Add a better Cody client server-sent configuration mechanism (#63591)

Signed-off-by: Stephen Gutekanst <stephen@sourcegraph.com>

client/cody-shared/src/sourcegraph-api/graphql/client.ts

@@ -151,6 +151,7 @@ export interface CodyLLMSiteConfiguration {
     completionModelMaxTokens?: number
     provider?: string
     smartContextWindow?: boolean
+    disableClientConfigAPI?: boolean
 }
 
 interface IsContextRequiredForChatQueryResponse {

cmd/frontend/graphqlbackend/schema.graphql

@@ -7496,6 +7496,11 @@ type CodyLLMConfiguration {
     """
     smartContextWindow: String!
     """
+    Disable Cody clients from using the new server-side config API. This is an escape-hatch for any issues
+    that may arise. This field will be removed in the future.
+    """
+    disableClientConfigAPI: Boolean!
+    """
     Name of the model being used for fast chat.
     """
     fastChatModel: String!

cmd/frontend/graphqlbackend/site.go

@@ -666,6 +666,9 @@ func (c *codyLLMConfigurationResolver) SmartContextWindow() string {
 	}
 	return "enabled"
 }
+func (c *codyLLMConfigurationResolver) DisableClientConfigAPI() bool {
+	return c.config.DisableClientConfigAPI
+}
 
 func (c *codyLLMConfigurationResolver) FastChatModel() string { return c.config.FastChatModel }
 func (c *codyLLMConfigurationResolver) FastChatModelMaxTokens() *int32 {

cmd/frontend/internal/clientconfig/BUILD.bazel

@@ -0,0 +1,38 @@
+load("//dev:go_defs.bzl", "go_test")
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "clientconfig",
+    srcs = [
+        "clientconfig.go",
+        "httpapi.go",
+    ],
+    importpath = "github.com/sourcegraph/sourcegraph/cmd/frontend/internal/clientconfig",
+    visibility = ["//cmd/frontend:__subpackages__"],
+    deps = [
+        "//cmd/frontend/internal/cody",
+        "//internal/actor",
+        "//internal/clientconfig",
+        "//internal/conf",
+        "//internal/database",
+        "@com_github_sourcegraph_log//:log",
+    ],
+)
+
+go_test(
+    name = "clientconfig_test",
+    srcs = ["httpapi_test.go"],
+    embed = [":clientconfig"],
+    tags = ["requires-network"],
+    deps = [
+        "//internal/actor",
+        "//internal/conf",
+        "//internal/database",
+        "//internal/database/dbtest",
+        "//internal/license",
+        "//internal/licensing",
+        "//schema",
+        "@com_github_hexops_autogold_v2//:autogold",
+        "@com_github_sourcegraph_log//logtest",
+    ],
+)

cmd/frontend/internal/clientconfig/clientconfig.go

@@ -0,0 +1,47 @@
+package clientconfig
+
+import (
+	"context"
+
+	"github.com/sourcegraph/log"
+
+	"github.com/sourcegraph/sourcegraph/cmd/frontend/internal/cody"
+	"github.com/sourcegraph/sourcegraph/internal/actor"
+	"github.com/sourcegraph/sourcegraph/internal/clientconfig"
+	"github.com/sourcegraph/sourcegraph/internal/conf"
+	"github.com/sourcegraph/sourcegraph/internal/database"
+)
+
+func GetForActor(ctx context.Context, logger log.Logger, db database.DB, actor *actor.Actor) (*clientconfig.ClientConfig, error) {
+	c := clientconfig.ClientConfig{
+		// TODO(chrsmith): TODO(slimsag): Set this to `true` when and only when clients should use
+		// the new LLM models httpapi endpoint being added in e.g. https://github.com/sourcegraph/sourcegraph/pull/63507
+		ModelsAPIEnabled: false,
+	}
+
+	// 🚨 SECURITY: This code lets site admins restrict who has access to Cody at all via RBAC.
+	// https://sourcegraph.com/docs/cody/clients/enable-cody-enterprise#enable-cody-only-for-some-users
+	c.CodyEnabled, _ = cody.IsCodyEnabled(ctx, db)
+
+	// 🚨 SECURITY: This code enforces that users do not have access to Cody features which
+	// site admins do not want them to have access to.
+	//
+	// Legacy admin-control configuration which should be moved to RBAC, not globally in site
+	// config. e.g. we should do it like https://github.com/sourcegraph/sourcegraph/pull/58831
+	features := conf.GetConfigFeatures(conf.Get().SiteConfig())
+	if features != nil { // nil -> Cody not enabled
+		c.ChatEnabled = features.Chat
+		c.AutoCompleteEnabled = features.AutoComplete
+		c.CustomCommandsEnabled = features.Commands
+		c.AttributionEnabled = features.Attribution
+	}
+
+	// Legacy feature-enablement configuration which should be moved to featureflag or RBAC,
+	// not exist in site config.
+	completionConfig := conf.GetCompletionsConfig(conf.Get().SiteConfig())
+	if completionConfig != nil { // nil -> Cody not enabled
+		c.SmartContextWindowEnabled = completionConfig.SmartContextWindow != "disabled"
+	}
+
+	return &c, nil
+}

cmd/frontend/internal/clientconfig/httpapi.go

@@ -0,0 +1,56 @@
+package clientconfig
+
+import (
+	"encoding/json"
+	"net/http"
+
+	"github.com/sourcegraph/log"
+
+	"github.com/sourcegraph/sourcegraph/internal/actor"
+	"github.com/sourcegraph/sourcegraph/internal/database"
+)
+
+// HTTP handlers for interacting with this Sourcegraph instance's
+// Cody client configuration. These handlers perform auth checks.
+type HTTPHandlers struct {
+	db     database.DB
+	logger log.Logger
+}
+
+func NewHandlers(db database.DB, logger log.Logger) *HTTPHandlers {
+	return &HTTPHandlers{
+		db:     db,
+		logger: logger,
+	}
+}
+
+// GetClientConfigHandler returns the current Sourcegraph instance's Cody client configuration
+// data as JSON. Requires that the calling user is an authenticated.
+func (h *HTTPHandlers) GetClientConfigHandler(w http.ResponseWriter, r *http.Request) {
+	ctx := r.Context()
+
+	h.logger.Info("fetching client config")
+
+	// Auth check.
+	callingActor := actor.FromContext(ctx)
+	if callingActor == nil || !callingActor.IsAuthenticated() {
+		h.logger.Warn("unauthenticated user requesting cody client config")
+		http.Error(w, http.StatusText(http.StatusUnauthorized), http.StatusUnauthorized)
+		return
+	}
+
+	currentConfig, err := GetForActor(r.Context(), h.logger, h.db, callingActor)
+	if err != nil {
+		h.logger.Error("fetching current cody client configuration", log.Error(err))
+		http.Error(w, http.StatusText(http.StatusInternalServerError), http.StatusInternalServerError)
+		return
+	}
+
+	rawJSON, err := json.MarshalIndent(currentConfig, "", "    ")
+	if err != nil {
+		h.logger.Error("marshalling configuration", log.Error(err))
+		http.Error(w, http.StatusText(http.StatusInternalServerError), http.StatusInternalServerError)
+		return
+	}
+	http.Error(w, string(rawJSON), http.StatusOK)
+}

