Add databricks v2 prompts (#579)

Author: arsenyinfo

klaudbiusz/cli/__init__.py

@@ -20,6 +20,7 @@
     query,
 )
 from dotenv import load_dotenv
+
 from cli.utils.shared import build_mcp_command, validate_mcp_manifest
 
 logger = logging.getLogger(__name__)
@@ -348,6 +349,7 @@ async def analyze_with_agent(
         system_prompt=base_instructions,
         permission_mode="bypassPermissions",
         disallowed_tools=disallowed_tools,
+        model="claude-opus-4-5",
         allowed_tools=[
             "Read",
             "Glob",

klaudbiusz/cli/analyze_trajectories.py

@@ -34,7 +34,7 @@
 
 # Load environment variables from .env file
 env_paths = [
-    Path(__file__).parent.parent.parent / "edda" / ".env",
+    Path(__file__).parent.parent.parent.parent / "edda" / ".env",
     Path(__file__).parent.parent / ".env",
 ]
 for env_path in env_paths:

klaudbiusz/cli/evaluation/evaluate_all.py

@@ -35,7 +35,7 @@
 
 # Load environment variables from .env file - try multiple locations
 env_paths = [
-    Path(__file__).parent.parent.parent / "edda" / ".env",
+    Path(__file__).parent.parent.parent.parent / "edda" / ".env",
     Path(__file__).parent.parent / ".env",
     Path(__file__).parent / ".env",
 ]

klaudbiusz/cli/evaluation/evaluate_app.py

@@ -45,7 +45,7 @@
 
 # Load environment variables
 env_paths = [
-    Path(__file__).parent.parent.parent / "edda" / ".env",
+    Path(__file__).parent.parent.parent.parent / "edda" / ".env",
     Path(__file__).parent.parent / ".env",
     Path(__file__).parent / ".env",
 ]

klaudbiusz/cli/evaluation/evaluate_app_dagger.py

@@ -46,10 +46,10 @@ def run_single_generation(
     backend: str,
     model: str | None,
     wipe_db: bool = False,
-    use_subagents: bool = False,
     suppress_logs: bool = True,
     mcp_binary: str | None = None,
     mcp_json: str | None = None,
+    mcp_args: list[str] | None = None,
     output_dir: str | None = None,
 ) -> RunResult:
     # re-apply litellm patch in worker process (joblib uses spawn/fork)
@@ -70,9 +70,9 @@ def timeout_handler(signum, frame):
                     app_name=app_name,
                     wipe_db=wipe_db,
                     suppress_logs=suppress_logs,
-                    use_subagents=use_subagents,
                     mcp_binary=mcp_binary,
                     mcp_json_path=mcp_json,
+                    mcp_args=mcp_args,
                     output_dir=output_dir,
                 )
                 metrics = codegen.run(prompt, wipe_db=wipe_db)
@@ -85,6 +85,7 @@ def timeout_handler(signum, frame):
                     model=model,
                     mcp_binary=mcp_binary,
                     mcp_json_path=mcp_json,
+                    mcp_args=mcp_args,
                     suppress_logs=suppress_logs,
                     output_dir=output_dir,
                 )
@@ -147,20 +148,20 @@ def main(
     model: str | None = None,
     wipe_db: bool = False,
     n_jobs: int = -1,
-    use_subagents: bool = False,
     mcp_binary: str | None = None,
     mcp_json: str | None = None,
+    mcp_args: list[str] | None = None,
     output_dir: str | None = None,
 ) -> None:
     """Bulk app generation from predefined prompt sets.
 
