cmd/cognitive-worker/main.go

@@ -97,7 +97,8 @@ func main() {
 	if gpuErr != nil {
 		log.Printf("optional GPU provider disabled: %v", gpuErr)
 	} else if gpuConfigured {
-		gpuAdapter, adapterErr := models.NewOpenAICompatibleLocalAdapterWithAPIKey(gpuURL, gpuKey, nil)
+		gpuTimeout := durationEnv("LOCAL_GPU_PROVIDER_TIMEOUT", 60*time.Second)
+		gpuAdapter, adapterErr := models.NewOpenAICompatibleLocalAdapterWithAPIKeyAndTimeout(gpuURL, gpuKey, nil, gpuTimeout)
 		if adapterErr != nil {
 			log.Printf("optional GPU provider disabled: %v", adapterErr)
 		} else {

cmd/cognitive-worker/main_test.go

@@ -4,6 +4,7 @@ import (
 	"os"
 	"path/filepath"
 	"testing"
+	"time"
 )
 
 func TestLifeboatProviderIsOptionalAndLocalAdapterBacked(t *testing.T) {
@@ -19,6 +20,21 @@ func TestLifeboatProviderIsOptionalAndLocalAdapterBacked(t *testing.T) {
 	}
 }
 
+func TestGPUProviderTimeoutUsesBoundedConfigurableDefault(t *testing.T) {
+	t.Setenv("LOCAL_GPU_PROVIDER_TIMEOUT", "")
+	if got := durationEnv("LOCAL_GPU_PROVIDER_TIMEOUT", 60*time.Second); got != 60*time.Second {
+		t.Fatalf("default GPU timeout=%s", got)
+	}
+	t.Setenv("LOCAL_GPU_PROVIDER_TIMEOUT", "31s")
+	if got := durationEnv("LOCAL_GPU_PROVIDER_TIMEOUT", 60*time.Second); got != 31*time.Second {
+		t.Fatalf("configured GPU timeout=%s", got)
+	}
+	t.Setenv("LOCAL_GPU_PROVIDER_TIMEOUT", "invalid")
+	if got := durationEnv("LOCAL_GPU_PROVIDER_TIMEOUT", 60*time.Second); got != 60*time.Second {
+		t.Fatalf("invalid GPU timeout=%s", got)
+	}
+}
+
 func TestLoadLocalCPUProviderAPIKeyUsesProtectedFileReference(t *testing.T) {
 	t.Setenv("LOCAL_CPU_PROVIDER_API_KEY", "")
 	_ = os.Unsetenv("LOCAL_CPU_PROVIDER_API_KEY")

deploy/docker-compose.prod.yml

@@ -254,6 +254,7 @@ services:
       LOCAL_MEDIUM_PROVIDER_URL: ${LOCAL_MEDIUM_PROVIDER_URL:-http://llm-local-api:8080/v1}
       LOCAL_MEDIUM_PROVIDER_API_KEY_FILE: ${LOCAL_MEDIUM_PROVIDER_API_KEY_FILE:-}
       LOCAL_MEDIUM_PROVIDER_CREDENTIAL_REF: ${LOCAL_MEDIUM_PROVIDER_CREDENTIAL_REF:-credential://domovoy/local-medium/provider-api-key/v1}
+      LOCAL_GPU_PROVIDER_TIMEOUT: ${LOCAL_GPU_PROVIDER_TIMEOUT:-60s}
       LOCAL_LIFEBOAT_PROVIDER_URL: http://cognitive-lifeboat:8086/v1
       COGNITIVE_OWNER_PRINCIPAL_ID: domovoy-creator-principal
       COGNITIVE_ALLOW_CANARY: "true"

dev/ACTIVE_TASK.json

@@ -1,5 +1,5 @@
 {
-  "task_file": "dev/tasks/DEV-000267-refresh-gpu-and-ui-gap-evidence.json",
-  "task_id": "DEV-000267",
+  "task_file": "dev/tasks/DEV-000268-configurable-gpu-qualification-timeout.json",
+  "task_id": "DEV-000268",
   "task_revision": 1
 }

dev/tasks/DEV-000268-configurable-gpu-qualification-timeout.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000268",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "3ca7ffa",
+  "risk_tier": "E",
+  "intent": "Make the GPU cognitive provider invocation deadline configurable and bounded so large local models are not rejected by an accidental ten-second cold-start limit.",
+  "requirement_ids": ["model fleet/failover", "quality separated from reachability", "hardware-neutral routing"],
+  "allowed_change_scope": {"modules": ["cognitive-worker", "deployment-governance"], "paths": ["cmd/cognitive-worker/**", "deploy/docker-compose.prod.yml", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000268-configurable-gpu-qualification-timeout.json", "docs/evidence/**", "docs/PROJECT_STATUS.md"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "contracts/**", "LLM-local/**"]},
+  "allowed_architecture_delta": {"new_cross_module_dependencies": 0, "new_canonical_owners": 0, "new_privileges": 1, "new_third_party_dependencies": 0, "new_database_owner_roles": 0},
+  "contracts": {"consume": ["existing OpenAI-compatible local provider adapter", "existing qualification route"], "may_change": ["deployment timeout configuration only"]},
+  "semantic_owner": "cognitive-worker invokes provider; capability remains qualification and promotion owner",
+  "canonical_state": ["existing qualification observations"],
+  "state_transitions": ["provider timeout window changes; no qualification or activation transition is inferred"],
+  "side_effects_and_privileges": ["bounded outbound GPU provider request through existing credential boundary"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["timeout remains an unrated/unavailable observation; CPU/LIFEBOAT fallback remains authoritative"],
+  "assumptions": ["60 seconds is a bounded cold-start window and qualification models still run in parallel", "no model is promoted automatically"],
+  "acceptance_scenarios": ["GPU adapter uses configured bounded timeout", "default is 60 seconds", "qualification remains unrated on timeout", "routing does not infer intelligence from GPU"],
+  "required_tests": ["go test ./...", "go vet ./...", "go run ./cmd/devctl preflight -active dev/ACTIVE_TASK.json", "go run ./cmd/devctl check-diff -active dev/ACTIVE_TASK.json", "go run ./cmd/devctl check-architecture", "go run ./cmd/devctl/independent", "docker compose config", "git diff --check", "production health smoke", "one Domovoy-owned GPU qualification probe"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT"]
+}