Atomizer/docs/plans/CLAUDE_CANVAS_PROJECT.md

# Claude + Canvas Integration Project

## Project Overview

**Project Name**: Unified Claude + Canvas Integration
**Goal**: Transform Atomizer's dashboard into a bi-directional Claude + Canvas experience where Claude and the user co-edit the same `atomizer_spec.json` in real-time.

**Core Principle**: `atomizer_spec.json` is the single source of truth. Both Claude (via tools) and the user (via canvas UI) read and write to it, with changes instantly reflected on both sides.

---

## Current State

### What Exists (✅ Working)

| Component | File | Status |
|-----------|------|--------|
| Power Mode WebSocket | `backend/api/routes/claude.py` | ✅ `/ws/power` endpoint |
| Write Tools | `backend/api/services/claude_agent.py` | ✅ 6 tools implemented |
| `spec_modified` Events | `backend/api/routes/claude.py:480-487` | ✅ Sent on tool use |
| Canvas Reload | `frontend/src/hooks/useChat.ts:247-262` | ✅ Triggers `reloadSpec()` |
| Spec Store | `frontend/src/hooks/useSpecStore.ts` | ✅ Manages spec state |
| Spec Renderer | `frontend/src/components/canvas/SpecRenderer.tsx` | ✅ Renders spec as nodes |

### What's Missing (❌ To Build)

| Feature | Priority | Effort |
|---------|----------|--------|
| Canvas state in Claude's context | P0 | Easy |
| Full spec in `spec_updated` payload | P0 | Easy |
| User edits notify Claude | P0 | Easy |
| Streaming responses | P1 | Medium |
| `create_study` tool | P1 | Medium |
| Interview Engine | P2 | Medium |
| Tool call UI indicators | P2 | Easy |
| Node animations | P3 | Easy |

---

## Architecture

### Target Data Flow

```
┌─────────────────────────────────────────────────────────────────────┐
│                         atomizer_spec.json                           │
│                      (Single Source of Truth)                        │
└───────────────────────────────┬─────────────────────────────────────┘
                                │
                    ┌───────────┴───────────┐
                    │                       │
                    ▼                       ▼
┌───────────────────────────┐   ┌───────────────────────────┐
│      Claude Agent         │   │         Canvas UI          │
│  (AtomizerClaudeAgent)    │   │      (SpecRenderer)        │
├───────────────────────────┤   ├───────────────────────────┤
│ • Reads spec for context  │   │ • Renders spec as nodes   │
│ • Writes via tools        │   │ • User edits nodes        │
│ • Sees user's edits       │   │ • Receives Claude's edits │
└───────────────────────────┘   └───────────────────────────┘
                    │                       │
                    │     WebSocket         │
                    │   (bi-directional)    │
                    │                       │
                    └───────────┬───────────┘
                                │
                    ┌───────────┴───────────┐
                    │   /api/claude/ws/power │
                    │   (Enhanced Endpoint)  │
                    └───────────────────────┘
```

### Message Protocol (Enhanced)

**Client → Server:**
```typescript
// Chat message
{ type: "message", content: "Add a thickness variable 2-10mm" }

// User edited canvas (NEW)
{ type: "canvas_edit", spec: { /* full spec */ } }

// Switch study
{ type: "set_study", study_id: "bracket_v1" }

// Heartbeat
{ type: "ping" }
```

**Server → Client:**
```typescript
// Streaming text (NEW - replaces single "text" message)
{ type: "text_delta", content: "Adding" }
{ type: "text_delta", content: " thickness" }
{ type: "text_delta", content: " variable..." }

// Tool started (NEW)
{ type: "tool_start", tool: "add_design_variable", input: {...} }

// Tool completed
{ type: "tool_result", tool: "add_design_variable", result: "✓ Added..." }

// Spec updated - NOW INCLUDES FULL SPEC (CHANGED)
{ type: "spec_updated", spec: { /* full atomizer_spec.json */ } }

// Response complete
{ type: "done" }

// Error
{ type: "error", message: "..." }

// Heartbeat response
{ type: "pong" }
```

---

## Implementation Tasks

### Phase 1: Core Bi-directional Sync (P0)

#### Task 1.1: Add Canvas State to Claude's Context
**File**: `backend/api/services/claude_agent.py`

**Current** (`_build_system_prompt`):
```python
if self.study_id and self.study_dir and self.study_dir.exists():
    context = self._get_study_context()
    base_prompt += f"\n## Current Study: {self.study_id}\n{context}\n"
```

