tactus 0.31.2__py3-none-any.whl

tactus/ide/server.py

@@ -0,0 +1,2223 @@
+"""
+Tactus IDE Backend Server.
+
+Provides HTTP-based LSP server for the Tactus IDE.
+"""
+
+import json
+import logging
+import os
+import queue
+import subprocess
+import threading
+import time
+from datetime import datetime
+from pathlib import Path
+from flask import Flask, request, jsonify, Response, stream_with_context
+from flask_cors import CORS
+from typing import Dict, Any, List, Optional
+
+from tactus.validation.validator import TactusValidator, ValidationMode
+from tactus.core.registry import ValidationMessage
+
+logger = logging.getLogger(__name__)
+
+# Workspace state
+WORKSPACE_ROOT = None
+
+
+class TactusLSPHandler:
+    """LSP handler for Tactus DSL."""
+
+    def __init__(self):
+        self.validator = TactusValidator()
+        self.documents: Dict[str, str] = {}
+        self.registries: Dict[str, Any] = {}
+
+    def validate_document(self, uri: str, text: str) -> List[Dict[str, Any]]:
+        """Validate document and return LSP diagnostics."""
+        self.documents[uri] = text
+
+        try:
+            result = self.validator.validate(text, ValidationMode.FULL)
+
+            if result.registry:
+                self.registries[uri] = result.registry
+
+            diagnostics = []
+            for error in result.errors:
+                diagnostic = self._convert_to_diagnostic(error, "Error")
+                if diagnostic:
+                    diagnostics.append(diagnostic)
+
+            for warning in result.warnings:
+                diagnostic = self._convert_to_diagnostic(warning, "Warning")
+                if diagnostic:
+                    diagnostics.append(diagnostic)
+
+            return diagnostics
+        except Exception as e:
+            logger.error(f"Error validating document {uri}: {e}", exc_info=True)
+            return []
+
+    def _convert_to_diagnostic(
+        self, message: ValidationMessage, severity_str: str
+    ) -> Optional[Dict[str, Any]]:
+        """Convert ValidationMessage to LSP diagnostic."""
+        severity = 1 if severity_str == "Error" else 2
+
+        line = message.location[0] - 1 if message.location else 0
+        col = message.location[1] - 1 if message.location and len(message.location) > 1 else 0
+
+        return {
+            "range": {
+                "start": {"line": line, "character": col},
+                "end": {"line": line, "character": col + 10},
+            },
+            "severity": severity,
+            "source": "tactus",
+            "message": message.message,
+        }
+
+    def close_document(self, uri: str):
+        """Close a document."""
+        self.documents.pop(uri, None)
+        self.registries.pop(uri, None)
+
+
+class LSPServer:
+    """Language Server Protocol server for Tactus DSL."""
+
+    def __init__(self):
+        self.handler = TactusLSPHandler()
+        self.initialized = False
+        self.client_capabilities = {}
+
+    def handle_message(self, message: Dict[str, Any]) -> Optional[Dict[str, Any]]:
+        """Handle LSP JSON-RPC message."""
+        method = message.get("method")
+        params = message.get("params", {})
+        msg_id = message.get("id")
+
+        try:
+            if method == "initialize":
+                result = self._handle_initialize(params)
+            else:
+                logger.warning(f"Unhandled LSP method: {method}")
+                return self._error_response(msg_id, -32601, f"Method not found: {method}")
+
+            if msg_id is not None:
+                return {"jsonrpc": "2.0", "id": msg_id, "result": result}
+        except Exception as e:
+            logger.error(f"Error handling {method}: {e}", exc_info=True)
+            return self._error_response(msg_id, -32603, str(e))
+
+    def _handle_initialize(self, params: Dict[str, Any]) -> Dict[str, Any]:
+        """Handle initialize request."""
+        self.client_capabilities = params.get("capabilities", {})
+        self.initialized = True
+
+        return {
+            "capabilities": {
+                "textDocumentSync": {"openClose": True, "change": 2, "save": {"includeText": True}},
+                "diagnosticProvider": {
+                    "interFileDependencies": False,
+                    "workspaceDiagnostics": False,
+                },
+            },
+            "serverInfo": {"name": "tactus-lsp-server", "version": "0.1.0"},
+        }
+
+    def _error_response(self, msg_id: Optional[int], code: int, message: str) -> Dict[str, Any]:
+        """Create LSP error response."""
+        return {"jsonrpc": "2.0", "id": msg_id, "error": {"code": code, "message": message}}
+
+
+def _resolve_workspace_path(relative_path: str) -> Path:
+    """
+    Resolve a relative path within the workspace root.
+    Raises ValueError if path escapes workspace or workspace not set.
+    """
+    global WORKSPACE_ROOT
+
+    if not WORKSPACE_ROOT:
+        raise ValueError("No workspace folder selected")
+
+    # Normalize the relative path
+    workspace = Path(WORKSPACE_ROOT).resolve()
+    target = (workspace / relative_path).resolve()
+
+    # Ensure target is within workspace (prevent path traversal)
+    try:
+        target.relative_to(workspace)
+    except ValueError:
+        raise ValueError(f"Path '{relative_path}' escapes workspace")
+
+    return target
+
+
+def create_app(initial_workspace: Optional[str] = None, frontend_dist_dir: Optional[str] = None):
+    """Create and configure the Flask app.
+
+    Args:
+        initial_workspace: Initial workspace directory. If not provided, uses current directory.
+        frontend_dist_dir: Path to frontend dist directory. If provided, serves frontend from Flask.
+    """
+    global WORKSPACE_ROOT
+
+    # Configure Flask to serve frontend static files if provided
+    if frontend_dist_dir:
+        app = Flask(__name__, static_folder=frontend_dist_dir, static_url_path="")
+    else:
+        app = Flask(__name__)
+    CORS(app)
+
+    # Set initial workspace if provided
+    if initial_workspace:
+        WORKSPACE_ROOT = str(Path(initial_workspace).resolve())
+
+    # Initialize LSP server
+    lsp_server = LSPServer()
+
+    @app.route("/health", methods=["GET"])
+    def health():
+        """Health check endpoint."""
+        return jsonify({"status": "ok", "service": "tactus-ide-backend"})
+
+    @app.route("/api/workspace/cwd", methods=["GET"])
+    def get_cwd():
+        """Get current working directory (returns the initial workspace if set)."""
+        if WORKSPACE_ROOT:
+            return jsonify({"cwd": WORKSPACE_ROOT})
+        return jsonify({"cwd": str(Path.cwd())})
+
+    @app.route("/api/about", methods=["GET"])
+    def get_about_info():
+        """Get application version and metadata."""
+        from tactus import __version__
+
+        return jsonify(
+            {
+                "version": __version__,
+                "name": "Tactus IDE",
+                "description": "A Lua-based DSL for agentic workflows",
+                "author": "Ryan Porter",
+                "license": "MIT",
+                "repository": "https://github.com/AnthusAI/Tactus",
+                "documentation": "https://github.com/AnthusAI/Tactus/tree/main/docs",
+                "issues": "https://github.com/AnthusAI/Tactus/issues",
+            }
+        )
+
+    @app.route("/api/workspace", methods=["GET", "POST"])
+    def workspace_operations():
+        """Handle workspace operations."""
+        global WORKSPACE_ROOT
+
+        if request.method == "GET":
+            if not WORKSPACE_ROOT:
+                return jsonify({"root": None, "name": None})
+
+            workspace_path = Path(WORKSPACE_ROOT)
+            return jsonify({"root": str(workspace_path), "name": workspace_path.name})
+
+        elif request.method == "POST":
+            data = request.json
+            root = data.get("root")
+
+            if not root:
+                return jsonify({"error": "Missing 'root' parameter"}), 400
+
+            try:
+                root_path = Path(root).resolve()
+
+                if not root_path.exists():
+                    return jsonify({"error": f"Path does not exist: {root}"}), 404
+
+                if not root_path.is_dir():
+                    return jsonify({"error": f"Path is not a directory: {root}"}), 400
+
+                # Set workspace root and change working directory
+                WORKSPACE_ROOT = str(root_path)
+                os.chdir(WORKSPACE_ROOT)
+
+                logger.info(f"Workspace set to: {WORKSPACE_ROOT}")
+
+                return jsonify({"success": True, "root": WORKSPACE_ROOT, "name": root_path.name})
+            except Exception as e:
+                logger.error(f"Error setting workspace {root}: {e}")
+                return jsonify({"error": str(e)}), 500
+
+    @app.route("/api/tree", methods=["GET"])
+    def tree_operations():
+        """List directory contents within the workspace."""
+        global WORKSPACE_ROOT
+
+        if not WORKSPACE_ROOT:
+            return jsonify({"error": "No workspace folder selected"}), 400
+
+        relative_path = request.args.get("path", "")
+
+        try:
+            target_path = _resolve_workspace_path(relative_path)
+
+            if not target_path.exists():
+                return jsonify({"error": f"Path not found: {relative_path}"}), 404
+
+            if not target_path.is_dir():
+                return jsonify({"error": f"Path is not a directory: {relative_path}"}), 400
+
+            # List directory contents
+            entries = []
+            for item in sorted(
+                target_path.iterdir(), key=lambda x: (not x.is_dir(), x.name.lower())
+            ):
+                entry = {
+                    "name": item.name,
+                    "path": str(item.relative_to(WORKSPACE_ROOT)),
+                    "type": "directory" if item.is_dir() else "file",
+                }
+
+                # Add extension for files
+                if item.is_file():
+                    entry["extension"] = item.suffix
+
+                entries.append(entry)
+
+            return jsonify({"path": relative_path, "entries": entries})
+        except ValueError as e:
+            return jsonify({"error": str(e)}), 400
+        except Exception as e:
+            logger.error(f"Error listing directory {relative_path}: {e}")
+            return jsonify({"error": str(e)}), 500
+
+    @app.route("/api/file", methods=["GET", "POST"])
+    def file_operations():
+        """Handle file operations (read/write files within workspace)."""