     Args:
-        prompts: Prompt set to use ("databricks" or "test", default: "databricks")
+        prompts: Prompt set to use ("databricks", "databricks_v2", or "test", default: "databricks")
         backend: Backend to use ("claude" or "litellm", default: "claude")
         model: LLM model (required if backend=litellm, e.g., "openrouter/minimax/minimax-m2")
         wipe_db: Whether to wipe database on start
         n_jobs: Number of parallel jobs (-1 for all cores)
-        use_subagents: Whether to enable subagent delegation (claude backend only)
+        mcp_args: Optional list of args passed to the MCP server (overrides defaults)
         mcp_binary: Optional path to pre-built edda-mcp binary (default: use cargo run)
         mcp_json: Optional path to JSON config file for edda_mcp
         output_dir: Custom output directory for generated apps (default: ./app)
@@ -169,6 +170,9 @@ def main(
         # Claude backend (default) with databricks prompts (default)
         python bulk_run.py
 
+        # Claude backend with databricks_v2 prompts
+        python bulk_run.py --prompts=databricks_v2
+
         # Claude backend with test prompts
         python bulk_run.py --prompts=test
 
@@ -195,10 +199,12 @@ def main(
     match prompts:
         case "databricks":
             from cli.generation.prompts.databricks import PROMPTS as selected_prompts
+        case "databricks_v2":
+            from cli.generation.prompts.databricks_v2 import PROMPTS as selected_prompts
         case "test":
             from cli.generation.prompts.web import PROMPTS as selected_prompts
         case _:
-            raise ValueError(f"Unknown prompt set: {prompts}. Use 'databricks' or 'test'")
+            raise ValueError(f"Unknown prompt set: {prompts}. Use 'databricks', 'databricks_v2', or 'test'")
 
     # validate backend-specific requirements
     if backend == "litellm" and not model:
@@ -216,14 +222,13 @@ def main(
     print(f"Parallel jobs: {n_jobs}")
     if backend == "claude":
         print(f"Wipe DB: {wipe_db}")
-        print(f"Use subagents: {use_subagents}")
     print(f"MCP binary: {mcp_binary if mcp_binary else 'cargo run (default)'}")
     print(f"Output dir: {output_dir if output_dir else './app (default)'}\n")
 
     # generate all apps
     results: list[RunResult] = Parallel(n_jobs=n_jobs, backend="loky", verbose=10)(  # type: ignore[assignment]
         delayed(run_single_generation)(
-            app_name, prompt, backend, model, wipe_db, use_subagents, suppress_logs, mcp_binary, mcp_json, output_dir
+            app_name, prompt, backend, model, wipe_db, suppress_logs, mcp_binary, mcp_json, mcp_args, output_dir
         )
         for app_name, prompt in selected_prompts.items()
     )

klaudbiusz/cli/generation/bulk_run.py

@@ -8,7 +8,6 @@
 from uuid import UUID, uuid4
 
 from claude_agent_sdk import (
-    AgentDefinition,
     AssistantMessage,
     ClaudeAgentOptions,
     ResultMessage,
@@ -19,6 +18,7 @@
     query,
 )
 from dotenv import load_dotenv
+
 from cli.utils.shared import ScaffoldTracker, Tracker, build_mcp_command, setup_logging, validate_mcp_manifest
 
 try:
@@ -44,35 +44,6 @@ class ToolInput:
     prompt: str = ""
 
 
-def _parse_agent_definition(agent_file: Path) -> tuple[dict[str, str], str] | None:
-    """Parse agent markdown file with YAML frontmatter.
-
-    Returns:
-        Tuple of (frontmatter_dict, content) or None if parsing fails
-    """
-    if not agent_file.exists():
-        return None
-
-    content = agent_file.read_text()
-
-    # frontmatter must start and end with ---
-    if not content.startswith("---"):
-        return None
-
-    parts = content.split("---", 2)
-    if len(parts) < 3:
-        return None
-
-    # parse simple yaml-like frontmatter manually
-    frontmatter = {}
-    for line in parts[1].strip().split("\n"):
-        if ":" in line:
-            key, value = line.split(":", 1)
-            frontmatter[key.strip()] = value.strip()
-
-    return frontmatter, parts[2].strip()
-
-
 class GenerationMetrics(TypedDict):
     cost_usd: float
     input_tokens: int
@@ -88,22 +59,22 @@ def __init__(
         app_name: str,
         wipe_db: bool = True,
         suppress_logs: bool = False,
-        use_subagents: bool = False,
         mcp_binary: str | None = None,
         mcp_json_path: str | None = None,
+        mcp_args: list[str] | None = None,
         output_dir: str | None = None,
     ):
         load_dotenv()
-        self.project_root = Path(__file__).parent.parent.parent
+        self.project_root = Path(__file__).parent.parent.parent.parent
         self.mcp_manifest = validate_mcp_manifest(mcp_binary, self.project_root)
 