**Change**: Add method to format current spec as context, call it from WebSocket handler.

```python
def set_canvas_state(self, spec: Dict[str, Any]) -> None:
    """Update the current canvas state for context"""
    self.canvas_state = spec

def _format_canvas_context(self) -> str:
    """Format current canvas state for Claude's system prompt"""
    if not self.canvas_state:
        return ""

    spec = self.canvas_state
    lines = ["\n## Current Canvas State\n"]
    lines.append("The user can see this canvas. When you modify it, they see changes in real-time.\n")

    # Design Variables
    dvs = spec.get('design_variables', [])
    if dvs:
        lines.append(f"**Design Variables ({len(dvs)}):**")
        for dv in dvs:
            bounds = dv.get('bounds', {})
            lines.append(f"  - `{dv.get('id')}`: {dv.get('name')} [{bounds.get('min')}, {bounds.get('max')}]")

    # Extractors
    exts = spec.get('extractors', [])
    if exts:
        lines.append(f"\n**Extractors ({len(exts)}):**")
        for ext in exts:
            lines.append(f"  - `{ext.get('id')}`: {ext.get('name')} ({ext.get('type')})")

    # Objectives
    objs = spec.get('objectives', [])
    if objs:
        lines.append(f"\n**Objectives ({len(objs)}):**")
        for obj in objs:
            lines.append(f"  - `{obj.get('id')}`: {obj.get('name')} ({obj.get('direction')})")

    # Constraints
    cons = spec.get('constraints', [])
    if cons:
        lines.append(f"\n**Constraints ({len(cons)}):**")
        for con in cons:
            lines.append(f"  - `{con.get('id')}`: {con.get('name')} {con.get('operator')} {con.get('threshold')}")

    # Model
    model = spec.get('model', {})
    if model.get('sim', {}).get('path'):
        lines.append(f"\n**Model**: {model['sim']['path']}")

    return "\n".join(lines)
```

**Update `_build_system_prompt`**:
```python
def _build_system_prompt(self) -> str:
    base_prompt = """..."""  # existing prompt

    # Add study context
    if self.study_id and self.study_dir and self.study_dir.exists():
        context = self._get_study_context()
        base_prompt += f"\n## Current Study: {self.study_id}\n{context}\n"

    # Add canvas state (NEW)
    canvas_context = self._format_canvas_context()
    if canvas_context:
        base_prompt += canvas_context

    return base_prompt
```

---

#### Task 1.2: Send Full Spec in `spec_updated`
**File**: `backend/api/routes/claude.py`

**Current** (line 483-487):
```python
await websocket.send_json({
    "type": "spec_modified",
    "tool": tool_call["tool"],
    "changes": tool_call["result_preview"],
})
```

**Change**: Send full spec instead of just changes.

```python
# After any write tool completes, send full spec
if tool_call["tool"] in ["add_design_variable", "add_extractor",
                          "add_objective", "add_constraint",
                          "update_spec_field", "remove_node"]:
    # Load the updated spec
    spec = agent.load_current_spec()
    await websocket.send_json({
        "type": "spec_updated",
        "tool": tool_call["tool"],
        "spec": spec,  # Full spec!
    })
```

**Add to `AtomizerClaudeAgent`**:
```python
def load_current_spec(self) -> Optional[Dict[str, Any]]:
    """Load the current atomizer_spec.json"""
    if not self.study_dir:
        return None
    spec_path = self.study_dir / "atomizer_spec.json"
    if not spec_path.exists():
        return None
    with open(spec_path, 'r', encoding='utf-8') as f:
        return json.load(f)
```

---

#### Task 1.3: Handle User Canvas Edits
**File**: `backend/api/routes/claude.py`

**Add to `power_mode_websocket`** (after line 524):
```python
elif data.get("type") == "canvas_edit":
    # User made a manual edit to the canvas
    spec = data.get("spec")
    if spec:
        # Update agent's canvas state so Claude sees the change
        agent.set_canvas_state(spec)
        # Optionally save to file if the frontend already saved
        # (or let frontend handle saving)
        await websocket.send_json({
            "type": "canvas_edit_received",
            "acknowledged": True
        })
```