+        if request.method == "GET":
+            file_path = request.args.get("path")
+            if not file_path:
+                return jsonify({"error": "Missing 'path' parameter"}), 400
+
+            try:
+                path = _resolve_workspace_path(file_path)
+
+                if not path.exists():
+                    return jsonify({"error": f"File not found: {file_path}"}), 404
+
+                if not path.is_file():
+                    return jsonify({"error": f"Path is not a file: {file_path}"}), 400
+
+                content = path.read_text()
+                return jsonify(
+                    {
+                        "path": file_path,
+                        "absolutePath": str(path),
+                        "content": content,
+                        "name": path.name,
+                    }
+                )
+            except ValueError as e:
+                return jsonify({"error": str(e)}), 400
+            except Exception as e:
+                logger.error(f"Error reading file {file_path}: {e}")
+                return jsonify({"error": str(e)}), 500
+
+        elif request.method == "POST":
+            data = request.json
+            file_path = data.get("path")
+            content = data.get("content")
+
+            if not file_path or content is None:
+                return jsonify({"error": "Missing 'path' or 'content'"}), 400
+
+            try:
+                path = _resolve_workspace_path(file_path)
+
+                # Create parent directories if needed
+                path.parent.mkdir(parents=True, exist_ok=True)
+                path.write_text(content)
+
+                return jsonify({"success": True, "path": file_path, "absolutePath": str(path)})
+            except ValueError as e:
+                return jsonify({"error": str(e)}), 400
+            except Exception as e:
+                logger.error(f"Error writing file {file_path}: {e}")
+                return jsonify({"error": str(e)}), 500
+
+    @app.route("/api/procedure/metadata", methods=["GET"])
+    def get_procedure_metadata():
+        """
+        Get metadata about a procedure file using TactusValidator.
+
+        Query params:
+        - path: workspace-relative path to procedure file (required)
+
+        Returns:
+            {
+                "success": true,
+                "metadata": {
+                    "description": str | null,
+                    "input": { name: ParameterDeclaration },
+                    "output": { name: OutputFieldDeclaration },
+                    "agents": { name: AgentDeclaration },
+                    "toolsets": { name: dict },
+                    "tools": [str]  # Flattened list of all tools
+                }
+            }
+        """
+        file_path = request.args.get("path")
+
+        if not file_path:
+            return jsonify({"error": "Missing 'path' parameter"}), 400
+
+        try:
+            # Resolve path
+            path = _resolve_workspace_path(file_path)
+
+            if not path.exists():
+                return jsonify({"error": f"File not found: {file_path}"}), 404
+
+            # Validate with FULL mode to get registry
+            validator = TactusValidator()
+            result = validator.validate_file(str(path), ValidationMode.FULL)
+
+            if not result.registry:
+                # Validation failed or no registry
+                return (
+                    jsonify(
+                        {
+                            "success": False,
+                            "error": "Failed to extract metadata",
+                            "validation_errors": [
+                                {
+                                    "message": e.message,
+                                    "line": e.location[0] if e.location else None,
+                                }
+                                for e in result.errors
+                            ],
+                        }
+                    ),
+                    400,
+                )
+
+            registry = result.registry
+
+            # Extract tools from agents, toolsets, and lua_tools
+            all_tools = set()
+            for agent in registry.agents.values():
+                all_tools.update(agent.tools)
+            for toolset in registry.toolsets.values():
+                if isinstance(toolset, dict) and "tools" in toolset:
+                    all_tools.update(toolset["tools"])
+            # Include Lua-defined tools (from tool() function calls)
+            if hasattr(registry, "lua_tools") and registry.lua_tools:
+                all_tools.update(registry.lua_tools.keys())
+
+            # Parse specifications if present
+            specifications_data = None
+            if registry.gherkin_specifications:
+                import re
+
+                gherkin_text = registry.gherkin_specifications
+
+                # Count scenarios
+                scenarios = re.findall(r"^\s*Scenario:", gherkin_text, re.MULTILINE)
+
+                # Extract feature name
+                feature_match = re.search(r"Feature:\s*(.+)", gherkin_text)
+                feature_name = feature_match.group(1).strip() if feature_match else None
+
+                specifications_data = {
+                    "text": gherkin_text,
+                    "feature_name": feature_name,
+                    "scenario_count": len(scenarios),
+                }
+
+            # Extract evaluations summary
+            evaluations_data = None
+            if registry.pydantic_evaluations:
+                evals = registry.pydantic_evaluations
+                dataset_count = 0
+                evaluator_count = 0
+
+                if isinstance(evals, dict):
+                    # Count dataset items
+                    if "dataset" in evals and isinstance(evals["dataset"], list):
+                        dataset_count = len(evals["dataset"])
+
+                    # Count evaluators
+                    if "evaluators" in evals and isinstance(evals["evaluators"], list):
+                        evaluator_count = len(evals["evaluators"])
+
+                evaluations_data = {
+                    "dataset_count": dataset_count,
+                    "evaluator_count": evaluator_count,
+                    "runs": evals.get("runs", 1) if isinstance(evals, dict) else 1,
+                    "parallel": evals.get("parallel", False) if isinstance(evals, dict) else False,
+                }
+
+            # Build metadata response
+            metadata = {
+                "description": registry.description,
+                "input": registry.input_schema if registry.input_schema else {},
+                "output": registry.output_schema if registry.output_schema else {},
+                "agents": {
+                    name: {
+                        "name": agent.name,
+                        "provider": agent.provider,
+                        "model": agent.model if isinstance(agent.model, str) else str(agent.model),
+                        "system_prompt": (
+                            agent.system_prompt
+                            if isinstance(agent.system_prompt, str)
+                            else "[Dynamic Prompt]"
+                        ),
+                        "tools": agent.tools,
+                    }
+                    for name, agent in registry.agents.items()
+                },
+                "toolsets": {name: toolset for name, toolset in registry.toolsets.items()},
+                "tools": sorted(list(all_tools)),
+                "specifications": specifications_data,
+                "evaluations": evaluations_data,
+            }
+
+            return jsonify({"success": True, "metadata": metadata})
+
+        except Exception as e:
+            logger.error(f"Error extracting procedure metadata: {e}", exc_info=True)
+            return jsonify({"error": str(e)}), 500
+
+    @app.route("/api/validate", methods=["POST"])
+    def validate_procedure():
+        """Validate Tactus procedure code."""
+        data = request.json
+        content = data.get("content")
+
+        if content is None:
+            return jsonify({"error": "Missing 'content' parameter"}), 400
+
+        try:
+            validator = TactusValidator()
+            result = validator.validate(content)
+
+            return jsonify(
+                {
+                    "valid": result.valid,
+                    "errors": [
+                        {
+                            "message": err.message,
+                            "line": err.location[0] if err.location else None,
+                            "column": err.location[1] if err.location else None,
+                            "level": err.level,
+                        }
+                        for err in result.errors
+                    ],
+                    "warnings": [
+                        {
+                            "message": warn.message,
+                            "line": warn.location[0] if warn.location else None,
+                            "column": warn.location[1] if warn.location else None,
+                            "level": warn.level,
+                        }
+                        for warn in result.warnings
+                    ],
+                }
+            )
+        except Exception as e:
+            logger.error(f"Error validating code: {e}")
+            return jsonify({"error": str(e)}), 500
+
+    @app.route("/api/validate/stream", methods=["GET"])
+    def validate_stream():
+        """Validate Tactus code with SSE streaming output."""
+        file_path = request.args.get("path")
+
+        if not file_path:
+            return jsonify({"error": "Missing 'path' parameter"}), 400
+
+        try:
+            # Resolve path within workspace
+            path = _resolve_workspace_path(file_path)
+
+            # Ensure file exists
+            if not path.exists():
+                return jsonify({"error": f"File not found: {file_path}"}), 404
+
+            def generate_events():
+                """Generator function that yields SSE validation events."""
+                try:
+                    import json
+                    from datetime import datetime
+
+                    # Read and validate file
+                    content = path.read_text()
+                    validator = TactusValidator()
+                    result = validator.validate(content)
+
+                    # Emit validation event
+                    validation_event = {
+                        "event_type": "validation",
+                        "valid": result.valid,
+                        "errors": [
+                            {
+                                "message": err.message,
+                                "line": err.location[0] if err.location else None,
+                                "column": err.location[1] if err.location else None,
+                                "level": err.level,
+                            }
+                            for err in result.errors
+                        ],
+                        "warnings": [
+                            {
+                                "message": warn.message,
+                                "line": warn.location[0] if warn.location else None,
+                                "column": warn.location[1] if warn.location else None,
+                                "level": warn.level,
+                            }
+                            for warn in result.warnings
+                        ],
+                        "timestamp": datetime.utcnow().isoformat() + "Z",
+                    }
+                    yield f"data: {json.dumps(validation_event)}\n\n"
+
+                except Exception as e:
+                    logger.error(f"Error in validation: {e}", exc_info=True)
+                    error_event = {
+                        "event_type": "execution",
+                        "lifecycle_stage": "error",
+                        "timestamp": datetime.utcnow().isoformat() + "Z",
+                        "details": {"error": str(e)},
+                    }
+                    yield f"data: {json.dumps(error_event)}\n\n"
+
+            return Response(
+                stream_with_context(generate_events()),
+                mimetype="text/event-stream",
+                headers={
+                    "Cache-Control": "no-cache",
+                    "X-Accel-Buffering": "no",
+                    "Connection": "keep-alive",
+                },
+            )
+
+        except ValueError as e:
+            return jsonify({"error": str(e)}), 400
+        except Exception as e:
+            logger.error(f"Error setting up validation: {e}", exc_info=True)
+            return jsonify({"error": str(e)}), 500
+
+    @app.route("/api/run", methods=["POST"])
+    def run_procedure():
+        """Run a Tactus procedure."""
