=False)
+
+    with um.patch("app.services.meal_plan.llm_router.task_allocate",
+                  return_value=failing_ctx), \
+         um.patch("app.services.meal_plan.llm_router.CFOrchClient") as MockClient:
+        MockClient.return_value.allocate.return_value = direct_ctx
+        from app.services.meal_plan.llm_router import get_meal_plan_router, _OrchTextRouter
+        router, ctx = get_meal_plan_router()
+
+    assert isinstance(router, _OrchTextRouter)
+    assert router._base_url == "http://node:9003"
+
+
+def test_tier2_none_alloc_releases_ctx_and_falls_through(monkeypatch):
+    """get_meal_plan_router() releases Tier 2 ctx and falls through when alloc is None."""
+    monkeypatch.setenv("CF_ORCH_URL", "http://coord:7700")
+
+    import unittest.mock as um
+
+    none_alloc_ctx = MagicMock()
+    none_alloc_ctx.__enter__ = MagicMock(return_value=None)
+    none_alloc_ctx.__exit__ = MagicMock(return_value=False)
+
+    mock_lr = MagicMock()
+
+    with um.patch("app.services.meal_plan.llm_router.task_allocate",
+                  return_value=_make_task_ctx_not_registered()), \
+         um.patch("app.services.meal_plan.llm_router.CFOrchClient") as MockClient, \
+         um.patch("app.services.meal_plan.llm_router.LLMRouter", return_value=mock_lr):
+        MockClient.return_value.allocate.return_value = none_alloc_ctx
+        from app.services.meal_plan.llm_router import get_meal_plan_router
+        router, ctx = get_meal_plan_router()
+
+    assert router is mock_lr
+    none_alloc_ctx.__exit__.assert_called_once_with(None, None, None)