---

#### Task 1.4: Frontend - Send Canvas Edits to Claude
**File**: `frontend/src/hooks/useChat.ts`

**Add to state/hook**:
```typescript
// Add to sendMessage or create new function
const notifyCanvasEdit = useCallback((spec: AtomizerSpec) => {
  if (wsRef.current?.readyState === WebSocket.OPEN) {
    wsRef.current.send(JSON.stringify({
      type: 'canvas_edit',
      spec: spec
    }));
  }
}, []);
```

**Return from hook**:
```typescript
return {
  // ... existing
  notifyCanvasEdit,  // NEW
};
```

---

#### Task 1.5: Frontend - Use Full Spec from `spec_updated`
**File**: `frontend/src/hooks/useChat.ts`

**Current** (line 247-262):
```typescript
case 'spec_modified':
  console.log('[useChat] Spec was modified by assistant:', data.tool, data.changes);
  if (onCanvasModification) {
    onCanvasModification({
      action: 'add_node',
      data: { _refresh: true, tool: data.tool, changes: data.changes },
    });
  }
  break;
```

**Change**: Use the spec directly instead of triggering reload.

```typescript
case 'spec_updated':
  console.log('[useChat] Spec updated by assistant:', data.tool);
  // Directly update spec store instead of triggering HTTP reload
  if (data.spec && onSpecUpdated) {
    onSpecUpdated(data.spec);
  }
  break;
```

**Add callback to hook options**:
```typescript
interface UseChatOptions {
  // ... existing
  onSpecUpdated?: (spec: AtomizerSpec) => void;  // NEW
}
```

---

#### Task 1.6: Wire Canvas to Use Direct Spec Updates
**File**: `frontend/src/pages/CanvasView.tsx`

**Current** (line 57-64):
```typescript
onCanvasModification: chatPowerMode ? (modification) => {
  console.log('Canvas modification from Claude:', modification);
  showNotification(`Claude: ${modification.action}...`);
  reloadSpec();
} : undefined,
```

**Change**: Use `onSpecUpdated` callback.

```typescript
const { setSpec } = useSpecStore();

// In useChat options:
onSpecUpdated: chatPowerMode ? (spec) => {
  console.log('Spec updated by Claude');
  setSpec(spec);  // Direct update, no HTTP reload
  showNotification('Canvas updated by Claude');
} : undefined,
```

---

### Phase 2: Streaming Responses (P1)

#### Task 2.1: Implement Streaming in Claude Agent
**File**: `backend/api/services/claude_agent.py`

**Add new method**:
```python
async def chat_stream(
    self,
    message: str,
    conversation_history: List[Dict[str, Any]]
) -> AsyncGenerator[Dict[str, Any], None]:
    """
    Stream chat response with tool calls.
    Yields events: text_delta, tool_start, tool_result, done
    """
    # Rebuild system prompt with current canvas state
    self.system_prompt = self._build_system_prompt()

    messages = conversation_history + [{"role": "user", "content": message}]

    # Use streaming API
    with self.client.messages.stream(
        model="claude-sonnet-4-20250514",
        max_tokens=4096,
        system=self.system_prompt,
        messages=messages,
        tools=self.tools
    ) as stream:
        for event in stream:
            if event.type == "content_block_delta":
                if hasattr(event.delta, "text"):
                    yield {"type": "text_delta", "content": event.delta.text}

        # Get final response for tool handling
        response = stream.get_final_message()

        # Process tool calls
        for block in response.content:
            if block.type == "tool_use":
                yield {"type": "tool_start", "tool": block.name, "input": block.input}

                # Execute tool
                result = self._execute_tool_sync(block.name, block.input)

                yield {"type": "tool_result", "tool": block.name, "result": result}

                # If spec changed, yield updated spec
                if block.name in ["add_design_variable", "add_extractor",
                                  "add_objective", "add_constraint",
                                  "update_spec_field", "remove_node"]:
                    spec = self.load_current_spec()
                    if spec:
                        yield {"type": "spec_updated", "spec": spec}

        yield {"type": "done"}
```

---

#### Task 2.2: Use Streaming in WebSocket Handler
**File**: `backend/api/routes/claude.py`

**Change `power_mode_websocket`** (line 464-501):
```python
if data.get("type") == "message":
    content = data.get("content", "")
    if not content:
        continue

    try:
        # Stream the response
        async for event in agent.chat_stream(content, conversation_history):
            await websocket.send_json(event)

        # Update conversation history
        # (need to track this differently with streaming)

    except Exception as e:
        import traceback
        traceback.print_exc()
        await websocket.send_json({
            "type": "error",
            "message": str(e),
        })
```