+        data = request.json
+        file_path = data.get("path")
+        content = data.get("content")
+
+        if not file_path:
+            return jsonify({"error": "Missing 'path' parameter"}), 400
+
+        try:
+            # Resolve path within workspace
+            path = _resolve_workspace_path(file_path)
+
+            # Save content if provided
+            if content is not None:
+                path.parent.mkdir(parents=True, exist_ok=True)
+                path.write_text(content)
+
+            # Ensure file exists
+            if not path.exists():
+                return jsonify({"error": f"File not found: {file_path}"}), 404
+
+            # Run the procedure using tactus CLI
+            result = subprocess.run(
+                ["tactus", "run", str(path)],
+                capture_output=True,
+                text=True,
+                timeout=30,
+                cwd=WORKSPACE_ROOT,
+            )
+
+            return jsonify(
+                {
+                    "success": result.returncode == 0,
+                    "exitCode": result.returncode,
+                    "stdout": result.stdout,
+                    "stderr": result.stderr,
+                }
+            )
+        except subprocess.TimeoutExpired:
+            return jsonify({"error": "Procedure execution timed out (30s)"}), 408
+        except ValueError as e:
+            return jsonify({"error": str(e)}), 400
+        except Exception as e:
+            logger.error(f"Error running procedure {file_path}: {e}")
+            return jsonify({"error": str(e)}), 500
+
+    @app.route("/api/run/stream", methods=["GET", "POST"])
+    def run_procedure_stream():
+        """
+        Run a Tactus procedure with SSE streaming output.
+
+        For GET:
+        - path: workspace-relative path to procedure file (required, query param)
+        - inputs: JSON-encoded input parameters (optional, query param)
+
+        For POST:
+        - path: workspace-relative path to procedure file (required, JSON body)
+        - content: optional file content to save before running (JSON body)
+        - inputs: input parameters as object (optional, JSON body)
+        """
+        if request.method == "POST":
+            data = request.json or {}
+            file_path = data.get("path")
+            content = data.get("content")
+            inputs = data.get("inputs", {})
+        else:
+            file_path = request.args.get("path")
+            content = None
+            inputs_json = request.args.get("inputs", "{}")
+            # Parse inputs JSON for GET
+            try:
+                inputs = json.loads(inputs_json) if inputs_json else {}
+            except json.JSONDecodeError as e:
+                return jsonify({"error": f"Invalid 'inputs' JSON: {e}"}), 400
+
+        if not file_path:
+            return jsonify({"error": "Missing 'path' parameter"}), 400
+
+        try:
+            # Resolve path within workspace
+            path = _resolve_workspace_path(file_path)
+
+            # Save content if provided (POST requests can include file content)
+            if content is not None:
+                path.parent.mkdir(parents=True, exist_ok=True)
+                path.write_text(content)
+
+            # Ensure file exists
+            if not path.exists():
+                return jsonify({"error": f"File not found: {file_path}"}), 404
+
+            procedure_id = f"ide-{path.stem}"
+
+            def generate_events():
+                """Generator function that yields SSE events."""
+                log_handler = None
+                all_events = []  # Collect all events to save at the end
+                try:
+                    # Send start event
+                    import json
+                    from datetime import datetime
+                    from tactus.adapters.ide_log import IDELogHandler
+                    from tactus.core.runtime import TactusRuntime
+                    from tactus.adapters.file_storage import FileStorage
+                    from nanoid import generate
+
+                    # Generate unique run_id for this execution
+                    run_id = generate(size=21)
+
+                    start_event = {
+                        "event_type": "execution",
+                        "lifecycle_stage": "start",
+                        "procedure_id": procedure_id,
+                        "run_id": run_id,
+                        "timestamp": datetime.utcnow().isoformat() + "Z",
+                        "details": {"path": file_path},
+                        "inputs": inputs,  # Include inputs in start event
+                    }
+                    all_events.append(start_event)
+                    yield f"data: {json.dumps(start_event)}\n\n"
+
+                    # Create IDE log handler to collect structured events
+                    log_handler = IDELogHandler()
+
+                    # Create storage backend
+                    from pathlib import Path as PathLib
+
+                    storage_dir = (
+                        str(PathLib(WORKSPACE_ROOT) / ".tac" / "storage")
+                        if WORKSPACE_ROOT
+                        else "~/.tactus/storage"
+                    )
+                    storage_backend = FileStorage(storage_dir=storage_dir)
+
+                    # Create runtime with log handler and run_id
+                    runtime = TactusRuntime(
+                        procedure_id=procedure_id,
+                        storage_backend=storage_backend,
+                        hitl_handler=None,  # No HITL in IDE streaming mode
+                        log_handler=log_handler,
+                        run_id=run_id,
+                        source_file_path=str(path),
+                    )
+
+                    # Read procedure source
+                    source = path.read_text()
+
+                    # Check Docker availability for sandbox execution
+                    from tactus.sandbox import is_docker_available, SandboxConfig, ContainerRunner
+
+                    docker_available, docker_reason = is_docker_available()
+                    # Enable dev_mode by default in IDE for live code mounting
+                    sandbox_config = SandboxConfig(dev_mode=True)
+                    use_sandbox = docker_available and not sandbox_config.is_explicitly_disabled()
+
+                    if use_sandbox:
+                        logger.info("[SANDBOX] Docker available, using container execution")
+                    else:
+                        logger.info(
+                            f"[SANDBOX] Direct execution (Docker: {docker_available}, reason: {docker_reason})"
+                        )
+
+                    # Create event queue for sandbox event streaming (if using sandbox)
+                    sandbox_event_queue = None
+                    if use_sandbox:
+                        sandbox_event_queue = queue.Queue()
+
+                        # Emit container starting event
+                        container_starting_event = {
+                            "event_type": "container_status",
+                            "status": "starting",
+                            "execution_id": run_id,
+                            "timestamp": datetime.utcnow().isoformat() + "Z",
+                        }
+                        all_events.append(container_starting_event)
+                        yield f"data: {json.dumps(container_starting_event)}\n\n"
+
+                    # Run in a thread to avoid blocking
+                    import asyncio
+
+                    result_container = {
+                        "result": None,
+                        "error": None,
+                        "done": False,
+                        "container_ready": False,
+                    }
+
+                    # Capture inputs in closure scope for the thread
+                    procedure_inputs = inputs
+
+                    def run_procedure():
+                        try:
+                            # Create new event loop for this thread
+                            loop = asyncio.new_event_loop()
+                            asyncio.set_event_loop(loop)
+
+                            if use_sandbox:
+                                # Use sandbox execution (events streamed via broker over UDS)
+                                runner = ContainerRunner(sandbox_config)
+                                exec_result = loop.run_until_complete(
+                                    runner.run(
+                                        source=source,
+                                        params=procedure_inputs,
+                                        source_file_path=str(path),
+                                        format="lua",
+                                        event_handler=(
+                                            sandbox_event_queue.put if sandbox_event_queue else None
+                                        ),
+                                    )
+                                )
+
+                                # Mark container as ready after first response
+                                if not result_container["container_ready"]:
+                                    result_container["container_ready"] = True
+
+                                # Extract result from ExecutionResult
+                                if exec_result.status.value == "success":
+                                    result_container["result"] = exec_result.result
+                                else:
+                                    raise Exception(exec_result.error or "Sandbox execution failed")
+                            else:
+                                # Direct execution (no sandbox)
+                                result = loop.run_until_complete(
+                                    runtime.execute(source, context=procedure_inputs, format="lua")
+                                )
+                                result_container["result"] = result
+                        except Exception as e:
+                            result_container["error"] = e
+                        finally:
+                            result_container["done"] = True
+                            loop.close()
+
+                    exec_thread = threading.Thread(target=run_procedure)
+                    exec_thread.daemon = True
+                    exec_thread.start()
+
+                    # Emit container running event after starting
+                    if use_sandbox:
+                        container_running_event = {
+                            "event_type": "container_status",
+                            "status": "running",
+                            "execution_id": run_id,
+                            "timestamp": datetime.utcnow().isoformat() + "Z",
+                        }
+                        all_events.append(container_running_event)
+                        yield f"data: {json.dumps(container_running_event)}\n\n"
+
+                    # Stream events based on execution mode
+                    while not result_container["done"]:
+                        if use_sandbox and sandbox_event_queue:
+                            # Stream from sandbox callback queue
+                            try:
+                                event_dict = sandbox_event_queue.get(timeout=0.1)
+                                all_events.append(event_dict)
+                                yield f"data: {json.dumps(event_dict)}\n\n"
+                            except queue.Empty:
+                                pass
+                        else:
+                            # Stream from IDELogHandler (direct execution)
+                            events = log_handler.get_events(timeout=0.1)
+                            for event in events:
+                                try:
+                                    # Serialize with ISO format for datetime
+                                    event_dict = event.model_dump(mode="json")
+                                    # Format timestamp: add 'Z' only if no timezone info present
+                                    iso_string = event.timestamp.isoformat()
+                                    if not (
+                                        iso_string.endswith("Z")
+                                        or "+" in iso_string
+                                        or iso_string.count("-") > 2
+                                    ):
+                                        iso_string += "Z"
+                                    event_dict["timestamp"] = iso_string
+                                    all_events.append(event_dict)
+                                    yield f"data: {json.dumps(event_dict)}\n\n"
