Daggerize klaudbiusz bulk generation

klaudbiusz/.dockerignore

@@ -0,0 +1,25 @@
+# ignore generated apps and artifacts
+app/
+app-eval/
+results/
+
+# ignore python cache
+__pycache__/
+*.py[cod]
+*$py.class
+*.so
+.Python
+.ruff_cache/
+
+# ignore git
+.git/
+.gitignore
+
+# ignore env files (will be mounted)
+.env
+
+# ignore venv
+.venv/
+
+# ignore build artifacts
+.DS_Store

klaudbiusz/Dockerfile

@@ -0,0 +1,37 @@
+FROM python:3.12-slim
+
+# install databricks cli and node.js (for claude agent sdk)
+RUN apt-get update && \
+    apt-get install -y curl unzip git && \
+    curl -fsSL https://raw.githubusercontent.com/databricks/setup-cli/main/install.sh | sh && \
+    curl -fsSL https://deb.nodesource.com/setup_22.x | bash - && \
+    apt-get install -y nodejs && \
+    npm install -g @anthropic-ai/claude-code && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
+
+WORKDIR /workspace
+
+# install uv for package management
+RUN pip install uv
+
+# install dependencies (cached until pyproject.toml changes)
+COPY pyproject.toml README.md ./
+RUN mkdir -p cli && touch cli/__init__.py && \
+    uv pip install --system -e . && \
+    rm -rf cli
+
+# copy only generation-related source code (exclude evaluation to prevent reward hacking)
+COPY cli/__init__.py ./cli/
+COPY cli/trajectory.py ./cli/
+COPY cli/generation/ ./cli/generation/
+COPY cli/utils/ ./cli/utils/
+
+# create non-root user (claude agent sdk requires non-root for security)
+RUN useradd -m -s /bin/bash klaudbiusz && \
+    chown -R klaudbiusz:klaudbiusz /workspace
+
+USER klaudbiusz
+
+# set working directory for app generation
+ENV APP_OUTPUT_DIR=/workspace/app

klaudbiusz/README.md

@@ -21,42 +21,55 @@ cp .env.example .env
 ```
 
 ### Generate Applications
+
+Generation runs inside Dagger containers for isolation and reproducibility.
+
+**Prerequisites:**
+- Docker running
+- Linux build of edda_mcp binary (for Dagger containers)
+- Databricks CLI OAuth configured (`~/.databrickscfg` + `~/.databricks/token-cache.json`)
+
 ```bash
 cd klaudbiusz
 
-
 # make sure app folder is empty
 cli/archive_evaluation.sh
 cli/cleanup_evaluation.sh
 
-# Generate a single app (Claude backend, default)
-uv run cli/generation/single_run.py "Create a customer churn analysis dashboard"
-
-# Use LiteLLM backend with specific model
+# Generate a single app via Dagger (requires Linux binary)
 uv run cli/generation/single_run.py "Create a customer churn analysis dashboard" \
-  --backend=litellm --model=openrouter/minimax/minimax-m2
+  --mcp_binary=/path/to/linux/edda_mcp \
+  --mcp_args='["experimental", "apps-mcp"]'
 
-# Batch generate from prompts (databricks set by default)
-uv run cli/generation/bulk_run.py
+# Batch generate from prompts
+uv run cli/generation/bulk_run.py \
+  --mcp_binary=/path/to/linux/edda_mcp \
+  --mcp_args='["experimental", "apps-mcp"]'
 
-# Batch generate with test prompts
-uv run cli/generation/bulk_run.py --prompts=test
+# Use LiteLLM backend with specific model
+uv run cli/generation/single_run.py "Create a customer churn analysis dashboard" \
+  --backend=litellm --model=gemini/gemini-2.5-pro \
+  --mcp_binary=/path/to/linux/edda_mcp
+```
 
-# Batch generate with LiteLLM backend
-uv run cli/generation/bulk_run.py --backend=litellm --model=gemini/gemini-2.5-pro
+**Building Linux binary (for macOS users):**
+```bash
+cd /path/to/cli
+GOOS=linux GOARCH=arm64 go build -o cli-linux .
+# Then use --mcp_binary=/path/to/cli-linux
+```
 
-# Custom output directory
-uv run cli/generation/bulk_run.py --output-dir=/path/to/custom/folder
+### Local Debugging (without Dagger)
 
-# Custom MCP binary (for testing modified edda_mcp)
-uv run cli/generation/bulk_run.py --mcp-binary=/path/to/custom/edda_mcp
+For faster iteration during development, run directly on host:
 