---

#### Task 2.3: Frontend - Handle Streaming Text
**File**: `frontend/src/hooks/useChat.ts`

**Add state for streaming**:
```typescript
const [streamingText, setStreamingText] = useState<string>('');
```

**Handle `text_delta`**:
```typescript
case 'text_delta':
  setStreamingText(prev => prev + data.content);
  break;

case 'done':
  // Finalize the streaming message
  if (streamingText) {
    setState(prev => ({
      ...prev,
      messages: [...prev.messages, {
        id: Date.now().toString(),
        role: 'assistant',
        content: streamingText
      }]
    }));
    setStreamingText('');
  }
  setState(prev => ({ ...prev, isThinking: false }));
  break;
```

---

### Phase 3: Study Creation (P1)

#### Task 3.1: Add `create_study` Tool
**File**: `backend/api/services/claude_agent.py`

**Add to `_define_tools()`**:
```python
{
    "name": "create_study",
    "description": "Create a new optimization study with directory structure and atomizer_spec.json. Use this when the user wants to start a new optimization from scratch.",
    "input_schema": {
        "type": "object",
        "properties": {
            "name": {
                "type": "string",
                "description": "Study name in snake_case (e.g., 'bracket_mass_v1')"
            },
            "category": {
                "type": "string",
                "description": "Category folder (e.g., 'Simple_Bracket', 'M1_Mirror'). Optional."
            },
            "model_path": {
                "type": "string",
                "description": "Path to NX simulation file (.sim). Optional, can be set later."
            },
            "description": {
                "type": "string",
                "description": "Brief description of the optimization goal"
            }
        },
        "required": ["name"]
    }
},
```

**Add implementation**:
```python
def _tool_create_study(self, params: Dict[str, Any]) -> str:
    """Create a new study with directory structure and atomizer_spec.json"""
    study_name = params['name']
    category = params.get('category', '')
    model_path = params.get('model_path', '')
    description = params.get('description', '')

    # Build study path
    if category:
        study_dir = STUDIES_DIR / category / study_name
    else:
        study_dir = STUDIES_DIR / study_name

    # Check if exists
    if study_dir.exists():
        return f"✗ Study '{study_name}' already exists at {study_dir}"

    # Create directory structure
    study_dir.mkdir(parents=True, exist_ok=True)
    (study_dir / "1_setup").mkdir(exist_ok=True)
    (study_dir / "2_iterations").mkdir(exist_ok=True)
    (study_dir / "3_results").mkdir(exist_ok=True)

    # Create initial spec
    spec = {
        "meta": {
            "version": "2.0",
            "study_name": study_name,
            "description": description,
            "created_at": datetime.now().isoformat(),
            "created_by": "claude_agent"
        },
        "model": {
            "sim": {
                "path": model_path,
                "solver": "nastran"
            }
        },
        "design_variables": [],
        "extractors": [],
        "objectives": [],
        "constraints": [],
        "optimization": {
            "algorithm": {"type": "TPE"},
            "budget": {"max_trials": 100}
        },
        "canvas": {
            "edges": [],
            "layout_version": "2.0"
        }
    }

    # Save spec
    spec_path = study_dir / "atomizer_spec.json"
    with open(spec_path, 'w', encoding='utf-8') as f:
        json.dump(spec, f, indent=2)

    # Update agent context
    self.study_id = f"{category}/{study_name}" if category else study_name
    self.study_dir = study_dir
    self.canvas_state = spec

    return f"✓ Created study '{study_name}' at {study_dir}\n\nThe canvas is now showing this empty study. You can start adding design variables, extractors, and objectives."
```

**Add to tool dispatcher**:
```python
elif tool_name == "create_study":
    return self._tool_create_study(tool_input)
```