+                                except Exception as e:
+                                    logger.error(f"Error serializing event: {e}", exc_info=True)
+                                    logger.error(f"Event type: {type(event)}, Event: {event}")
+
+                        time.sleep(0.05)
+
+                    # Get any remaining events
+                    if use_sandbox and sandbox_event_queue:
+                        # Drain sandbox event queue
+                        while True:
+                            try:
+                                event_dict = sandbox_event_queue.get_nowait()
+                                all_events.append(event_dict)
+                                yield f"data: {json.dumps(event_dict)}\n\n"
+                            except queue.Empty:
+                                break
+
+                        # Emit container stopped event
+                        container_stopped_event = {
+                            "event_type": "container_status",
+                            "status": "stopped",
+                            "execution_id": run_id,
+                            "timestamp": datetime.utcnow().isoformat() + "Z",
+                        }
+                        all_events.append(container_stopped_event)
+                        yield f"data: {json.dumps(container_stopped_event)}\n\n"
+                    else:
+                        # Drain IDELogHandler events (direct execution)
+                        events = log_handler.get_events(timeout=0.1)
+                        for event in events:
+                            try:
+                                # Serialize with ISO format for datetime
+                                event_dict = event.model_dump(mode="json")
+                                # Format timestamp: add 'Z' only if no timezone info present
+                                iso_string = event.timestamp.isoformat()
+                                if not (
+                                    iso_string.endswith("Z")
+                                    or "+" in iso_string
+                                    or iso_string.count("-") > 2
+                                ):
+                                    iso_string += "Z"
+                                event_dict["timestamp"] = iso_string
+                                all_events.append(event_dict)
+                                yield f"data: {json.dumps(event_dict)}\n\n"
+                            except Exception as e:
+                                logger.error(f"Error serializing event: {e}", exc_info=True)
+                                logger.error(f"Event type: {type(event)}, Event: {event}")
+
+                    # Wait for thread to finish
+                    exec_thread.join(timeout=1)
+
+                    # Send completion event
+                    if result_container["error"]:
+                        complete_event = {
+                            "event_type": "execution",
+                            "lifecycle_stage": "error",
+                            "procedure_id": procedure_id,
+                            "exit_code": 1,
+                            "timestamp": datetime.utcnow().isoformat() + "Z",
+                            "details": {"success": False, "error": str(result_container["error"])},
+                        }
+                    else:
+                        complete_event = {
+                            "event_type": "execution",
+                            "lifecycle_stage": "complete",
+                            "procedure_id": procedure_id,
+                            "exit_code": 0,
+                            "timestamp": datetime.utcnow().isoformat() + "Z",
+                            "details": {"success": True},
+                        }
+                    all_events.append(complete_event)
+                    yield f"data: {json.dumps(complete_event)}\n\n"
+
+                    # Consolidate streaming chunks before saving to disk
+                    # Keep only the final accumulated text for each agent
+                    consolidated_events = []
+                    stream_chunks_by_agent = {}
+
+                    for event in all_events:
+                        if event.get("event_type") == "agent_stream_chunk":
+                            # Track by agent name, keeping only the latest
+                            agent_name = event.get("agent_name")
+                            stream_chunks_by_agent[agent_name] = event
+                        else:
+                            consolidated_events.append(event)
+
+                    # Add the final consolidated chunks
+                    consolidated_events.extend(stream_chunks_by_agent.values())
+
+                    # Save consolidated events to disk
+                    try:
+                        from pathlib import Path as PathLib
+
+                        events_dir = PathLib(storage_dir) / "events"
+                        events_dir.mkdir(parents=True, exist_ok=True)
+                        events_file = events_dir / f"{run_id}.json"
+                        with open(events_file, "w") as f:
+                            json.dump(consolidated_events, f, indent=2)
+                    except Exception as e:
+                        logger.error(f"Failed to save events for run {run_id}: {e}", exc_info=True)
+
+                except Exception as e:
+                    logger.error(f"Error in streaming execution: {e}", exc_info=True)
+                    error_event = {
+                        "event_type": "execution",
+                        "lifecycle_stage": "error",
+                        "procedure_id": procedure_id,
+                        "timestamp": datetime.utcnow().isoformat() + "Z",
+                        "details": {"error": str(e)},
+                    }
+                    yield f"data: {json.dumps(error_event)}\n\n"
+
+            return Response(
+                stream_with_context(generate_events()),
+                mimetype="text/event-stream",
+                headers={
+                    "Cache-Control": "no-cache",
+                    "X-Accel-Buffering": "no",
+                    "Connection": "keep-alive",
+                },
+            )
+
+        except ValueError as e:
+            return jsonify({"error": str(e)}), 400
+        except Exception as e:
+            logger.error(f"Error setting up streaming execution: {e}", exc_info=True)
+            return jsonify({"error": str(e)}), 500
+
+    @app.route("/api/test/stream", methods=["GET"])
+    def test_procedure_stream():
+        """
+        Run BDD tests with SSE streaming output.
+
+        Query params:
+        - path: procedure file path (required)
+        - mock: use mock mode (optional, default true)
+        - scenario: specific scenario name (optional)
+        - parallel: run in parallel (optional, default false)
+        """
+        file_path = request.args.get("path")
+
+        if not file_path:
+            return jsonify({"error": "Missing 'path' parameter"}), 400
+
+        # Get options
+        mock = request.args.get("mock", "true").lower() == "true"
+        parallel = request.args.get("parallel", "false").lower() == "true"
+
+        try:
+            # Resolve path within workspace
+            path = _resolve_workspace_path(file_path)
+
+            # Ensure file exists
+            if not path.exists():
+                return jsonify({"error": f"File not found: {file_path}"}), 404
+
+            procedure_id = path.stem
+
+            def generate_events():
+                """Generator function that yields SSE test events."""
+                try:
+                    import json
+                    from datetime import datetime
+                    from tactus.validation import TactusValidator
+                    from tactus.testing import TactusTestRunner, GherkinParser
+
+                    # Validate and extract specifications
+                    validator = TactusValidator()
+                    validation_result = validator.validate_file(str(path))
+
+                    if not validation_result.valid:
+                        # Emit validation error
+                        error_event = {
+                            "event_type": "execution",
+                            "lifecycle_stage": "error",
+                            "procedure_id": procedure_id,
+                            "timestamp": datetime.utcnow().isoformat() + "Z",
+                            "details": {
+                                "error": "Validation failed",
+                                "errors": [
+                                    {"message": e.message, "level": e.level}
+                                    for e in validation_result.errors
+                                ],
+                            },
+                        }
+                        yield f"data: {json.dumps(error_event)}\n\n"
+                        return
+
+                    if (
+                        not validation_result.registry
+                        or not validation_result.registry.gherkin_specifications
+                    ):
+                        # No specifications found
+                        error_event = {
+                            "event_type": "execution",
+                            "lifecycle_stage": "error",
+                            "procedure_id": procedure_id,
+                            "timestamp": datetime.utcnow().isoformat() + "Z",
+                            "details": {"error": "No specifications found in procedure"},
+                        }
+                        yield f"data: {json.dumps(error_event)}\n\n"
+                        return
+
+                    # Clear Behave's global step registry before each test run
+                    # This prevents conflicts when running multiple tests in the same Flask process
+                    try:
+                        from behave import step_registry
+
+                        # Clear all registered steps (each step_type maps to a list)
+                        step_registry.registry.steps = {
+                            "given": [],
+                            "when": [],
+                            "then": [],
+                            "step": [],
+                        }
+                        # Recreate the decorators
+                        from behave.step_registry import setup_step_decorators
+
+                        setup_step_decorators()
+                    except Exception as e:
+                        logger.warning(f"Could not reset Behave step registry: {e}")
+
+                    # Setup test runner with mocks from registry
+                    mock_tools = None
+                    if mock:
+                        # Start with default done mock
+                        mock_tools = {"done": {"status": "ok"}}
+                        # Add tool mocks from Mocks {} block in .tac file
+                        if validation_result.registry.mocks:
+                            for tool_name, mock_config in validation_result.registry.mocks.items():
+                                # Extract output/response from mock config
+                                if isinstance(mock_config, dict) and "output" in mock_config:
+                                    mock_tools[tool_name] = mock_config["output"]
+                                else:
+                                    mock_tools[tool_name] = mock_config
+                    runner = TactusTestRunner(path, mock_tools=mock_tools, mocked=mock)
+                    runner.setup(validation_result.registry.gherkin_specifications)
+
+                    # Get parsed feature to count scenarios
+                    parser = GherkinParser()
+                    parsed_feature = parser.parse(validation_result.registry.gherkin_specifications)
+                    total_scenarios = len(parsed_feature.scenarios)
+
+                    # Emit started event
+                    start_event = {
+                        "event_type": "test_started",
+                        "procedure_file": str(path),
+                        "total_scenarios": total_scenarios,
+                        "timestamp": datetime.utcnow().isoformat() + "Z",
+                    }
+                    yield f"data: {json.dumps(start_event)}\n\n"
+
+                    # Run tests
+                    test_result = runner.run_tests(parallel=parallel)
+
+                    # Emit scenario completion events
+                    for feature in test_result.features:
+                        for scenario in feature.scenarios:
+                            scenario_event = {