tests/services/test_task_inference.py

@@ -0,0 +1,164 @@
+"""Tests for app/services/task_inference.py"""
+from __future__ import annotations
+
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+
+def _ok_resp(url: str = "http://node:8080", allocation_id: str = "alloc-123") -> MagicMock:
+    m = MagicMock()
+    m.status_code = 200
+    m.is_success = True
+    m.json.return_value = {
+        "url": url,
+        "allocation_id": allocation_id,
+        "gpu_id": 0,
+        "started": True,
+        "warm": False,
+    }
+    return m
+
+
+def _err_resp(status_code: int, text: str = "error") -> MagicMock:
+    m = MagicMock()
+    m.status_code = status_code
+    m.is_success = False
+    m.text = text
+    return m
+
+
+def test_task_allocate_yields_allocation_on_200(monkeypatch):
+    """task_allocate() yields Allocation with url, allocation_id, service on 200."""
+    monkeypatch.setenv("CF_ORCH_URL", "http://coord:7700")
+    with patch("app.services.task_inference.httpx.post", return_value=_ok_resp()) as mock_post, \
+         patch("app.services.task_inference.httpx.delete") as mock_del:
+        from app.services.task_inference import task_allocate
+        with task_allocate("kiwi", "meal_plan", service_hint="cf-text") as alloc:
+            assert alloc.url == "http://node:8080"
+            assert alloc.allocation_id == "alloc-123"
+            assert alloc.service == "cf-text"
+        called_url = mock_post.call_args[0][0]
+        assert called_url == "http://coord:7700/api/inference/task"
+        mock_del.assert_called_once()
+
+
+def test_task_allocate_uses_service_from_response_when_present(monkeypatch):
+    """task_allocate() uses service from response dict over service_hint when available."""
+    monkeypatch.setenv("CF_ORCH_URL", "http://coord:7700")
+    resp = _ok_resp()
+    resp.json.return_value["service"] = "cf-vision"
+    with patch("app.services.task_inference.httpx.post", return_value=resp), \
+         patch("app.services.task_inference.httpx.delete"):
+        from app.services.task_inference import task_allocate
+        with task_allocate("kiwi", "ocr", service_hint="cf-docuvision") as alloc:
+            assert alloc.service == "cf-vision"
+
+
+def test_task_allocate_404_raises_task_not_registered(monkeypatch):
+    """task_allocate() raises TaskNotRegistered on coordinator 404."""
+    monkeypatch.setenv("CF_ORCH_URL", "http://coord:7700")
+    with patch("app.services.task_inference.httpx.post", return_value=_err_resp(404)):
+        from app.services.task_inference import task_allocate, TaskNotRegistered
+        with pytest.raises(TaskNotRegistered):
+            with task_allocate("kiwi", "meal_plan", service_hint="cf-text"):
+                pass
+
+
+def test_task_allocate_503_raises_runtime_error(monkeypatch):
+    """task_allocate() raises RuntimeError on non-404 coordinator errors."""
+    monkeypatch.setenv("CF_ORCH_URL", "http://coord:7700")
+    with patch("app.services.task_inference.httpx.post", return_value=_err_resp(503, "no GPU")):
+        from app.services.task_inference import task_allocate
+        with pytest.raises(RuntimeError, match="HTTP 503"):
+            with task_allocate("kiwi", "meal_plan", service_hint="cf-text"):
+                pass
+
+
+def test_task_allocate_release_called_on_clean_exit(monkeypatch):
+    """task_allocate() DELETEs the allocation on clean context exit."""
+    monkeypatch.setenv("CF_ORCH_URL", "http://coord:7700")
+    with patch("app.services.task_inference.httpx.post", return_value=_ok_resp(allocation_id="xyz")), \
+         patch("app.services.task_inference.httpx.delete") as mock_del:
+        from app.services.task_inference import task_allocate
+        with task_allocate("kiwi", "meal_plan", service_hint="cf-text"):
+            pass
+        release_url = mock_del.call_args[0][0]
+        assert "cf-text" in release_url
+        assert "xyz" in release_url
+
+
+def test_task_allocate_release_called_when_inner_block_raises(monkeypatch):
+    """task_allocate() DELETEs the allocation even when the inner block raises."""
+    monkeypatch.setenv("CF_ORCH_URL", "http://coord:7700")
+    with patch("app.services.task_inference.httpx.post", return_value=_ok_resp(allocation_id="abc")), \
+         patch("app.services.task_inference.httpx.delete") as mock_del:
+        from app.services.task_inference import task_allocate
+        with pytest.raises(ValueError):
+            with task_allocate("kiwi", "meal_plan", service_hint="cf-text"):
+                raise ValueError("inner error")
+        mock_del.assert_called_once()
+
+
+def test_task_allocate_release_failure_is_swallowed(monkeypatch):
+    """task_allocate() does not propagate DELETE failures."""
+    import httpx as _httpx
+    monkeypatch.setenv("CF_ORCH_URL", "http://coord:7700")
+    with patch("app.services.task_inference.httpx.post", return_value=_ok_resp()), \
+         patch("app.services.task_inference.httpx.delete",
+               side_effect=_httpx.RequestError("gone", request=MagicMock())):
+        from app.services.task_inference import task_allocate
+        with task_allocate("kiwi", "meal_plan", service_hint="cf-text") as alloc:
+            assert alloc.url == "http://node:8080"
+        # no exception raised
+
+
+def test_task_allocate_no_orch_url_raises_runtime_error(monkeypatch):
+    """task_allocate() raises RuntimeError when CF_ORCH_URL is not set."""
+    monkeypatch.delenv("CF_ORCH_URL", raising=False)
+    from app.services.task_inference import task_allocate
+    with pytest.raises(RuntimeError, match="CF_ORCH_URL"):
+        with task_allocate("kiwi", "meal_plan", service_hint="cf-text"):
+            pass
+
+
+def test_task_allocate_network_error_raises_runtime_error(monkeypatch):
+    """task_allocate() wraps httpx.RequestError in RuntimeError."""
+    import httpx as _httpx
+    monkeypatch.setenv("CF_ORCH_URL", "http://coord:7700")
+    with patch("app.services.task_inference.httpx.post",
+               side_effect=_httpx.RequestError("timeout", request=MagicMock())):
+        from app.services.task_inference import task_allocate
+        with pytest.raises(RuntimeError, match="unreachable"):
+            with task_allocate("kiwi", "meal_plan", service_hint="cf-text"):
+                pass
+
+
+def test_task_allocate_malformed_json_raises_runtime_error(monkeypatch):
+    """task_allocate() raises RuntimeError when coordinator returns non-JSON on 200."""
+    monkeypatch.setenv("CF_ORCH_URL", "http://coord:7700")
+    bad_resp = MagicMock()
+    bad_resp.status_code = 200
+    bad_resp.is_success = True
+    bad_resp.text = "<html>proxy error</html>"
+    bad_resp.json.side_effect = ValueError("not json")
+    with patch("app.services.task_inference.httpx.post", return_value=bad_resp):
+        from app.services.task_inference import task_allocate
+        with pytest.raises(RuntimeError, match="malformed"):
+            with task_allocate("kiwi", "meal_plan", service_hint="cf-text"):
+                pass
+
+
+def test_task_allocate_missing_url_field_raises_runtime_error(monkeypatch):
+    """task_allocate() raises RuntimeError when coordinator response is missing url field."""
+    monkeypatch.setenv("CF_ORCH_URL", "http://coord:7700")
+    bad_resp = MagicMock()
+    bad_resp.status_code = 200
+    bad_resp.is_success = True
+    bad_resp.text = '{"allocation_id": "x"}'
+    bad_resp.json.return_value = {"allocation_id": "x"}  # missing "url"
+    with patch("app.services.task_inference.httpx.post", return_value=bad_resp):
+        from app.services.task_inference import task_allocate
+        with pytest.raises(RuntimeError, match="malformed"):
+            with task_allocate("kiwi", "meal_plan", service_hint="cf-text"):
+                pass