---

### Phase 4: Interview Engine (P2)

#### Task 4.1: Create Interview Engine Class
**File**: `backend/api/services/interview_engine.py` (NEW)

```python
"""
Interview Engine for guided study creation.

Walks the user through creating an optimization study step-by-step,
building the atomizer_spec.json incrementally.
"""

from typing import Dict, List, Optional, Any
from dataclasses import dataclass, field
from enum import Enum
import re


class InterviewPhase(Enum):
    WELCOME = "welcome"
    MODEL = "model"
    OBJECTIVES = "objectives"
    DESIGN_VARS = "design_vars"
    CONSTRAINTS = "constraints"
    METHOD = "method"
    REVIEW = "review"
    COMPLETE = "complete"


@dataclass
class InterviewState:
    phase: InterviewPhase = InterviewPhase.WELCOME
    collected: Dict[str, Any] = field(default_factory=dict)
    spec: Dict[str, Any] = field(default_factory=dict)
    model_expressions: List[Dict] = field(default_factory=list)


class InterviewEngine:
    """Guided study creation through conversation"""

    PHASE_QUESTIONS = {
        InterviewPhase.WELCOME: "What kind of optimization do you want to set up? (e.g., minimize mass of a bracket, reduce wavefront error of a mirror)",
        InterviewPhase.MODEL: "What's the path to your NX simulation file (.sim)?\n(You can type the path or I can help you find it)",
        InterviewPhase.OBJECTIVES: "What do you want to optimize?\n\nCommon objectives:\n- Minimize mass/weight\n- Minimize displacement (maximize stiffness)\n- Minimize stress\n- Minimize wavefront error (WFE)\n\nYou can have multiple objectives (multi-objective optimization).",
        InterviewPhase.DESIGN_VARS: "Which parameters should vary during optimization?\n\n{suggestions}",
        InterviewPhase.CONSTRAINTS: "Any constraints to respect?\n\nExamples:\n- Maximum stress ≤ 200 MPa\n- Minimum frequency ≥ 50 Hz\n- Maximum mass ≤ 5 kg\n\n(Say 'none' if no constraints)",
        InterviewPhase.METHOD: "Based on your setup, I recommend **{method}**.\n\nReason: {reason}\n\nShould I use this method?",
        InterviewPhase.REVIEW: "Here's your configuration:\n\n{summary}\n\nReady to create the study? (yes/no)",
    }

    def __init__(self):
        self.state = InterviewState()
        self._init_spec()

    def _init_spec(self):
        """Initialize empty spec structure"""
        self.state.spec = {
            "meta": {"version": "2.0"},
            "model": {"sim": {"path": "", "solver": "nastran"}},
            "design_variables": [],
            "extractors": [],
            "objectives": [],
            "constraints": [],
            "optimization": {
                "algorithm": {"type": "TPE"},
                "budget": {"max_trials": 100}
            },
            "canvas": {"edges": [], "layout_version": "2.0"}
        }

    def get_current_question(self) -> str:
        """Get the question for the current phase"""
        question = self.PHASE_QUESTIONS.get(self.state.phase, "")

        # Dynamic substitutions
        if self.state.phase == InterviewPhase.DESIGN_VARS:
            suggestions = self._format_dv_suggestions()
            question = question.format(suggestions=suggestions)
        elif self.state.phase == InterviewPhase.METHOD:
            method, reason = self._recommend_method()
            question = question.format(method=method, reason=reason)
        elif self.state.phase == InterviewPhase.REVIEW:
            summary = self._format_summary()
            question = question.format(summary=summary)

        return question

    def process_answer(self, answer: str) -> Dict[str, Any]:
        """Process user's answer and advance interview"""
        phase = self.state.phase
        result = {
            "phase": phase.value,
            "spec_changes": [],
            "next_phase": None,
            "question": None,
            "complete": False,
            "error": None
        }

        try:
            if phase == InterviewPhase.WELCOME:
                self._process_welcome(answer)
            elif phase == InterviewPhase.MODEL:
                self._process_model(answer, result)
            elif phase == InterviewPhase.OBJECTIVES:
                self._process_objectives(answer, result)
            elif phase == InterviewPhase.DESIGN_VARS:
                self._process_design_vars(answer, result)
            elif phase == InterviewPhase.CONSTRAINTS:
                self._process_constraints(answer, result)
            elif phase == InterviewPhase.METHOD:
                self._process_method(answer, result)
            elif phase == InterviewPhase.REVIEW:
                self._process_review(answer, result)

            # Advance to next phase
            self._advance_phase()

            if self.state.phase == InterviewPhase.COMPLETE:
                result["complete"] = True
            else:
                result["next_phase"] = self.state.phase.value
                result["question"] = self.get_current_question()

        except Exception as e:
            result["error"] = str(e)

        return result

    def _advance_phase(self):
        """Move to next phase"""
        phases = list(InterviewPhase)
        current_idx = phases.index(self.state.phase)
        if current_idx < len(phases) - 1:
            self.state.phase = phases[current_idx + 1]

    def _process_welcome(self, answer: str):
        """Extract optimization type from welcome"""
        self.state.collected["goal"] = answer

        # Try to infer study name
        words = answer.lower().split()
        if "bracket" in words:
            self.state.collected["geometry_type"] = "bracket"