+                                "event_type": "test_scenario_completed",
+                                "scenario_name": scenario.name,
+                                "status": scenario.status,
+                                "duration": scenario.duration,
+                                "total_cost": scenario.total_cost,
+                                "total_tokens": scenario.total_tokens,
+                                "llm_calls": scenario.llm_calls,
+                                "iterations": scenario.iterations,
+                                "tools_used": scenario.tools_used,
+                                "timestamp": datetime.utcnow().isoformat() + "Z",
+                            }
+                            yield f"data: {json.dumps(scenario_event)}\n\n"
+
+                    # Emit completed event
+                    complete_event = {
+                        "event_type": "test_completed",
+                        "result": {
+                            "total_scenarios": test_result.total_scenarios,
+                            "passed_scenarios": test_result.passed_scenarios,
+                            "failed_scenarios": test_result.failed_scenarios,
+                            "total_cost": test_result.total_cost,
+                            "total_tokens": test_result.total_tokens,
+                            "total_llm_calls": test_result.total_llm_calls,
+                            "total_iterations": test_result.total_iterations,
+                            "unique_tools_used": test_result.unique_tools_used,
+                            "features": [
+                                {
+                                    "name": f.name,
+                                    "scenarios": [
+                                        {
+                                            "name": s.name,
+                                            "status": s.status,
+                                            "duration": s.duration,
+                                            "steps": [
+                                                {
+                                                    "keyword": step.keyword,
+                                                    "text": step.message,
+                                                    "status": step.status,
+                                                    "error_message": step.error_message,
+                                                }
+                                                for step in s.steps
+                                            ],
+                                        }
+                                        for s in f.scenarios
+                                    ],
+                                }
+                                for f in test_result.features
+                            ],
+                        },
+                        "timestamp": datetime.utcnow().isoformat() + "Z",
+                    }
+                    yield f"data: {json.dumps(complete_event)}\n\n"
+
+                    # Cleanup
+                    runner.cleanup()
+
+                except Exception as e:
+                    logger.error(f"Error in test execution: {e}", exc_info=True)
+                    error_event = {
+                        "event_type": "execution",
+                        "lifecycle_stage": "error",
+                        "procedure_id": procedure_id,
+                        "timestamp": datetime.utcnow().isoformat() + "Z",
+                        "details": {"error": str(e)},
+                    }
+                    yield f"data: {json.dumps(error_event)}\n\n"
+
+            return Response(
+                stream_with_context(generate_events()),
+                mimetype="text/event-stream",
+                headers={
+                    "Cache-Control": "no-cache",
+                    "X-Accel-Buffering": "no",
+                    "Connection": "keep-alive",
+                },
+            )
+
+        except ValueError as e:
+            return jsonify({"error": str(e)}), 400
+        except Exception as e:
+            logger.error(f"Error setting up test execution: {e}", exc_info=True)
+            return jsonify({"error": str(e)}), 500
+
+    @app.route("/api/evaluate/stream", methods=["GET"])
+    def evaluate_procedure_stream():
+        """
+        Run BDD evaluation with SSE streaming output.
+
+        Query params:
+        - path: procedure file path (required)
+        - runs: number of runs per scenario (optional, default 10)
+        - mock: use mock mode (optional, default true)
+        - scenario: specific scenario name (optional)
+        - parallel: run in parallel (optional, default true)
+        """
+        file_path = request.args.get("path")
+
+        if not file_path:
+            return jsonify({"error": "Missing 'path' parameter"}), 400
+
+        # Get options
+        runs = int(request.args.get("runs", "10"))
+        mock = request.args.get("mock", "true").lower() == "true"
+        parallel = request.args.get("parallel", "true").lower() == "true"
+
+        try:
+            # Resolve path within workspace
+            path = _resolve_workspace_path(file_path)
+
+            # Ensure file exists
+            if not path.exists():
+                return jsonify({"error": f"File not found: {file_path}"}), 404
+
+            procedure_id = path.stem
+
+            def generate_events():
+                """Generator function that yields SSE evaluation events."""
+                try:
+                    import json
+                    from datetime import datetime
+                    from tactus.validation import TactusValidator
+                    from tactus.testing import TactusEvaluationRunner, GherkinParser
+
+                    # Validate and extract specifications
+                    validator = TactusValidator()
+                    validation_result = validator.validate_file(str(path))
+
+                    if not validation_result.valid:
+                        error_event = {
+                            "event_type": "execution",
+                            "lifecycle_stage": "error",
+                            "procedure_id": procedure_id,
+                            "timestamp": datetime.utcnow().isoformat() + "Z",
+                            "details": {
+                                "error": "Validation failed",
+                                "errors": [
+                                    {"message": e.message, "level": e.level}
+                                    for e in validation_result.errors
+                                ],
+                            },
+                        }
+                        yield f"data: {json.dumps(error_event)}\n\n"
+                        return
+
+                    if (
+                        not validation_result.registry
+                        or not validation_result.registry.gherkin_specifications
+                    ):
+                        error_event = {
+                            "event_type": "execution",
+                            "lifecycle_stage": "error",
+                            "procedure_id": procedure_id,
+                            "timestamp": datetime.utcnow().isoformat() + "Z",
+                            "details": {"error": "No specifications found in procedure"},
+                        }
+                        yield f"data: {json.dumps(error_event)}\n\n"
+                        return
+
+                    # Setup evaluation runner
+                    mock_tools = {"done": {"status": "ok"}} if mock else None
+                    evaluator = TactusEvaluationRunner(path, mock_tools=mock_tools)
+                    evaluator.setup(validation_result.registry.gherkin_specifications)
+
+                    # Get parsed feature to count scenarios
+                    parser = GherkinParser()
+                    parsed_feature = parser.parse(validation_result.registry.gherkin_specifications)
+                    total_scenarios = len(parsed_feature.scenarios)
+
+                    # Emit started event
+                    start_event = {
+                        "event_type": "evaluation_started",
+                        "procedure_file": str(path),
+                        "total_scenarios": total_scenarios,
+                        "runs_per_scenario": runs,
+                        "timestamp": datetime.utcnow().isoformat() + "Z",
+                    }
+                    yield f"data: {json.dumps(start_event)}\n\n"
+
+                    # Run evaluation
+                    eval_results = evaluator.evaluate_all(runs=runs, parallel=parallel)
+
+                    # Emit progress/completion events for each scenario
+                    for eval_result in eval_results:
+                        progress_event = {
+                            "event_type": "evaluation_progress",
+                            "scenario_name": eval_result.scenario_name,
+                            "completed_runs": eval_result.total_runs,
+                            "total_runs": eval_result.total_runs,
+                            "timestamp": datetime.utcnow().isoformat() + "Z",
+                        }
+                        yield f"data: {json.dumps(progress_event)}\n\n"
+
+                    # Emit completed event
+                    complete_event = {
+                        "event_type": "evaluation_completed",
+                        "results": [
+                            {
+                                "scenario_name": r.scenario_name,
+                                "total_runs": r.total_runs,
+                                "successful_runs": r.successful_runs,
+                                "failed_runs": r.failed_runs,
+                                "success_rate": r.success_rate,
+                                "consistency_score": r.consistency_score,
+                                "is_flaky": r.is_flaky,
+                                "avg_duration": r.avg_duration,
+                                "std_duration": r.std_duration,
+                            }
+                            for r in eval_results
+                        ],
+                        "timestamp": datetime.utcnow().isoformat() + "Z",
+                    }
+                    yield f"data: {json.dumps(complete_event)}\n\n"
+
+                    # Cleanup
+                    evaluator.cleanup()
+
+                except Exception as e:
+                    logger.error(f"Error in evaluation execution: {e}", exc_info=True)
+                    error_event = {
+                        "event_type": "execution",
+                        "lifecycle_stage": "error",
+                        "procedure_id": procedure_id,
+                        "timestamp": datetime.utcnow().isoformat() + "Z",
+                        "details": {"error": str(e)},
+                    }
+                    yield f"data: {json.dumps(error_event)}\n\n"
+
+            return Response(
+                stream_with_context(generate_events()),
+                mimetype="text/event-stream",
+                headers={
+                    "Cache-Control": "no-cache",
+                    "X-Accel-Buffering": "no",
+                    "Connection": "keep-alive",
+                },
+            )
+
+        except ValueError as e:
+            return jsonify({"error": str(e)}), 400
+        except Exception as e:
+            logger.error(f"Error setting up evaluation execution: {e}", exc_info=True)
+            return jsonify({"error": str(e)}), 500
+
+    @app.route("/api/pydantic-eval/stream", methods=["GET"])
+    def pydantic_eval_stream():
+        """
+        Run Pydantic Evals with SSE streaming output.
+
+        Query params:
+        - path: procedure file path (required)
+        - runs: number of runs per case (optional, default 1)
+        """
+        logger.info(f"Pydantic eval stream request: args={request.args}")
+
+        file_path = request.args.get("path")
+        if not file_path:
+            logger.error("Missing 'path' parameter")
+            return jsonify({"error": "Missing 'path' parameter"}), 400
+
+        runs = int(request.args.get("runs", "1"))
+
+        try:
+            # Resolve path within workspace
+            logger.info(f"Resolving path: {file_path}")
+            path = _resolve_workspace_path(file_path)
+            logger.info(f"Resolved to: {path}")
+
+            if not path.exists():
+                return jsonify({"error": f"File not found: {file_path}"}), 404
+
+            procedure_id = path.stem
+
+            def generate_events():
+                """Generator function that yields SSE evaluation events."""