cmd/frontend/internal/clientconfig/httpapi_test.go

@@ -0,0 +1,87 @@
+package clientconfig
+
+import (
+	"context"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"github.com/hexops/autogold/v2"
+	"github.com/sourcegraph/log/logtest"
+
+	"github.com/sourcegraph/sourcegraph/internal/actor"
+	"github.com/sourcegraph/sourcegraph/internal/conf"
+	"github.com/sourcegraph/sourcegraph/internal/database"
+	"github.com/sourcegraph/sourcegraph/internal/database/dbtest"
+	"github.com/sourcegraph/sourcegraph/internal/license"
+	"github.com/sourcegraph/sourcegraph/internal/licensing"
+	"github.com/sourcegraph/sourcegraph/schema"
+)
+
+func TestAPI(t *testing.T) {
+	logger := logtest.Scoped(t)
+	db := database.NewDB(logger, dbtest.NewDB(t))
+	ctx := context.Background()
+
+	// Enable Cody (and all other license features)
+	oldLicensingMock := licensing.MockCheckFeature
+	licensing.MockCheckFeature = func(feature licensing.Feature) error {
+		return nil
+	}
+	t.Cleanup(func() { licensing.MockCheckFeature = oldLicensingMock })
+
+	// Mock the site configuration
+	truePtr := true
+	falsePtr := false
+	licenseKey := "theasdfkey"
+	licenseAccessToken := license.GenerateLicenseKeyBasedAccessToken(licenseKey)
+	conf.Mock(&conf.Unified{
+		SiteConfiguration: schema.SiteConfiguration{
+			CodyEnabled:     &truePtr,
+			CodyPermissions: &falsePtr, // disable RBAC Cody permissions
+			Completions: &schema.Completions{
+				AccessToken: licenseAccessToken,
+			},
+		},
+	})
+	t.Cleanup(func() { conf.Mock(nil) })
+
+	// Grab HTTP handlers
+	handlers := NewHandlers(db, logger)
+
+	// Note: all the mechanics of conf.GetConfigFeatures, RBAC cody access via cody.IsCodyEnabled,
+	// and conf.GetCompletionsConfig are tested independently at their implementations. We really
+	// only test that those properties are relayed correctly by the HTTP API here.
+
+	t.Run("unauthenticated", func(t *testing.T) {
+		req, _ := http.NewRequest("GET", "", nil)
+		recorder := httptest.NewRecorder()
+		handlers.GetClientConfigHandler(recorder, req)
+
+		autogold.Expect(int(401)).Equal(t, recorder.Code)
+		autogold.Expect("Unauthorized\n").Equal(t, recorder.Body.String())
+	})
+
+	t.Run("authenticated_defaults", func(t *testing.T) {
+		req, _ := http.NewRequest("GET", "", nil)
+		req = req.WithContext(
+			actor.WithActor(ctx, &actor.Actor{
+				UID: 99,
+			}),
+		)
+		recorder := httptest.NewRecorder()
+		handlers.GetClientConfigHandler(recorder, req)
+
+		autogold.Expect(int(200)).Equal(t, recorder.Code)
+		autogold.Expect(`{
+    "codyEnabled": true,
+    "chatEnabled": true,
+    "autoCompleteEnabled": true,
+    "customCommandsEnabled": true,
+    "attributionEnabled": false,
+    "smartContextWindowEnabled": true,
+    "modelsAPIEnabled": false
+}
+`).Equal(t, recorder.Body.String())
+	})
+}

cmd/frontend/internal/httpapi/BUILD.bazel

@@ -26,6 +26,7 @@ go_library(
         "//cmd/frontend/backend",
         "//cmd/frontend/enterprise",
         "//cmd/frontend/graphqlbackend",
+        "//cmd/frontend/internal/clientconfig",
         "//cmd/frontend/internal/cody",
         "//cmd/frontend/internal/handlerutil",
         "//cmd/frontend/internal/httpapi/releasecache",

cmd/frontend/internal/httpapi/httpapi.go

@@ -22,6 +22,7 @@ import (
 	"github.com/sourcegraph/sourcegraph/cmd/frontend/backend"
 	"github.com/sourcegraph/sourcegraph/cmd/frontend/enterprise"
 	"github.com/sourcegraph/sourcegraph/cmd/frontend/graphqlbackend"
+	"github.com/sourcegraph/sourcegraph/cmd/frontend/internal/clientconfig"
 	"github.com/sourcegraph/sourcegraph/cmd/frontend/internal/handlerutil"
 	"github.com/sourcegraph/sourcegraph/cmd/frontend/internal/httpapi/releasecache"
 	"github.com/sourcegraph/sourcegraph/cmd/frontend/internal/httpapi/webhookhandlers"
@@ -174,6 +175,10 @@ func NewHandler(
 	m.Path("/completions/stream").Methods("POST").Handler(handlers.NewChatCompletionsStreamHandler())
 	m.Path("/completions/code").Methods("POST").Handler(handlers.NewCodeCompletionsHandler())
 
