feat: redesign chat UI and add SSE token streaming

.claude/launch.json

@@ -0,0 +1,23 @@
+{
+  "version": "0.0.1",
+  "configurations": [
+    {
+      "name": "Web Frontend",
+      "runtimeExecutable": "npm",
+      "runtimeArgs": ["--prefix", "apps/web", "run", "dev"],
+      "port": 5173
+    },
+    {
+      "name": "Backend API",
+      "runtimeExecutable": "bash",
+      "runtimeArgs": ["-c", "cd apps/backend && make all && make run"],
+      "port": 8007
+    },
+    {
+      "name": "Backend API (live reload)",
+      "runtimeExecutable": "bash",
+      "runtimeArgs": ["-c", "cd apps/backend && air"],
+      "port": 8007
+    }
+  ]
+}

.claude/settings.local.json

@@ -0,0 +1,10 @@
+{
+  "permissions": {
+    "allow": [
+      "Bash(go version *)",
+      "Bash(npm --version)",
+      "Bash(docker --version)",
+      "Bash(ollama --version)"
+    ]
+  }
+}

apps/backend/servers/echo/handlers/http.go

@@ -117,6 +117,7 @@ func Error(c echo.Context, logger *slog.Logger, err error) error {
 	return echo.NewHTTPError(status, ErrorResponse{Error: message})
 }
 
+
 func logInternalError(c echo.Context, logger *slog.Logger, err error) {
 	if logger == nil {
 		logger = slog.Default()

apps/backend/services/chat/internal/chat/llm/generate_sql_helpers.go

@@ -5,6 +5,7 @@ import (
 	"fmt"
 	"strings"
 
+	chaterrors "github.com/Uncensored-Developer/datalk/apps/backend/services/chat/pkg/errors"
 	llmtypes "github.com/Uncensored-Developer/datalk/apps/backend/services/chat/pkg/llm"
 	schematypes "github.com/Uncensored-Developer/datalk/apps/backend/services/schemas/pkg/schemas"
 	"github.com/mdobak/go-xerrors"
@@ -114,7 +115,11 @@ func ParseGenerateSQLResponse(rawRequest, rawResponse []byte, payloadText string
 	payload.SQL = strings.TrimSpace(payload.SQL)
 	payload.Explanation = strings.TrimSpace(payload.Explanation)
 	if payload.SQL == "" {
-		return nil, xerrors.New("structured SQL payload did not include sql")
+		explanation := strings.TrimSpace(payload.Explanation)
+		if explanation != "" {
+			return nil, xerrors.Newf("%s: %w", explanation, chaterrors.ErrInvalidSQL)
+		}
+		return nil, xerrors.Newf("the model could not generate SQL for this question: %w", chaterrors.ErrInvalidSQL)
 	}
 
 	return &llmtypes.GenerateSQLResponse{

apps/backend/services/chat/internal/chat/llm/openai/client.go

@@ -338,11 +338,18 @@ func openAIInputMessages(req llmtypes.GenerateSQLRequest) []inputMessage {
 	})
 
 	for _, message := range promptMessages {
+		role := normalizeOpenAIRole(message.Role)
+		// OpenAI Responses API requires "output_text" for assistant turns
+		// and "input_text" for user/developer turns.
+		contentType := "input_text"
+		if role == "assistant" {
+			contentType = "output_text"
+		}
 		messages = append(messages, inputMessage{
-			Role: normalizeOpenAIRole(message.Role),
+			Role: role,
 			Content: []contentPart{
 				{
-					Type: "input_text",
+					Type: contentType,
 					Text: message.Content,
 				},
 			},

apps/backend/services/chat/internal/chat/llm/registry.go

@@ -246,17 +246,22 @@ func normalizeProviderModels(provider llmtypes.Provider, models []llmtypes.Model
 
 		qualifiedModelID := QualifiedModelID(provider, rawModelID)
 		if strings.Contains(rawModelID, ":") {
+			// Only treat the colon as a provider:model separator if the part
+			// before it is actually a known provider. Ollama model IDs use ":"
+			// as a tag separator (e.g. "llama3.2:1b"), which must not be
+			// mistaken for a qualified provider prefix.
 			parsedProvider, parsedModelID, err := ParseQualifiedModelID(rawModelID)
-			if err != nil {
-				return nil, err
+			if err == nil {
+				// Providers are allowed to return already-qualified ids, but the
+				// embedded provider prefix still has to match the provider being queried.
+				if parsedProvider != provider {
+					return nil, xerrors.Newf("provider %s returned model id for different provider %s", provider, parsedProvider)
+				}
+				qualifiedModelID = QualifiedModelID(parsedProvider, parsedModelID)
+				rawModelID = parsedModelID
 			}
-			// Providers are allowed to return already-qualified ids, but the
-			// embedded provider prefix still has to match the provider being queried.
-			if parsedProvider != provider {
-				return nil, xerrors.Newf("provider %s returned model id for different provider %s", provider, parsedProvider)
-			}
-			qualifiedModelID = QualifiedModelID(parsedProvider, parsedModelID)
-			rawModelID = parsedModelID
+			// If parsing failed the colon is part of the model's own id (e.g. an
+			// Ollama tag). Fall through and use the already-computed qualifiedModelID.
 		}
 
 		// Deduplicate after normalization so "gpt-5.2" and "openai:gpt-5.2"

apps/web/src/components/layout/AppShell.tsx

@@ -242,7 +242,7 @@ export function AppShell({ title, children }: AppShellProps) {
           </Toolbar>
         </Box>
 
-        <Box component="main" sx={{ p: { xs: 2, sm: 3 }, maxWidth: 1280 }}>
+        <Box component="main" sx={{ p: { xs: 2, sm: 3 }, maxWidth: 1280, mx: "auto", width: "100%" }}>
           {children}
         </Box>
       </Box>