feat(workflow-generator): enhance the AI auto-creation flow end-to-end (#38175)

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: autofix-ci[bot] <114827586+autofix-ci[bot]@users.noreply.github.com>
Co-authored-by: Copilot <198982749+Copilot@users.noreply.github.com>

api/controllers/console/app/generator.py

@@ -1,4 +1,5 @@
-from collections.abc import Sequence
+import json
+from collections.abc import Generator, Sequence
 from typing import Any, Literal
 
 from flask_restx import Resource
@@ -24,8 +25,10 @@ from core.helper.code_executor.javascript.javascript_code_provider import Javasc
 from core.helper.code_executor.python3.python3_code_provider import Python3CodeProvider
 from core.llm_generator.entities import RuleCodeGeneratePayload, RuleGeneratePayload, RuleStructuredOutputPayload
 from core.llm_generator.llm_generator import LLMGenerator
+from core.workflow.generator.types import WorkflowGenerateErrorCode
 from graphon.model_runtime.entities.llm_entities import LLMMode
 from graphon.model_runtime.errors.invoke import InvokeError
+from libs.helper import compact_generate_response
 from libs.login import login_required
 from models import App
 from services.workflow_generator_service import WorkflowGeneratorService
@@ -65,7 +68,10 @@ class WorkflowGeneratePayload(BaseModel):
     can reuse its existing handler.
     """
 
-    mode: Literal["workflow", "advanced-chat"] = Field(..., description="Target app mode for the generated graph")
+    mode: Literal["workflow", "advanced-chat", "auto"] = Field(
+        ...,
+        description="Target app mode for the generated graph; 'auto' lets the backend classify the instruction",
+    )
     instruction: str = Field(..., description="Natural-language workflow description")
     ideal_output: str = Field(default="", description="Optional sample output for grounding")
     model_config_data: ModelConfig = Field(
@@ -79,6 +85,19 @@ class WorkflowGeneratePayload(BaseModel):
     )
 
 
+class WorkflowInstructionSuggestionsPayload(BaseModel):
+    """Payload for the workflow-generator instruction-suggestions endpoint.
+
+    Runs before the user picks a model, so the suggestions come from the
+    tenant's default model. The underlying generator never raises — an empty
+    ``suggestions`` list is a valid 200 (soft-fail).
+    """
+
+    mode: Literal["workflow", "advanced-chat"] = Field(..., description="Target app mode for the suggestions")
+    language: str | None = Field(default=None, description="Optional language to write the suggestions in")
+    count: int = Field(default=4, ge=1, le=6, description="Number of suggestions to return (1-6)")
+
+
 class GeneratorResponse(RootModel[Any]):
     root: Any
 
@@ -92,6 +111,7 @@ register_schema_models(
     InstructionGeneratePayload,
     InstructionTemplatePayload,
     WorkflowGeneratePayload,
+    WorkflowInstructionSuggestionsPayload,
     ModelConfig,
 )
 register_response_schema_models(console_ns, GeneratorResponse, SimpleDataResponse)
@@ -316,6 +336,34 @@ class InstructionGenerationTemplateApi(Resource):
                 raise ValueError(f"Invalid type: {args.type}")
 
 
+def _workflow_instruction_guard(args: WorkflowGeneratePayload) -> tuple[dict, int] | None:
+    """Shared boundary guard for the workflow-generate endpoints.
+
+    Returns a ``(body, 400)`` tuple when the instruction is empty / whitespace
+    or either free-text field exceeds the cap, else ``None``. Pydantic only
+    validates the field is a str; a whitespace-only or pasted-document input
+    would otherwise waste a slow planner+builder roundtrip on a response the
+    validator rejects anyway. Both the blocking and streaming endpoints call
+    this so they reject identical inputs.
+    """
+    if not args.instruction.strip():
+        return {
+            "error": "Instruction is required",
+            "errors": [{"code": WorkflowGenerateErrorCode.EMPTY_INSTRUCTION, "detail": "Instruction is required"}],
+        }, 400
+    if len(args.instruction) > _MAX_INSTRUCTION_LENGTH or len(args.ideal_output) > _MAX_INSTRUCTION_LENGTH:
+        return {
+            "error": "Instruction is too long",
+            "errors": [
+                {
+                    "code": WorkflowGenerateErrorCode.INSTRUCTION_TOO_LONG,
+                    "detail": f"Instruction and ideal output must each be at most {_MAX_INSTRUCTION_LENGTH} characters",
+                }
+            ],
+        }, 400
+    return None
+
+
 @console_ns.route("/workflow-generate")
 class WorkflowGenerateApi(Resource):
     """Generate a Workflow / Chatflow draft graph from a natural-language description.
@@ -338,31 +386,11 @@ class WorkflowGenerateApi(Resource):
     def post(self, current_tenant_id: str):
         args = WorkflowGeneratePayload.model_validate(console_ns.payload)
 