+	// HTTP endpoints related to Cody client configuration.
+	clientConfigHandlers := clientconfig.NewHandlers(db, logger)
+	m.Path("/client-config").Methods("GET").HandlerFunc(clientConfigHandlers.GetClientConfigHandler)
+
 	// HTTP endpoints related to LLM model configuration.
 	modelConfigHandlers := modelconfig.NewHandlers(db, logger)
 	m.Path("/modelconfig/supported-models.json").Methods("GET").HandlerFunc(modelConfigHandlers.GetSupportedModelsHandler)

internal/clientconfig/BUILD.bazel

@@ -0,0 +1,8 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "clientconfig",
+    srcs = ["types.go"],
+    importpath = "github.com/sourcegraph/sourcegraph/internal/clientconfig",
+    visibility = ["//:__subpackages__"],
+)

internal/clientconfig/types.go

@@ -0,0 +1,38 @@
+package clientconfig
+
+// This is the JSON object which all clients request after authentication to determine how
+// they should behave, e.g. if a site admin has restricted chat/autocomplete/other functionality,
+// if experimental features are available, etc.
+//
+// The configuration is always specific to a single authenticated user.
+//
+// Adding new fields here is fine, but you cannot make backwards-incompatible changes (removing
+// fields or change the meaning of fields in backwars-incompatible ways.) If you need to do that,
+// then read up on https://github.com/sourcegraph/sourcegraph/pull/63591#discussion_r1663211601
+//
+// After adding a field here, you can implement it in cmd/frontend/internal/clientconfig/clientconfig.go
+// GetForActor method.
+type ClientConfig struct {
+	// Whether the site admin allows this user to make use of Cody at all.
+	CodyEnabled bool `json:"codyEnabled"`
+
+	// Whether the site admin allows this user to make use of the Cody chat feature.
+	ChatEnabled bool `json:"chatEnabled"`
+
+	// Whether the site admin allows this user to make use of the Cody autocomplete feature.
+	AutoCompleteEnabled bool `json:"autoCompleteEnabled"`
+
+	// Whether the site admin allows the user to make use of the **custom** Cody commands feature.
+	CustomCommandsEnabled bool `json:"customCommandsEnabled"`
+
+	// Whether the site admin allows this user to make use of the Cody attribution feature.
+	AttributionEnabled bool `json:"attributionEnabled"`
+
+	// Whether the 'smart context window' feature should be enabled, and whether the Sourcegraph
+	// instance supports various new GraphQL APIs needed to make it work.
+	SmartContextWindowEnabled bool `json:"smartContextWindowEnabled"`
+
+	// Whether the new Sourcegraph backend LLM models API endpoint should be used to query which
+	// models are available.
+	ModelsAPIEnabled bool `json:"modelsAPIEnabled"`
+}

internal/conf/computed.go

@@ -931,12 +931,15 @@ func GetCompletionsConfig(siteConfig schema.SiteConfiguration) (c *conftypes.Com
 		completionsConfig.SmartContextWindow = "enabled"
 	}
 
+	disableClientConfigAPI := completionsConfig.DisableClientConfigAPI != nil && *completionsConfig.DisableClientConfigAPI
+
 	computedConfig := &conftypes.CompletionsConfig{
 		Provider:               conftypes.CompletionsProviderName(completionsConfig.Provider),
 		AccessToken:            completionsConfig.AccessToken,
 		ChatModel:              completionsConfig.ChatModel,
 		ChatModelMaxTokens:     completionsConfig.ChatModelMaxTokens,
 		SmartContextWindow:     completionsConfig.SmartContextWindow,
+		DisableClientConfigAPI: disableClientConfigAPI,
 		FastChatModel:          completionsConfig.FastChatModel,
 		FastChatModelMaxTokens: completionsConfig.FastChatModelMaxTokens,
 		AzureUseDeprecatedCompletionsAPIForOldModels: completionsConfig.AzureUseDeprecatedCompletionsAPIForOldModels,