        elif "mirror" in words:
            self.state.collected["geometry_type"] = "mirror"
        elif "beam" in words:
            self.state.collected["geometry_type"] = "beam"

    def _process_model(self, answer: str, result: Dict):
        """Extract model path"""
        # Extract path from answer
        path = answer.strip().strip('"').strip("'")
        self.state.spec["model"]["sim"]["path"] = path
        self.state.collected["model_path"] = path
        result["spec_changes"].append(f"Set model path: {path}")

    def _process_objectives(self, answer: str, result: Dict):
        """Extract objectives from natural language"""
        answer_lower = answer.lower()
        objectives = []
        extractors = []

        # Mass/weight
        if any(w in answer_lower for w in ["mass", "weight", "light"]):
            extractors.append({
                "id": "ext_mass",
                "name": "Mass",
                "type": "bdf_mass",
                "enabled": True
            })
            objectives.append({
                "id": "obj_mass",
                "name": "Mass",
                "direction": "minimize",
                "source": {"extractor_id": "ext_mass", "output_key": "mass"},
                "enabled": True
            })
            result["spec_changes"].append("Added objective: minimize mass")

        # Displacement/stiffness
        if any(w in answer_lower for w in ["displacement", "stiff", "deflection"]):
            extractors.append({
                "id": "ext_disp",
                "name": "Max Displacement",
                "type": "displacement",
                "config": {"node_set": "all", "direction": "magnitude"},
                "enabled": True
            })
            objectives.append({
                "id": "obj_disp",
                "name": "Max Displacement",
                "direction": "minimize",
                "source": {"extractor_id": "ext_disp", "output_key": "max_displacement"},
                "enabled": True
            })
            result["spec_changes"].append("Added objective: minimize displacement")

        # Stress
        if "stress" in answer_lower and "constraint" not in answer_lower:
            extractors.append({
                "id": "ext_stress",
                "name": "Max Stress",
                "type": "stress",
                "config": {"stress_type": "von_mises"},
                "enabled": True
            })
            objectives.append({
                "id": "obj_stress",
                "name": "Max Stress",
                "direction": "minimize",
                "source": {"extractor_id": "ext_stress", "output_key": "max_stress"},
                "enabled": True
            })
            result["spec_changes"].append("Added objective: minimize stress")

        # WFE/wavefront
        if any(w in answer_lower for w in ["wfe", "wavefront", "optical", "zernike"]):
            extractors.append({
                "id": "ext_wfe",
                "name": "Wavefront Error",
                "type": "zernike",
                "config": {"terms": [4, 5, 6, 7, 8, 9, 10, 11]},
                "enabled": True
            })
            objectives.append({
                "id": "obj_wfe",
                "name": "WFE RMS",
                "direction": "minimize",
                "source": {"extractor_id": "ext_wfe", "output_key": "wfe_rms"},
                "enabled": True
            })
            result["spec_changes"].append("Added objective: minimize wavefront error")

        self.state.spec["extractors"].extend(extractors)
        self.state.spec["objectives"].extend(objectives)
        self.state.collected["objectives"] = [o["name"] for o in objectives]

    def _process_design_vars(self, answer: str, result: Dict):
        """Extract design variables"""
        answer_lower = answer.lower()
        dvs = []

        # Parse patterns like "thickness 2-10mm" or "thickness from 2 to 10"
        # Pattern: name [range]
        patterns = [
            r'(\w+)\s+(\d+(?:\.\d+)?)\s*[-–to]+\s*(\d+(?:\.\d+)?)\s*(mm|deg|°)?',
            r'(\w+)\s+\[(\d+(?:\.\d+)?),?\s*(\d+(?:\.\d+)?)\]',
        ]

        for pattern in patterns:
            matches = re.findall(pattern, answer_lower)
            for match in matches:
                name = match[0]
                min_val = float(match[1])
                max_val = float(match[2])
                unit = match[3] if len(match) > 3 else ""

                dv = {
                    "id": f"dv_{name}",
                    "name": name.replace("_", " ").title(),
                    "expression_name": name,
                    "type": "continuous",
                    "bounds": {"min": min_val, "max": max_val},
                    "baseline": (min_val + max_val) / 2,
                    "enabled": True
                }
                if unit:
                    dv["units"] = unit
                dvs.append(dv)
                result["spec_changes"].append(f"Added design variable: {name} [{min_val}, {max_val}]")

        # If no pattern matched, try to use suggestions
        if not dvs and self.state.model_expressions:
            # User might have said "yes" or named expressions without ranges
            for expr in self.state.model_expressions[:3]:  # Use top 3
                if expr["name"].lower() in answer_lower or "yes" in answer_lower or "all" in answer_lower:
                    val = expr["value"]
                    dv = {
                        "id": f"dv_{expr['name']}",
                        "name": expr["name"],
                        "expression_name": expr["name"],