+                try:
+                    from tactus.testing.pydantic_eval_runner import TactusPydanticEvalRunner
+                    from tactus.testing.eval_models import (
+                        EvaluationConfig,
+                        EvalCase,
+                        EvaluatorConfig,
+                    )
+
+                    # Validate and extract evaluations
+                    validator = TactusValidator()
+                    validation_result = validator.validate_file(str(path))
+
+                    if not validation_result.valid:
+                        error_event = {
+                            "event_type": "execution",
+                            "lifecycle_stage": "error",
+                            "procedure_id": procedure_id,
+                            "timestamp": datetime.utcnow().isoformat() + "Z",
+                            "details": {
+                                "error": "Validation failed",
+                                "errors": [e.message for e in validation_result.errors],
+                            },
+                        }
+                        yield f"data: {json.dumps(error_event)}\n\n"
+                        return
+
+                    if (
+                        not validation_result.registry
+                        or not validation_result.registry.pydantic_evaluations
+                    ):
+                        error_event = {
+                            "event_type": "execution",
+                            "lifecycle_stage": "error",
+                            "procedure_id": procedure_id,
+                            "timestamp": datetime.utcnow().isoformat() + "Z",
+                            "details": {"error": "No evaluations found in procedure"},
+                        }
+                        yield f"data: {json.dumps(error_event)}\n\n"
+                        return
+
+                    # Parse evaluation config FIRST (before start event)
+                    eval_dict = validation_result.registry.pydantic_evaluations
+                    dataset_cases = [EvalCase(**c) for c in eval_dict.get("dataset", [])]
+                    evaluators = [EvaluatorConfig(**e) for e in eval_dict.get("evaluators", [])]
+
+                    # Parse thresholds if present
+                    from tactus.testing.eval_models import EvaluationThresholds
+
+                    thresholds = None
+                    if "thresholds" in eval_dict:
+                        thresholds = EvaluationThresholds(**eval_dict["thresholds"])
+
+                    # Use runs from file if specified, otherwise use query param
+                    actual_runs = eval_dict.get("runs", runs)
+
+                    # Emit start event (after actual_runs is defined)
+                    start_event = {
+                        "event_type": "execution",
+                        "lifecycle_stage": "started",
+                        "procedure_id": procedure_id,
+                        "timestamp": datetime.utcnow().isoformat() + "Z",
+                        "details": {"type": "pydantic_eval", "runs": actual_runs},
+                    }
+                    yield f"data: {json.dumps(start_event)}\n\n"
+
+                    eval_config = EvaluationConfig(
+                        dataset=dataset_cases,
+                        evaluators=evaluators,
+                        runs=actual_runs,
+                        parallel=False,  # Sequential for IDE streaming
+                        thresholds=thresholds,
+                    )
+
+                    # Run evaluation
+                    runner = TactusPydanticEvalRunner(
+                        procedure_file=path,
+                        eval_config=eval_config,
+                        openai_api_key=os.environ.get("OPENAI_API_KEY"),
+                    )
+
+                    report = runner.run_evaluation()
+
+                    # Emit results
+                    result_details = {
+                        "type": "pydantic_eval",
+                        "total_cases": len(report.cases) if hasattr(report, "cases") else 0,
+                    }
+
+                    if hasattr(report, "cases"):
+                        result_details["cases"] = []
+                        for case in report.cases:
+                            # Convert case to dict, handling non-serializable objects
+                            def make_serializable(obj):
+                                """Recursively convert objects to JSON-serializable types."""
+                                if isinstance(obj, (str, int, float, bool, type(None))):
+                                    return obj
+                                elif isinstance(obj, dict):
+                                    return {k: make_serializable(v) for k, v in obj.items()}
+                                elif isinstance(obj, (list, tuple)):
+                                    return [make_serializable(item) for item in obj]
+                                elif hasattr(obj, "__dict__"):
+                                    # Convert object with __dict__ to dict
+                                    return {
+                                        k: make_serializable(v)
+                                        for k, v in obj.__dict__.items()
+                                        if not k.startswith("_")
+                                    }
+                                else:
+                                    return str(obj)
+
+                            case_dict = {
+                                "name": str(case.name),
+                                "inputs": make_serializable(case.inputs),
+                                "output": make_serializable(case.output),
+                                "assertions": make_serializable(case.assertions),
+                                "scores": make_serializable(case.scores),
+                                "labels": make_serializable(case.labels),
+                                "duration": (
+                                    float(case.task_duration)
+                                    if hasattr(case, "task_duration")
+                                    else 0.0
+                                ),
+                            }
+                            result_details["cases"].append(case_dict)
+
+                    # Check thresholds
+                    passed, violations = runner.check_thresholds(report)
+                    result_details["thresholds_passed"] = passed
+                    if violations:
+                        result_details["threshold_violations"] = violations
+
+                    result_event = {
+                        "event_type": "execution",
+                        "lifecycle_stage": "complete",
+                        "procedure_id": procedure_id,
+                        "timestamp": datetime.utcnow().isoformat() + "Z",
+                        "details": result_details,
+                    }
+                    yield f"data: {json.dumps(result_event)}\n\n"
+
+                except ImportError as e:
+                    error_event = {
+                        "event_type": "execution",
+                        "lifecycle_stage": "error",
+                        "procedure_id": procedure_id,
+                        "timestamp": datetime.utcnow().isoformat() + "Z",
+                        "details": {"error": f"pydantic_evals not installed: {e}"},
+                    }
+                    yield f"data: {json.dumps(error_event)}\n\n"
+                except Exception as e:
+                    logger.error(f"Error running Pydantic Evals: {e}", exc_info=True)
+                    error_event = {
+                        "event_type": "execution",
+                        "lifecycle_stage": "error",
+                        "procedure_id": procedure_id,
+                        "timestamp": datetime.utcnow().isoformat() + "Z",
+                        "details": {"error": str(e)},
+                    }
+                    yield f"data: {json.dumps(error_event)}\n\n"
+
+            return Response(
+                stream_with_context(generate_events()),
+                mimetype="text/event-stream",
+                headers={
+                    "Cache-Control": "no-cache",
+                    "X-Accel-Buffering": "no",
+                    "Connection": "keep-alive",
+                },
+            )
+
+        except Exception as e:
+            logger.error(f"Error setting up Pydantic Evals: {e}", exc_info=True)
+            return jsonify({"error": str(e)}), 500
+
+    @app.route("/api/traces/runs", methods=["GET"])
+    def list_trace_runs():
+        """List all execution runs by grouping checkpoints by run_id."""
+        try:
+            from pathlib import Path as PathLib
+            from tactus.adapters.file_storage import FileStorage
+            from collections import defaultdict
+
+            # Get optional query params
+            procedure = request.args.get("procedure")
+            limit = int(request.args.get("limit", "50"))
+
+            # Create storage backend
+            storage_dir = (
+                str(PathLib(WORKSPACE_ROOT) / ".tac" / "storage")
+                if WORKSPACE_ROOT
+                else "~/.tactus/storage"
+            )
+            storage_backend = FileStorage(storage_dir=storage_dir)
+
+            # Load procedure metadata
+            if not procedure:
+                return jsonify({"runs": []})
+
+            metadata = storage_backend.load_procedure_metadata(procedure)
+
+            # Group checkpoints by run_id
+            runs_dict = defaultdict(list)
+            for checkpoint in metadata.execution_log:
+                runs_dict[checkpoint.run_id].append(checkpoint)
+
+            # Build runs list
+            runs_data = []
+            for run_id, checkpoints in runs_dict.items():
+                # Sort checkpoints by position
+                checkpoints.sort(key=lambda c: c.position)
+
+                # Get start/end times
+                start_time = checkpoints[0].timestamp if checkpoints else None
+                end_time = checkpoints[-1].timestamp if checkpoints else None
+
+                runs_data.append(
+                    {
+                        "run_id": run_id,
+                        "procedure_name": procedure,
+                        "start_time": start_time.isoformat() if start_time else None,
+                        "end_time": end_time.isoformat() if end_time else None,
+                        "status": "COMPLETED",  # Can be enhanced later
+                        "checkpoint_count": len(checkpoints),
+                    }
+                )
+
+            # Sort by start_time (most recent first)
+            runs_data.sort(key=lambda r: r["start_time"] or "", reverse=True)
+
+            # Apply limit
+            runs_data = runs_data[:limit]
+
+            return jsonify({"runs": runs_data})
+        except Exception as e:
+            logger.error(f"Error listing trace runs: {e}", exc_info=True)
+            return jsonify({"error": str(e)}), 500
+
+    @app.route("/api/traces/runs/<run_id>", methods=["GET"])
+    def get_trace_run(run_id: str):
+        """Get a specific execution run by filtering checkpoints by run_id."""