-# Combined example
-uv run cli/generation/bulk_run.py \
-  --backend=litellm \
-  --model=gemini/gemini-2.5-pro \
-  --output-dir=./my-apps \
-  --mcp-binary=../edda/target/release/edda_mcp
+```bash
+# Local run with macOS binary
+uv run python cli/generation/container_runner.py "Create a dashboard" \
+  --app_name=debug-test \
+  --mcp_binary=/usr/local/bin/edda_mcp \
+  --mcp_args='["experimental", "apps-mcp"]' \
+  --output_dir=./app
 ```
 
 ### Evaluate Generated Apps
@@ -155,8 +168,10 @@ klaudbiusz/
 │   │   ├── prompts/               # Prompt collections
 │   │   ├── codegen.py             # Claude Agent SDK backend
 │   │   ├── codegen_multi.py       # LiteLLM backend
-│   │   ├── single_run.py          # Single app generation
-│   │   ├── bulk_run.py            # Batch app generation
+│   │   ├── dagger_run.py          # Dagger container orchestration
+│   │   ├── container_runner.py    # Runner script (inside container or local)
+│   │   ├── single_run.py          # Single app generation (via Dagger)
+│   │   ├── bulk_run.py            # Batch app generation (via Dagger)
 │   │   └── screenshot.py          # Batch screenshotting
 │   ├── evaluation/                 # App evaluation
 │   │   ├── evaluate_all.py        # Batch evaluation

klaudbiusz/cli/analyze_trajectories.py

@@ -136,7 +136,12 @@ async def analyze_single_trajectory(trajectory_md: str, app_name: str, model: st
     return response.choices[0].message.content  # type: ignore[attr-defined]
 
 
-def get_mcp_tools_description(mcp_binary: str | None, project_root: Path, mcp_json_path: str | None = None) -> str:
+def get_mcp_tools_description(
+    mcp_binary: str | None,
+    project_root: Path,
+    mcp_json_path: str | None = None,
+    mcp_args: list[str] | None = None,
+) -> str:
     """Extract MCP tool definitions by querying the MCP server.
 
     Returns empty string if mcp_binary is not provided.
@@ -145,7 +150,7 @@ def get_mcp_tools_description(mcp_binary: str | None, project_root: Path, mcp_js
         return ""
 
     mcp_manifest = validate_mcp_manifest(mcp_binary, project_root)
-    command, args = build_mcp_command(mcp_binary, mcp_manifest, mcp_json_path)
+    command, args = build_mcp_command(mcp_binary, mcp_manifest, mcp_json_path, mcp_args)
 
     proc = subprocess.Popen(
         [command, *args],
@@ -403,6 +408,7 @@ async def analyze_trajectories_async(
     trajectories_pattern: str = "./app/*/trajectory.jsonl",
     eval_report_path: str | None = None,
     mcp_json_path: str | None = None,
+    mcp_args: list[str] | None = None,
 ):
     """Analyze trajectories using map-reduce approach with LLM, then agent-based analysis."""
     litellm.drop_params = True
@@ -414,7 +420,7 @@ async def analyze_trajectories_async(
     mcp_tools_doc = ""
     if mcp_binary:
         logger.info("📋 Extracting MCP tool definitions")
-        mcp_tools_doc = get_mcp_tools_description(mcp_binary, project_root, mcp_json_path)
+        mcp_tools_doc = get_mcp_tools_description(mcp_binary, project_root, mcp_json_path, mcp_args)
 
     eval_report = ""
     if eval_report_path:
@@ -460,6 +466,7 @@ def cli(
     map_model: str = "anthropic/claude-haiku-4-5",
     eval_report: str | None = None,
     mcp_json: str | None = None,
+    mcp_args: list[str] | None = None,
 ):
     """Analyze agent trajectories to find friction points and patterns.
 
@@ -472,6 +479,7 @@ def cli(
         map_model: LiteLLM model identifier for individual trajectory analysis
         eval_report: Path to evaluation report JSON (optional)
         mcp_json: Optional path to JSON config file for edda_mcp
+        mcp_args: Optional list of args passed to the MCP server (overrides defaults)
     """
     coloredlogs.install(
         level=logging.INFO,
@@ -490,6 +498,7 @@ def cli(
             trajectories_pattern,
             eval_report,
             mcp_json,
+            mcp_args,
         )
     )
 

klaudbiusz/cli/evaluation/evaluate_app_dagger.py

@@ -7,6 +7,7 @@
 
 import asyncio
 import json
+import os
 import sys
 import time
 from dataclasses import asdict
@@ -55,6 +56,11 @@
         break
 
 
+def _restore_terminal_cursor() -> None:
+    """Restore terminal cursor after Dagger run (workaround for dagger/dagger#7160)."""
+    os.system("tput cnorm 2>/dev/null || true")
+
+
 async def evaluate_app_async(
     client: dagger.Client,
     app_dir: Path,
@@ -390,7 +396,10 @@ async def main_async():
 
 def main():
     """Sync wrapper for async main."""
-    asyncio.run(main_async())
+    try:
+        asyncio.run(main_async())
+    finally:
+        _restore_terminal_cursor()
 
 
 if __name__ == "__main__":