         self.wipe_db = wipe_db
         self.run_id: UUID = uuid4()
         self.app_name = app_name
-        self.use_subagents = use_subagents
         self.suppress_logs = suppress_logs
         self.mcp_binary = mcp_binary
         self.mcp_json_path = mcp_json_path
+        self.mcp_args = mcp_args
         self.output_dir = Path(output_dir) if output_dir else Path.cwd() / "app"
         self.tracker = Tracker(self.run_id, app_name, suppress_logs)
         self.scaffold_tracker = ScaffoldTracker()
@@ -115,43 +86,23 @@ async def run_async(self, prompt: str) -> GenerationMetrics:
         await self.tracker.init(wipe_db=self.wipe_db)
 
         agents = {}
-        if self.use_subagents:
-            agents_dir = self.project_root / "klaudbiusz" / "agents"
-            dataresearch_file = agents_dir / "dataresearch.md"
-
-            if parsed := _parse_agent_definition(dataresearch_file):
-                frontmatter, content = parsed
-                tools_str = frontmatter.get("tools", "")
-                tools = [t.strip() for t in tools_str.split(",")] if tools_str else None
-
-                agents["dataresearch"] = AgentDefinition(
-                    description=frontmatter.get("description", ""),
-                    prompt=content,
-                    tools=tools,
-                    model=frontmatter.get("model"),  # type: ignore[arg-type]
-                )
 
         # workflow and template best practices are now in the MCP tool description
-        base_instructions = "Use Edda MCP tools to scaffold, build, and test the app as needed.\n Use data from Databricks when relevant.\n"
-
-        if self.use_subagents:
-            base_instructions += """When you need to explore Databricks tables, schemas, or execute SQL queries, use the Task tool to delegate to the 'dataresearch' subagent. Do NOT use databricks_* tools directly.\n"""
-
-        base_instructions += """Be concise and to the point in your responses.\n
-Use up to 10 tools per call to speed up the process.\n"""
+        base_instructions = """Use MCP tools to scaffold, build, and test the app as needed.
+Use data from Databricks when relevant.
+Be concise and to the point in your responses.
+Use up to 10 tools per call to speed up the process.
+Never deploy the app, just scaffold and build it.
+"""
 
         disallowed_tools = [
             "NotebookEdit",
             "WebSearch",
             "WebFetch",
         ]
 
-        # NOTE: We cannot use disallowed_tools to block Databricks tools from the main agent
-        # because disallowed_tools applies globally to ALL agents (including subagents).
-        # The CLI doesn't support per-agent tool permissions yet.
-        # Instead, we rely on system prompt instructions to enforce delegation.
+        command, args = build_mcp_command(self.mcp_binary, self.mcp_manifest, self.mcp_json_path, self.mcp_args)
 
-        command, args = build_mcp_command(self.mcp_binary, self.mcp_manifest, self.mcp_json_path)
         mcp_config = {
             "type": "stdio",
             "command": command,

klaudbiusz/cli/generation/codegen.py

@@ -34,7 +34,7 @@ class MCPSession:
     def __init__(self, mcp_binary: str | None = None, mcp_json_path: str | None = None):
         self.mcp_binary = mcp_binary
         self.mcp_json_path = mcp_json_path
-        self.project_root = Path(__file__).parent.parent.parent
+        self.project_root = Path(__file__).parent.parent.parent.parent
         self.mcp_manifest = validate_mcp_manifest(mcp_binary, self.project_root)
 
         self._context = None