internal/conf/computed_test.go

@@ -321,6 +321,7 @@ func TestGetCompletionsConfig(t *testing.T) {
 		ChatModel:                "anthropic/claude-3-sonnet-20240229",
 		ChatModelMaxTokens:       12000,
 		SmartContextWindow:       "enabled",
+		DisableClientConfigAPI:   false,
 		FastChatModel:            "anthropic/claude-3-haiku-20240307",
 		FastChatModelMaxTokens:   12000,
 		CompletionModel:          "fireworks/starcoder",
@@ -411,6 +412,7 @@ func TestGetCompletionsConfig(t *testing.T) {
 				ChatModel:                "claude-3-sonnet-20240229",
 				ChatModelMaxTokens:       12000,
 				SmartContextWindow:       "enabled",
+				DisableClientConfigAPI:   false,
 				FastChatModel:            "claude-3-haiku-20240307",
 				FastChatModelMaxTokens:   12000,
 				CompletionModel:          "claude-3-haiku-20240307",
@@ -426,18 +428,20 @@ func TestGetCompletionsConfig(t *testing.T) {
 				CodyEnabled: pointers.Ptr(true),
 				LicenseKey:  licenseKey,
 				Completions: &schema.Completions{
-					Enabled:            pointers.Ptr(true),
-					Provider:           "anthropic",
-					AccessToken:        "asdf",
-					ChatModel:          "claude-3-opus-20240229",
-					SmartContextWindow: "disabled",
-					CompletionModel:    "claude-instant-1.2",
+					Enabled:                pointers.Ptr(true),
+					Provider:               "anthropic",
+					AccessToken:            "asdf",
+					ChatModel:              "claude-3-opus-20240229",
+					SmartContextWindow:     "disabled",
+					DisableClientConfigAPI: pointers.Ptr(false),
+					CompletionModel:        "claude-instant-1.2",
 				},
 			},
 			wantConfig: &conftypes.CompletionsConfig{
 				ChatModel:                "claude-3-opus-20240229",
 				ChatModelMaxTokens:       12000,
 				SmartContextWindow:       "disabled",
+				DisableClientConfigAPI:   false,
 				FastChatModel:            "claude-3-haiku-20240307",
 				FastChatModelMaxTokens:   12000,
 				CompletionModel:          "claude-instant-1.2",
@@ -472,6 +476,7 @@ func TestGetCompletionsConfig(t *testing.T) {
 				ChatModel:                "gpt-4",
 				ChatModelMaxTokens:       7000,
 				SmartContextWindow:       "enabled",
+				DisableClientConfigAPI:   false,
 				FastChatModel:            "gpt-3.5-turbo",
 				FastChatModelMaxTokens:   16000,
 				CompletionModel:          "gpt-3.5-turbo-instruct",
@@ -487,19 +492,21 @@ func TestGetCompletionsConfig(t *testing.T) {
 				CodyEnabled: pointers.Ptr(true),
 				LicenseKey:  licenseKey,
 				Completions: &schema.Completions{
-					Provider:           "azure-openai",
-					AccessToken:        "asdf",
-					Endpoint:           "https://acmecorp.openai.azure.com",
-					ChatModel:          "gpt4-deployment",
-					SmartContextWindow: "disabled",
-					FastChatModel:      "gpt35-turbo-deployment",
-					CompletionModel:    "gpt35-turbo-deployment",
+					Provider:               "azure-openai",
+					AccessToken:            "asdf",
+					Endpoint:               "https://acmecorp.openai.azure.com",
+					ChatModel:              "gpt4-deployment",
+					SmartContextWindow:     "disabled",
+					DisableClientConfigAPI: pointers.Ptr(false),
+					FastChatModel:          "gpt35-turbo-deployment",
+					CompletionModel:        "gpt35-turbo-deployment",
 				},
 			},
 			wantConfig: &conftypes.CompletionsConfig{
 				ChatModel:                "gpt4-deployment",
 				ChatModelMaxTokens:       7000,
 				SmartContextWindow:       "disabled",
+				DisableClientConfigAPI:   false,
 				FastChatModel:            "gpt35-turbo-deployment",
 				FastChatModelMaxTokens:   7000,
 				CompletionModel:          "gpt35-turbo-deployment",
@@ -523,6 +530,7 @@ func TestGetCompletionsConfig(t *testing.T) {
 				ChatModel:                "accounts/fireworks/models/llama-v2-7b",
 				ChatModelMaxTokens:       3000,
 				SmartContextWindow:       "enabled",
+				DisableClientConfigAPI:   false,
 				FastChatModel:            "accounts/fireworks/models/llama-v2-7b",
 				FastChatModelMaxTokens:   3000,
 				CompletionModel:          "starcoder",
@@ -546,6 +554,7 @@ func TestGetCompletionsConfig(t *testing.T) {
 				ChatModel:                "anthropic.claude-v2",
 				ChatModelMaxTokens:       12000,
 				SmartContextWindow:       "enabled",
+				DisableClientConfigAPI:   false,
 				FastChatModel:            "anthropic.claude-instant-v1",
 				FastChatModelMaxTokens:   9000,
 				CompletionModel:          "anthropic.claude-instant-v1",
@@ -571,6 +580,7 @@ func TestGetCompletionsConfig(t *testing.T) {
 				ChatModel:                "anthropic.claude-3-haiku-20240307-v1:0-100k/arn:aws:bedrock:us-west-2:012345678901:provisioned-model/abcdefghijkl",
 				ChatModelMaxTokens:       100_000,
 				SmartContextWindow:       "enabled",
+				DisableClientConfigAPI:   false,
 				FastChatModel:            "anthropic.claude-v2",
 				FastChatModelMaxTokens:   12000,
 				CompletionModel:          "anthropic.claude-instant-v1",
@@ -611,6 +621,7 @@ func TestGetCompletionsConfig(t *testing.T) {
 				ChatModel:                "anthropic/claude-v1.3",
 				ChatModelMaxTokens:       9000,
 				SmartContextWindow:       "enabled",
+				DisableClientConfigAPI:   false,
 				FastChatModel:            "anthropic/claude-instant-1.3",
 				FastChatModelMaxTokens:   9000,
 				CompletionModel:          "anthropic/claude-instant-1.3",