                        "type": "continuous",
                        "bounds": {"min": val * 0.5, "max": val * 1.5},
                        "baseline": val,
                        "enabled": True
                    }
                    dvs.append(dv)
                    result["spec_changes"].append(f"Added design variable: {expr['name']} [{val*0.5:.2f}, {val*1.5:.2f}]")

        self.state.spec["design_variables"].extend(dvs)
        self.state.collected["design_vars"] = [dv["name"] for dv in dvs]

    def _process_constraints(self, answer: str, result: Dict):
        """Extract constraints"""
        answer_lower = answer.lower()

        if answer_lower in ["none", "no", "skip", "n/a"]:
            return

        constraints = []

        # Pattern: "stress < 200" or "stress <= 200 MPa"
        stress_match = re.search(r'stress\s*([<>=≤≥]+)\s*(\d+(?:\.\d+)?)', answer_lower)
        if stress_match:
            op = stress_match.group(1).replace("≤", "<=").replace("≥", ">=")
            val = float(stress_match.group(2))

            # Add extractor if not exists
            if not any(e["type"] == "stress" for e in self.state.spec["extractors"]):
                self.state.spec["extractors"].append({
                    "id": "ext_stress_con",
                    "name": "Stress for Constraint",
                    "type": "stress",
                    "config": {"stress_type": "von_mises"},
                    "enabled": True
                })

            constraints.append({
                "id": "con_stress",
                "name": "Max Stress",
                "operator": op if op in ["<=", ">=", "<", ">", "=="] else "<=",
                "threshold": val,
                "source": {"extractor_id": "ext_stress_con", "output_key": "max_stress"},
                "enabled": True
            })
            result["spec_changes"].append(f"Added constraint: stress {op} {val}")

        # Similar patterns for frequency, mass, displacement...

        self.state.spec["constraints"].extend(constraints)

    def _process_method(self, answer: str, result: Dict):
        """Confirm or change optimization method"""
        answer_lower = answer.lower()

        if any(w in answer_lower for w in ["yes", "ok", "sure", "good", "proceed"]):
            # Keep recommended method
            pass
        elif "nsga" in answer_lower:
            self.state.spec["optimization"]["algorithm"]["type"] = "NSGA-II"
        elif "tpe" in answer_lower:
            self.state.spec["optimization"]["algorithm"]["type"] = "TPE"
        elif "cma" in answer_lower:
            self.state.spec["optimization"]["algorithm"]["type"] = "CMA-ES"

        result["spec_changes"].append(f"Set method: {self.state.spec['optimization']['algorithm']['type']}")

    def _process_review(self, answer: str, result: Dict):
        """Confirm or revise"""
        answer_lower = answer.lower()

        if any(w in answer_lower for w in ["yes", "ok", "create", "proceed", "looks good"]):
            result["complete"] = True
        else:
            # User wants changes - stay in review or go back
            result["error"] = "What would you like to change?"

    def _recommend_method(self) -> tuple:
        """Recommend optimization method based on problem"""
        n_obj = len(self.state.spec["objectives"])
        n_dv = len(self.state.spec["design_variables"])

        if n_obj > 1:
            return "NSGA-II", f"You have {n_obj} objectives, which requires multi-objective optimization"
        elif n_dv > 10:
            return "CMA-ES", f"With {n_dv} design variables, CMA-ES handles high dimensions well"
        else:
            return "TPE", "TPE (Bayesian optimization) is efficient for single-objective problems"

    def _format_dv_suggestions(self) -> str:
        """Format design variable suggestions"""
        if self.state.model_expressions:
            lines = ["I found these expressions in your model:"]
            for expr in self.state.model_expressions[:5]:
                lines.append(f"  - {expr['name']} = {expr['value']}")
            lines.append("\nWhich ones should vary? (or describe your own)")
            return "\n".join(lines)
        return "Describe the parameters and their ranges (e.g., 'thickness 2-10mm, width 5-20mm')"

    def _format_summary(self) -> str:
        """Format configuration summary"""
        spec = self.state.spec
        lines = []

        lines.append(f"**Model**: {spec['model']['sim']['path'] or 'Not set'}")

        lines.append(f"\n**Design Variables ({len(spec['design_variables'])}):**")
        for dv in spec["design_variables"]:
            b = dv["bounds"]
            lines.append(f"  - {dv['name']}: [{b['min']}, {b['max']}]")

        lines.append(f"\n**Objectives ({len(spec['objectives'])}):**")
        for obj in spec["objectives"]:
            lines.append(f"  - {obj['direction']} {obj['name']}")

        lines.append(f"\n**Constraints ({len(spec['constraints'])}):**")
        if spec["constraints"]:
            for con in spec["constraints"]:
                lines.append(f"  - {con['name']} {con['operator']} {con['threshold']}")
        else:
            lines.append("  - None")

        lines.append(f"\n**Method**: {spec['optimization']['algorithm']['type']}")

        return "\n".join(lines)

    def get_spec(self) -> Dict[str, Any]:
        """Get the built spec"""
        return self.state.spec

    def set_model_expressions(self, expressions: List[Dict]):
        """Set model expressions for DV suggestions"""
        self.state.model_expressions = expressions
```