+        try:
+            from pathlib import Path as PathLib
+            from tactus.adapters.file_storage import FileStorage
+
+            # Get procedure name from query param or try to find it
+            procedure = request.args.get("procedure")
+            if not procedure:
+                return jsonify({"error": "procedure parameter required"}), 400
+
+            # Create storage backend
+            storage_dir = (
+                str(PathLib(WORKSPACE_ROOT) / ".tac" / "storage")
+                if WORKSPACE_ROOT
+                else "~/.tactus/storage"
+            )
+            storage_backend = FileStorage(storage_dir=storage_dir)
+
+            # Load procedure metadata
+            metadata = storage_backend.load_procedure_metadata(procedure)
+
+            # Filter checkpoints by run_id
+            run_checkpoints = [cp for cp in metadata.execution_log if cp.run_id == run_id]
+
+            if not run_checkpoints:
+                return jsonify({"error": f"Run not found: {run_id}"}), 404
+
+            # Sort by position
+            run_checkpoints.sort(key=lambda c: c.position)
+
+            # Get start/end times
+            start_time = run_checkpoints[0].timestamp if run_checkpoints else None
+            end_time = run_checkpoints[-1].timestamp if run_checkpoints else None
+
+            # Convert to API format
+            run_dict = {
+                "run_id": run_id,
+                "procedure_name": procedure,
+                "file_path": "",
+                "start_time": start_time.isoformat() if start_time else None,
+                "end_time": end_time.isoformat() if end_time else None,
+                "status": "COMPLETED",
+                "execution_log": [
+                    {
+                        "position": cp.position,
+                        "type": cp.type,
+                        "result": cp.result,
+                        "timestamp": cp.timestamp.isoformat() if cp.timestamp else None,
+                        "duration_ms": cp.duration_ms,
+                        "source_location": (
+                            cp.source_location.model_dump() if cp.source_location else None
+                        ),
+                        "captured_vars": cp.captured_vars,
+                    }
+                    for cp in run_checkpoints
+                ],
+                "final_state": metadata.state,
+                "breakpoints": [],
+            }
+
+            return jsonify(run_dict)
+        except Exception as e:
+            logger.error(f"Error getting trace run {run_id}: {e}", exc_info=True)
+            return jsonify({"error": str(e)}), 500
+
+    @app.route("/api/traces/runs/<run_id>/checkpoints", methods=["GET"])
+    def get_run_checkpoints(run_id: str):
+        """Get all checkpoints for a specific run."""
+        try:
+            from pathlib import Path as PathLib
+            from tactus.adapters.file_storage import FileStorage
+
+            # Get procedure name from query param
+            procedure = request.args.get("procedure")
+            if not procedure:
+                return jsonify({"error": "procedure parameter required"}), 400
+
+            # Create storage backend
+            storage_dir = (
+                str(PathLib(WORKSPACE_ROOT) / ".tac" / "storage")
+                if WORKSPACE_ROOT
+                else "~/.tactus/storage"
+            )
+            storage_backend = FileStorage(storage_dir=storage_dir)
+
+            # Load procedure metadata
+            metadata = storage_backend.load_procedure_metadata(procedure)
+
+            # Filter checkpoints by run_id
+            run_checkpoints = [cp for cp in metadata.execution_log if cp.run_id == run_id]
+
+            # Sort by position
+            run_checkpoints.sort(key=lambda c: c.position)
+
+            # Convert to dict format
+            checkpoints_dict = [
+                {
+                    "run_id": cp.run_id,
+                    "position": cp.position,
+                    "name": cp.type,  # Use 'type' field as the name (e.g., "agent_turn")
+                    "timestamp": cp.timestamp.isoformat() if cp.timestamp else None,
+                    "source_location": (
+                        {
+                            "file": cp.source_location.file,
+                            "line": cp.source_location.line,
+                        }
+                        if cp.source_location
+                        else None
+                    ),
+                    "data": getattr(cp, "data", None),  # Not all checkpoints have 'data'
+                }
+                for cp in run_checkpoints
+            ]
+
+            return jsonify({"checkpoints": checkpoints_dict})
+        except Exception as e:
+            logger.error(f"Error getting checkpoints for run {run_id}: {e}", exc_info=True)
+            return jsonify({"error": str(e)}), 500
+
+    @app.route("/api/traces/runs/<run_id>/checkpoints/<int:position>", methods=["GET"])
+    def get_checkpoint(run_id: str, position: int):
+        """Get a specific checkpoint from a run by filtering by run_id."""
+        try:
+            from pathlib import Path as PathLib
+            from tactus.adapters.file_storage import FileStorage
+
+            # Get procedure name from query param
+            procedure = request.args.get("procedure")
+            if not procedure:
+                return jsonify({"error": "procedure parameter required"}), 400
+
+            # Create storage backend
+            storage_dir = (
+                str(PathLib(WORKSPACE_ROOT) / ".tac" / "storage")
+                if WORKSPACE_ROOT
+                else "~/.tactus/storage"
+            )
+            storage_backend = FileStorage(storage_dir=storage_dir)
+
+            # Load procedure metadata
+            metadata = storage_backend.load_procedure_metadata(procedure)
+
+            # Find checkpoint by run_id and position
+            checkpoint = next(
+                (
+                    cp
+                    for cp in metadata.execution_log
+                    if cp.run_id == run_id and cp.position == position
+                ),
+                None,
+            )
+
+            if not checkpoint:
+                return (
+                    jsonify({"error": f"Checkpoint position {position} not found in run {run_id}"}),
+                    404,
+                )
+
+            # Convert to API format
+            cp_dict = {
+                "position": checkpoint.position,
+                "type": checkpoint.type,
+                "result": checkpoint.result,
+                "timestamp": checkpoint.timestamp.isoformat() if checkpoint.timestamp else None,
+                "duration_ms": checkpoint.duration_ms,
+                "source_location": (
+                    checkpoint.source_location.model_dump() if checkpoint.source_location else None
+                ),
+                "captured_vars": checkpoint.captured_vars,
+            }
+
+            return jsonify(cp_dict)
+        except Exception as e:
+            logger.error(f"Error getting checkpoint {run_id}@{position}: {e}", exc_info=True)
+            return jsonify({"error": str(e)}), 500
+
+    @app.route("/api/traces/runs/<run_id>/statistics", methods=["GET"])
+    def get_run_statistics(run_id: str):
+        """Get statistics for a run by filtering checkpoints by run_id."""
+        try:
+            from pathlib import Path as PathLib
+            from tactus.adapters.file_storage import FileStorage
+            from collections import Counter
+
+            # Get procedure name from query param
+            procedure = request.args.get("procedure")
+            if not procedure:
+                return jsonify({"error": "procedure parameter required"}), 400
+
+            # Create storage backend
+            storage_dir = (
+                str(PathLib(WORKSPACE_ROOT) / ".tac" / "storage")
+                if WORKSPACE_ROOT
+                else "~/.tactus/storage"
+            )
+            storage_backend = FileStorage(storage_dir=storage_dir)
+
+            # Load procedure metadata
+            metadata = storage_backend.load_procedure_metadata(procedure)
+
+            # Filter checkpoints by run_id
+            run_checkpoints = [cp for cp in metadata.execution_log if cp.run_id == run_id]
+
+            if not run_checkpoints:
+                return jsonify({"error": f"Run not found: {run_id}"}), 404
+
+            # Calculate statistics
+            checkpoint_types = Counter(cp.type for cp in run_checkpoints)
+            total_duration = sum(cp.duration_ms or 0 for cp in run_checkpoints)
+            has_source_locations = sum(1 for cp in run_checkpoints if cp.source_location)
+
+            stats = {
+                "run_id": run_id,
+                "procedure": procedure,
+                "status": "COMPLETED",
+                "total_checkpoints": len(run_checkpoints),
+                "checkpoints_by_type": dict(checkpoint_types),
+                "total_duration_ms": total_duration,
+                "has_source_locations": has_source_locations,
+            }
+
+            return jsonify(stats)
+        except Exception as e:
+            logger.error(f"Error getting statistics for {run_id}: {e}", exc_info=True)
+            return jsonify({"error": str(e)}), 500
+
+    @app.route("/api/traces/runs/<run_id>/events", methods=["GET"])
+    def get_run_events(run_id: str):
+        """Get all SSE events for a specific run."""
+        try:
+            from pathlib import Path as PathLib
+
+            # Determine storage directory
+            storage_dir = (
+                str(PathLib(WORKSPACE_ROOT) / ".tac" / "storage")
+                if WORKSPACE_ROOT
+                else "~/.tactus/storage"
+            )
+            events_dir = PathLib(storage_dir) / "events"
+            events_file = events_dir / f"{run_id}.json"
+
+            if not events_file.exists():
+                return jsonify({"error": f"Events not found for run {run_id}"}), 404
+
+            # Load events from file
+            with open(events_file, "r") as f:
+                events = json.load(f)
+
+            return jsonify({"events": events})
+        except Exception as e:
+            logger.error(f"Error getting events for {run_id}: {e}", exc_info=True)
+            return jsonify({"error": str(e)}), 500
+
+    # Coding Assistant - persistent agent instance per session
+    coding_assistant = None
+
+    def get_or_create_assistant():
+        """Get or create the coding assistant instance."""
+        nonlocal coding_assistant
+        if coding_assistant is None and WORKSPACE_ROOT:
+            try:
+                from tactus.ide.coding_assistant import CodingAssistantAgent
+                from tactus.core.config_manager import ConfigManager
+
+                # Load configuration
+                config_manager = ConfigManager()
+                # For IDE, we don't have a procedure file, so use a dummy path
+                config = config_manager._load_from_environment()
+
+                # Try to load user config
+                for user_path in config_manager._get_user_config_paths():
+                    if user_path.exists():
+                        user_config = config_manager._load_yaml_file(user_path)
+                        if user_config:
+                            config = config_manager._deep_merge(config, user_config)
+                            break
+
+                coding_assistant = CodingAssistantAgent(WORKSPACE_ROOT, config)
+                logger.info("Coding assistant initialized")
+            except Exception as e:
+                logger.error(f"Failed to initialize coding assistant: {e}", exc_info=True)
+                raise
+        return coding_assistant
+
+    @app.route("/api/chat", methods=["POST"])
+    def chat_message():
+        """Handle chat messages from the user."""