tests/services/test_vl_model_task.py

@@ -0,0 +1,88 @@
+"""Tests for task-based routing added to _try_docuvision()."""
+from __future__ import annotations
+
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+
+def _mock_doc_result(text: str = "RECEIPT TEXT") -> MagicMock:
+    r = MagicMock()
+    r.text = text
+    return r
+
+
+def _make_task_ctx(url: str = "http://node:9010") -> MagicMock:
+    alloc = MagicMock()
+    alloc.url = url
+    alloc.allocation_id = "alloc-vis-1"
+    alloc.service = "cf-docuvision"
+    ctx = MagicMock()
+    ctx.__enter__ = MagicMock(return_value=alloc)
+    ctx.__exit__ = MagicMock(return_value=False)
+    return ctx
+
+
+def _make_task_not_registered() -> MagicMock:
+    from app.services.task_inference import TaskNotRegistered
+    ctx = MagicMock()
+    ctx.__enter__ = MagicMock(side_effect=TaskNotRegistered("not registered"))
+    ctx.__exit__ = MagicMock(return_value=False)
+    return ctx
+
+
+def _make_direct_alloc(url: str = "http://node:9011") -> MagicMock:
+    alloc = MagicMock()
+    alloc.url = url
+    ctx = MagicMock()
+    ctx.__enter__ = MagicMock(return_value=alloc)
+    ctx.__exit__ = MagicMock(return_value=False)
+    return ctx
+
+
+def test_try_docuvision_task_path_returns_text(monkeypatch, tmp_path):
+    """_try_docuvision() uses task allocation and returns extracted text on success."""
+    monkeypatch.setenv("CF_ORCH_URL", "http://coord:7700")
+    fake_image = tmp_path / "receipt.jpg"
+    fake_image.write_bytes(b"fake")
+
+    with patch("app.services.task_inference.task_allocate",
+               return_value=_make_task_ctx(url="http://node:9010")), \
+         patch("app.services.ocr.docuvision_client.DocuvisionClient") as MockDoc:
+        MockDoc.return_value.extract_text.return_value = _mock_doc_result("STORE $12.34")
+        from app.services.ocr.vl_model import _try_docuvision
+        result = _try_docuvision(str(fake_image))
+
+    assert result == "STORE $12.34"
+    MockDoc.assert_called_once_with("http://node:9010")
+
+
+def test_try_docuvision_falls_back_to_direct_on_task_not_registered(monkeypatch, tmp_path):
+    """_try_docuvision() falls back to direct cf-docuvision allocation on TaskNotRegistered."""
+    monkeypatch.setenv("CF_ORCH_URL", "http://coord:7700")
+    fake_image = tmp_path / "receipt.jpg"
+    fake_image.write_bytes(b"fake")
+
+    with patch("app.services.task_inference.task_allocate",
+               return_value=_make_task_not_registered()), \
+         patch("circuitforge_orch.client.CFOrchClient") as MockClient, \
+         patch("app.services.ocr.docuvision_client.DocuvisionClient") as MockDoc:
+        MockClient.return_value.allocate.return_value = _make_direct_alloc("http://node:9011")
+        MockDoc.return_value.extract_text.return_value = _mock_doc_result("FALLBACK TEXT")
+        from app.services.ocr.vl_model import _try_docuvision
+        result = _try_docuvision(str(fake_image))
+
+    assert result == "FALLBACK TEXT"
+    MockDoc.assert_called_once_with("http://node:9011")
+
+
+def test_try_docuvision_returns_none_without_cf_orch_url(monkeypatch, tmp_path):
+    """_try_docuvision() returns None immediately when CF_ORCH_URL is not set."""
+    monkeypatch.delenv("CF_ORCH_URL", raising=False)
+    fake_image = tmp_path / "receipt.jpg"
+    fake_image.write_bytes(b"fake")
+
+    from app.services.ocr.vl_model import _try_docuvision
+    result = _try_docuvision(str(fake_image))
+
+    assert result is None

tests/test_services/test_docuvision_client.py

@@ -17,12 +17,17 @@ from app.services.ocr.docuvision_client import DocuvisionClient, DocuvisionResul
 
 
 def test_extract_text_sends_base64_image(tmp_path: Path) -> None:
-    """extract_text() POSTs a base64-encoded image and returns parsed text."""
+    """extract_text() POSTs image_b64 and returns parsed raw_text."""
     image_file = tmp_path / "test.jpg"
     image_file.write_bytes(b"fake-image-bytes")
 
     mock_response = MagicMock()
-    mock_response.json.return_value = {"text": "Cheerios", "confidence": 0.95}
+    mock_response.json.return_value = {
+        "raw_text": "Cheerios",
+        "elements": [],
+        "tables": [],
+        "metadata": {"hint": "text", "confidence": 0.95},
+    }
     mock_response.raise_for_status.return_value = None
 
     with patch("httpx.Client") as mock_client_cls:
@@ -41,7 +46,8 @@ def test_extract_text_sends_base64_image(tmp_path: Path) -> None:
     assert call_kwargs[0][0] == "http://docuvision:8080/extract"
     posted_json = call_kwargs[1]["json"]
     expected_b64 = base64.b64encode(b"fake-image-bytes").decode()
-    assert posted_json["image"] == expected_b64
+    assert posted_json["image_b64"] == expected_b64
+    assert posted_json["hint"] == "text"
 
 
 def test_extract_text_raises_on_http_error(tmp_path: Path) -> None: