feat(config): add GPU_SERVER_URL alias for CF_ORCH_URL

Self-hoster-friendly env var name. Priority: GPU_SERVER_URL →
CF_ORCH_URL (compat) → https://orch.circuitforge.tech when
CF_LICENSE_KEY is present (Paid+ auto-default). Resolved value
written back to os.environ["CF_ORCH_URL"] at startup so all
service callers remain unchanged.

Bump version to 0.10.0.

.env.example

@@ -21,10 +21,12 @@ DATA_DIR=./data
 # IP this machine advertises to the coordinator (must be reachable from coordinator host)
 # CF_ORCH_ADVERTISE_HOST=10.1.10.71
 
-# CF-core hosted coordinator (managed cloud GPU inference — Paid+ tier)
-# Set CF_ORCH_URL to use a hosted cf-orch coordinator instead of self-hosting.
-# CF_LICENSE_KEY is read automatically by CFOrchClient for bearer auth.
-# CF_ORCH_URL=https://orch.circuitforge.tech
+# GPU inference server (cf-orch coordinator for recipe scan, LLM generation, etc.)
+# GPU_SERVER_URL: set to your local cf-orch coordinator (self-hosted rack).
+#   CF_ORCH_URL is the backward-compat alias — both are honoured.
+# Paid+ default: when CF_LICENSE_KEY is present and neither URL is set,
+#   the app automatically points to https://orch.circuitforge.tech.
+# GPU_SERVER_URL=http://10.1.10.71:7700
 # CF_LICENSE_KEY=CFG-KIWI-xxxx-xxxx-xxxx
 
 # LLM backend — env-var auto-config (no llm.yaml needed for bare-metal users)
@@ -57,9 +59,9 @@ CF_APP_NAME=kiwi
 # Unset = auto-detect: true if CLOUD_MODE or circuitforge_orch is installed (paid+ local).
 # Set false to force LocalScheduler even when cf-orch is present.
 # USE_ORCH_SCHEDULER=false
-# CF_ORCH_URL: coordinator endpoint. Required for recipe scan (cf-docuvision) and
-# LLM features when CF_ORCH_URL is the only backend. Local rack: http://10.1.10.71:7700
-# CF_ORCH_URL=http://10.1.10.71:7700
+# GPU_SERVER_URL: cf-orch coordinator endpoint. Required for recipe scan (cf-docuvision)
+# and LLM features on a self-hosted rack. CF_ORCH_URL is the backward-compat alias.
+# GPU_SERVER_URL=http://10.1.10.71:7700
 
 # Cloud mode (set in compose.cloud.yml; also set here for reference)
 # CLOUD_DATA_ROOT=/devl/kiwi-cloud-data

app/core/config.py

@@ -65,9 +65,24 @@ class Settings:
     # Quality
     MIN_QUALITY_SCORE: float = float(os.environ.get("MIN_QUALITY_SCORE", "50.0"))
 
-    # CF-core resource coordinator (VRAM lease management)
+    # CF-core resource coordinator (VRAM lease management — lease broker, not inference)
     COORDINATOR_URL: str = os.environ.get("COORDINATOR_URL", "http://localhost:7700")
 
+    # GPU inference server URL
+    # Priority: GPU_SERVER_URL env var → CF_ORCH_URL env var (backward compat)
+    #           → https://orch.circuitforge.tech when CF_LICENSE_KEY is present (Paid+)
+    # Resolved value is written back to os.environ["CF_ORCH_URL"] at startup so
+    # all service-layer callers that read CF_ORCH_URL directly see the right URL.
+    GPU_SERVER_URL: str | None = (
+        os.environ.get("GPU_SERVER_URL")
+        or os.environ.get("CF_ORCH_URL")
+        or (
+            "https://orch.circuitforge.tech"
+            if os.environ.get("CF_LICENSE_KEY")
+            else None
+        )
+    )
+
     # Hosted cf-orch coordinator — bearer token for managed cloud GPU inference (Paid+)
     # CFOrchClient reads CF_LICENSE_KEY automatically; exposed here for startup validation.
     CF_LICENSE_KEY: str | None = os.environ.get("CF_LICENSE_KEY")
@@ -108,3 +123,9 @@ class Settings:
 
 
 settings = Settings()
+
+# Normalise GPU_SERVER_URL into CF_ORCH_URL so every service-layer caller that
+# reads os.environ.get("CF_ORCH_URL") sees the resolved value, including the
+# Paid+ cloud default injected above.
+if settings.GPU_SERVER_URL:
+    os.environ["CF_ORCH_URL"] = settings.GPU_SERVER_URL

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "kiwi"
-version = "0.6.0"
+version = "0.10.0"
 description = "Pantry tracking + leftover recipe suggestions"
 readme = "README.md"
 requires-python = ">=3.11"