+        try:
+            data = request.json
+            message = data.get("message")
+
+            if not message:
+                return jsonify({"error": "Missing 'message' parameter"}), 400
+
+            if not WORKSPACE_ROOT:
+                return jsonify({"error": "No workspace folder selected"}), 400
+
+            # Get or create assistant
+            assistant = get_or_create_assistant()
+
+            # Process message
+            result = assistant.process_message(message)
+
+            return jsonify(
+                {
+                    "success": True,
+                    "response": result["response"],
+                    "tool_calls": result.get("tool_calls", []),
+                }
+            )
+
+        except Exception as e:
+            logger.error(f"Error handling chat message: {e}", exc_info=True)
+            return jsonify({"error": str(e)}), 500
+
+    @app.route("/api/chat/stream", methods=["POST"])
+    def chat_stream():
+        """
+        Stream chat responses with SSE using our working implementation.
+
+        Request body:
+        - workspace_root: Workspace path
+        - message: User's message
+        - config: Optional config with provider, model, etc.
+        """
+        try:
+            import sys
+            import os
+            import uuid
+            import asyncio
+
+            # Add backend directory to path so we can import our modules
+            backend_dir = os.path.join(
+                os.path.dirname(__file__), "..", "..", "tactus-ide", "backend"
+            )
+            if backend_dir not in sys.path:
+                sys.path.insert(0, backend_dir)
+
+            from assistant_service import AssistantService
+
+            data = request.json or {}
+            workspace_root = data.get("workspace_root") or WORKSPACE_ROOT
+            user_message = data.get("message")
+            config = data.get(
+                "config",
+                {"provider": "openai", "model": "gpt-4o", "temperature": 0.7, "max_tokens": 4000},
+            )
+
+            if not workspace_root or not user_message:
+                return jsonify({"error": "workspace_root and message required"}), 400
+
+            # Create service instance
+            conversation_id = str(uuid.uuid4())
+            service = AssistantService(workspace_root, config)
+
+            def generate():
+                """Generator function that yields SSE events."""
+                loop = asyncio.new_event_loop()
+                asyncio.set_event_loop(loop)
+
+                try:
+                    # Start conversation (configures DSPy LM internally)
+                    loop.run_until_complete(service.start_conversation(conversation_id))
+
+                    # Send immediate thinking indicator
+                    yield f"data: {json.dumps({'type': 'thinking', 'content': 'Processing your request...'})}\n\n"
+
+                    # Create async generator
+                    async_gen = service.send_message(user_message)
+
+                    # Consume events one at a time and yield immediately
+                    while True:
+                        try:
+                            event = loop.run_until_complete(async_gen.__anext__())
+                            yield f"data: {json.dumps(event)}\n\n"
+                        except StopAsyncIteration:
+                            break
+
+                except Exception as e:
+                    logger.error(f"Error streaming message: {e}", exc_info=True)
+                    yield f"data: {json.dumps({'type': 'error', 'error': str(e)})}\n\n"
+                finally:
+                    loop.close()
+
+            return Response(
+                stream_with_context(generate()),
+                mimetype="text/event-stream",
+                headers={
+                    "Cache-Control": "no-cache",
+                    "X-Accel-Buffering": "no",
+                    "Connection": "keep-alive",
+                },
+            )
+
+        except Exception as e:
+            logger.error(f"Error in stream endpoint: {e}", exc_info=True)
+            return jsonify({"error": str(e)}), 500
+
+    @app.route("/api/chat/reset", methods=["POST"])
+    def chat_reset():
+        """Reset the chat conversation."""
+        try:
+            assistant = get_or_create_assistant()
+            if assistant:
+                assistant.reset_conversation()
+                return jsonify({"success": True})
+            return jsonify({"error": "Assistant not initialized"}), 400
+        except Exception as e:
+            logger.error(f"Error resetting chat: {e}", exc_info=True)
+            return jsonify({"error": str(e)}), 500
+
+    @app.route("/api/chat/tools", methods=["GET"])
+    def chat_tools():
+        """Get available tools for the coding assistant."""
+        try:
+            assistant = get_or_create_assistant()
+            if assistant:
+                tools = assistant.get_available_tools()
+                return jsonify({"tools": tools})
+            return jsonify({"error": "Assistant not initialized"}), 400
+        except Exception as e:
+            logger.error(f"Error getting tools: {e}", exc_info=True)
+            return jsonify({"error": str(e)}), 500
+
+    @app.route("/api/lsp", methods=["POST"])
+    def lsp_request():
+        """Handle LSP requests via HTTP."""
+        try:
+            message = request.json
+            logger.debug(f"Received LSP message: {message.get('method')}")
+            response = lsp_server.handle_message(message)
+
+            if response:
+                return jsonify(response)
+            return jsonify({"jsonrpc": "2.0", "id": message.get("id"), "result": None})
+        except Exception as e:
+            logger.error(f"Error handling LSP message: {e}")
+            return (
+                jsonify(
+                    {
+                        "jsonrpc": "2.0",
+                        "id": message.get("id"),
+                        "error": {"code": -32603, "message": str(e)},
+                    }
+                ),
+                500,
+            )
+
+    @app.route("/api/lsp/notification", methods=["POST"])
+    def lsp_notification():
+        """Handle LSP notifications via HTTP and return diagnostics."""
+        try:
+            message = request.json
+            method = message.get("method")
+            params = message.get("params", {})
+
+            logger.debug(f"Received LSP notification: {method}")
+
+            # Handle notifications that produce diagnostics
+            diagnostics = []
+            if method == "textDocument/didOpen":
+                text_document = params.get("textDocument", {})
+                uri = text_document.get("uri")
+                text = text_document.get("text")
+                if uri and text:
+                    diagnostics = lsp_server.handler.validate_document(uri, text)
+            elif method == "textDocument/didChange":
+                text_document = params.get("textDocument", {})
+                content_changes = params.get("contentChanges", [])
+                uri = text_document.get("uri")
+                if uri and content_changes:
+                    text = content_changes[0].get("text") if content_changes else None
+                    if text:
+                        diagnostics = lsp_server.handler.validate_document(uri, text)
+            elif method == "textDocument/didClose":
+                text_document = params.get("textDocument", {})
+                uri = text_document.get("uri")
+                if uri:
+                    lsp_server.handler.close_document(uri)
+
+            # Return diagnostics if any
+            if diagnostics:
+                return jsonify({"status": "ok", "diagnostics": diagnostics})
+
+            return jsonify({"status": "ok"})
+        except Exception as e:
+            logger.error(f"Error handling LSP notification: {e}")
+            return jsonify({"error": str(e)}), 500
+
+    # Register config API routes
+    try:
+        import sys
+
+        # Add tactus-ide/backend to path for imports
+        # Path from tactus/ide/server.py -> project root -> tactus-ide/backend
+        backend_dir = Path(__file__).parent.parent.parent / "tactus-ide" / "backend"
+        if backend_dir.exists():
+            sys.path.insert(0, str(backend_dir))
+            from config_server import register_config_routes
+
+            register_config_routes(app)
+        else:
+            logger.warning(f"Config server backend directory not found: {backend_dir}")
+    except ImportError as e:
+        logger.warning(f"Could not register config routes: {e}")
+
+    # Serve frontend if dist directory is provided
+    if frontend_dist_dir:
+
+        @app.route("/")
+        def serve_frontend():
+            """Serve the frontend index.html."""
+            return app.send_static_file("index.html")
+
+        @app.route("/<path:path>")
+        def serve_static_or_frontend(path):
+            """Serve static files or index.html for client-side routing."""
+            # If the file exists, serve it
+            file_path = Path(frontend_dist_dir) / path
+            if file_path.exists() and file_path.is_file():
+                return app.send_static_file(path)
+            # Otherwise, serve index.html for client-side routing (unless it's an API call)
+            if not path.startswith("api/"):
+                return app.send_static_file("index.html")
+            # For API calls that don't match any route, return 404
+            return jsonify({"error": "Not found"}), 404
+
+    return app
+
+
+def main() -> None:
+    """
+    Run the IDE backend server.
+
+    This enables `python -m tactus.ide.server` which is useful for local development
+    and file-watcher based auto-reload workflows.
+
+    Environment variables:
+    - TACTUS_IDE_HOST: Host to bind to (default: 127.0.0.1)
+    - TACTUS_IDE_PORT: Port to bind to (default: 5001)
+    - TACTUS_IDE_WORKSPACE: Initial workspace directory (default: current directory)
+    - TACTUS_IDE_LOG_LEVEL: Logging level (default: INFO)
+    """
+    logging.basicConfig(level=os.environ.get("TACTUS_IDE_LOG_LEVEL", "INFO"))
+
+    host = os.environ.get("TACTUS_IDE_HOST", "127.0.0.1")
+    port_str = os.environ.get("TACTUS_IDE_PORT", "5001")
+    try:
+        port = int(port_str)
+    except ValueError:
+        raise SystemExit(f"Invalid TACTUS_IDE_PORT: {port_str!r}")
+
+    # Get initial workspace from environment or use current directory
+    initial_workspace = os.environ.get("TACTUS_IDE_WORKSPACE")
+    if initial_workspace:
+        logger.info(f"Setting initial workspace to: {initial_workspace}")
+
+    app = create_app(initial_workspace=initial_workspace)
+    # NOTE: We intentionally disable Flask's reloader here; external watchers (e.g. watchdog)
+    # should restart this process to avoid double-fork behavior.
+    app.run(host=host, port=port, debug=False, threaded=True, use_reloader=False)
+
+
+if __name__ == "__main__":
+    main()