internal/conf/conftypes/consts.go

@@ -6,7 +6,8 @@ type CompletionsConfig struct {
 	ChatModel          string
 	ChatModelMaxTokens int
 
-	SmartContextWindow string
+	SmartContextWindow     string
+	DisableClientConfigAPI bool
 
 	FastChatModel          string
 	FastChatModelMaxTokens int

schema/schema.go

@@ -680,6 +680,8 @@ type Completions struct {
 	CompletionModel string `json:"completionModel,omitempty"`
 	// CompletionModelMaxTokens description: The maximum number of tokens to use as client when talking to completionModel. If not set, clients need to set their own limit.
 	CompletionModelMaxTokens int `json:"completionModelMaxTokens,omitempty"`
+	// DisableClientConfigAPI description: Should not be set. If set to true, disables the use of the new client config API. This new API has no user-facing effect, this opt-out is provided only as an escape hatch in case of issues.
+	DisableClientConfigAPI *bool `json:"disableClientConfigAPI,omitempty"`
 	// Enabled description: DEPRECATED. Use cody.enabled instead to turn Cody on/off.
 	Enabled *bool `json:"enabled,omitempty"`
 	// Endpoint description: The endpoint under which to reach the provider. Currently only used for provider types "sourcegraph", "openai" and "anthropic". The default values are "https://cody-gateway.sourcegraph.com", "https://api.openai.com/v1/chat/completions", and "https://api.anthropic.com/v1/messages" for Sourcegraph, OpenAI, and Anthropic, respectively.

schema/site.schema.json

@@ -2926,6 +2926,14 @@
           "default": "enabled",
           "enum": ["enabled", "disabled"]
         },
+        "disableClientConfigAPI": {
+          "description": "Should not be set. If set to true, disables the use of the new client config API. This new API has no user-facing effect, this opt-out is provided only as an escape hatch in case of issues.",
+          "type": "boolean",
+          "!go": {
+            "pointer": true
+          },
+          "deprecationMessage": "This opt-out feature flag will be removed soon."
+        },
         "completionModel": {
           "description": "The model used for code completion. If using the default provider 'sourcegraph', a reasonable default model will be set.\n NOTE: The Anthropic messages API does not support model names like claude-2 or claude-instant-1 where only the major version is specified as they are retired. We recommend using a specific model identifier as specified here https://docs.anthropic.com/claude/docs/models-overview#model-comparison ",
           "type": "string",