-        # Reject obviously-empty instructions at the boundary — Pydantic only
-        # validates ``instruction`` is a str, but a whitespace-only string
-        # would still hit the LLM and waste a planner+builder roundtrip on a
-        # response that the postprocess validator would reject anyway.
-        if not args.instruction.strip():
-            return {
-                "error": "Instruction is required",
-                "errors": [{"code": "EMPTY_INSTRUCTION", "detail": "Instruction is required"}],
-            }, 400
-
-        # Bound the prompt at the boundary too: an arbitrarily long
-        # instruction (or pasted document) blows the planner/builder context
-        # window and fails with an opaque provider error after two slow LLM
-        # calls. The cap matches the frontend textarea's maxLength.
-        if len(args.instruction) > _MAX_INSTRUCTION_LENGTH or len(args.ideal_output) > _MAX_INSTRUCTION_LENGTH:
-            return {
-                "error": "Instruction is too long",
-                "errors": [
-                    {
-                        "code": "INSTRUCTION_TOO_LONG",
-                        "detail": f"Instruction and ideal output must each be at most "
-                        f"{_MAX_INSTRUCTION_LENGTH} characters",
-                    }
-                ],
-            }, 400
+        # Reject empty / over-length instructions at the boundary (shared with
+        # the streaming endpoint) before spending a planner+builder roundtrip.
+        guard = _workflow_instruction_guard(args)
+        if guard is not None:
+            return guard
 
         try:
             result = WorkflowGeneratorService.generate_workflow_graph(
@@ -383,3 +411,93 @@ class WorkflowGenerateApi(Resource):
             raise CompletionRequestError(e.description)
 
         return result
+
+
+@console_ns.route("/workflow-generate/suggestions")
+class WorkflowInstructionSuggestionsApi(Resource):
+    """Suggest short, buildable example instructions for the cmd+k generator.
+
+    Runs before a model is selected (uses the tenant's default model). The
+    underlying generator never raises, so an empty list is a valid 200 — the
+    frontend renders "no suggestions" rather than an error, so no provider-error
+    mapping is needed here.
+    """
+
+    @console_ns.doc("generate_workflow_instruction_suggestions")
+    @console_ns.doc(description="Suggest example workflow-generator instructions for the tenant")
+    @console_ns.expect(console_ns.models[WorkflowInstructionSuggestionsPayload.__name__])
+    @console_ns.response(200, "Suggestions generated successfully", console_ns.models[GeneratorResponse.__name__])
+    @console_ns.response(400, "Invalid request parameters")
+    @setup_required
+    @login_required
+    @account_initialization_required
+    @with_current_tenant_id
+    def post(self, current_tenant_id: str):
+        args = WorkflowInstructionSuggestionsPayload.model_validate(console_ns.payload)
+        suggestions = LLMGenerator.generate_workflow_instruction_suggestions(
+            tenant_id=current_tenant_id,
+            mode=args.mode,
+            language=args.language,
+            count=args.count,
+        )
+        return {"suggestions": suggestions}
+
+
+@console_ns.route("/workflow-generate/stream")
+class WorkflowGenerateStreamApi(Resource):
+    """Plan-first streaming variant of ``/workflow-generate`` (Server-Sent Events).
+
+    Emits a ``plan`` event (high-level node list + app metadata) as soon as the
+    planner returns, then a final ``result`` event with the full graph — the
+    SAME envelope ``/workflow-generate`` returns. Provider-init / invoke errors
+    are surfaced as a single ``result`` event (code ``MODEL_ERROR``) so the
+    frontend's stream parser always receives a result rather than a non-SSE HTTP
+    error.
+    """
+
+    @console_ns.doc("generate_workflow_graph_stream")
+    @console_ns.doc(description="Stream a Dify workflow graph (plan then result) via SSE")
+    @console_ns.expect(console_ns.models[WorkflowGeneratePayload.__name__])
+    @console_ns.response(200, "Server-Sent Events stream of plan/result events")
+    @console_ns.response(400, "Invalid request parameters")
+    @setup_required
+    @login_required
+    @account_initialization_required
+    @with_current_tenant_id
+    def post(self, current_tenant_id: str):
+        args = WorkflowGeneratePayload.model_validate(console_ns.payload)
+
+        # Same boundary guards as the blocking endpoint — return a normal 400
+        # JSON for these BEFORE opening the stream.
+        guard = _workflow_instruction_guard(args)
+        if guard is not None:
+            return guard
+
+        def generate() -> Generator[str, None, None]:
+            try:
+                for event_name, payload in WorkflowGeneratorService.generate_workflow_graph_stream(
+                    tenant_id=current_tenant_id,
+                    mode=args.mode,
+                    instruction=args.instruction,
+                    model_config=args.model_config_data,
+                    ideal_output=args.ideal_output,
+                    current_graph=args.current_graph,
+                ):
+                    body = {"event": event_name, **payload}
+                    yield f"data: {json.dumps(body)}\n\n"
+            except (ProviderTokenNotInitError, QuotaExceededError, ModelCurrentlyNotSupportError, InvokeError) as e:
+                # The model instance is resolved inside the service (lazily, on
+                # first iteration), so a provider / init error surfaces here.
+                # Emit it as a single SSE result event rather than a non-SSE
+                # error response so the frontend's stream parser always gets a
+                # result it can render.
+                detail = getattr(e, "description", None) or str(e) or "Model invocation failed"
+                error_body = {
+                    "event": "result",
+                    "graph": {"nodes": [], "edges": [], "viewport": {"x": 0.0, "y": 0.0, "zoom": 0.7}},
+                    "error": detail,
+                    "errors": [{"code": WorkflowGenerateErrorCode.MODEL_ERROR, "detail": detail}],
+                }
+                yield f"data: {json.dumps(error_body)}\n\n"
+
+        return compact_generate_response(generate())