---

#### Task 4.2: Add Interview Tools to Claude Agent
**File**: `backend/api/services/claude_agent.py`

**Add tools**:
```python
{
    "name": "start_interview",
    "description": "Start a guided interview to create a new optimization study. Use this when the user wants help setting up an optimization but hasn't provided full details.",
    "input_schema": {
        "type": "object",
        "properties": {},
        "required": []
    }
},
{
    "name": "interview_answer",
    "description": "Process the user's answer during an interview. Extract relevant information and advance the interview.",
    "input_schema": {
        "type": "object",
        "properties": {
            "answer": {
                "type": "string",
                "description": "The user's answer to the current interview question"
            }
        },
        "required": ["answer"]
    }
},
```

**Add state and implementations**:
```python
def __init__(self, study_id: Optional[str] = None):
    # ... existing
    self.interview: Optional[InterviewEngine] = None

def _tool_start_interview(self, params: Dict[str, Any]) -> str:
    """Start guided study creation"""
    from api.services.interview_engine import InterviewEngine
    self.interview = InterviewEngine()
    question = self.interview.get_current_question()
    return f"Let's set up your optimization step by step.\n\n{question}"

def _tool_interview_answer(self, params: Dict[str, Any]) -> str:
    """Process interview answer"""
    if not self.interview:
        return "No interview in progress. Use start_interview first."

    result = self.interview.process_answer(params["answer"])

    response_parts = []

    # Show what was extracted
    if result["spec_changes"]:
        response_parts.append("**Updated:**")
        for change in result["spec_changes"]:
            response_parts.append(f"  ✓ {change}")

    if result["error"]:
        response_parts.append(f"\n{result['error']}")
    elif result["complete"]:
        # Create the study
        spec = self.interview.get_spec()
        # ... create study directory and save spec
        response_parts.append("\n✓ **Interview complete!** Creating your study...")
        self.canvas_state = spec
    elif result["question"]:
        response_parts.append(f"\n{result['question']}")

    return "\n".join(response_parts)
```

---

### Phase 5: UI Polish (P2/P3)

#### Task 5.1: Tool Call Indicators
**File**: `frontend/src/components/chat/ToolIndicator.tsx` (NEW)

```typescript
import { Loader2, Check, Variable, Cpu, Target, Lock, FolderPlus, Search, Wrench } from 'lucide-react';

interface ToolIndicatorProps {
  tool: string;
  status: 'running' | 'complete';
  result?: string;
}

const TOOL_ICONS: Record<string, React.ComponentType<any>> = {
  add_design_variable: Variable,
  add_extractor: Cpu,
  add_objective: Target,
  add_constraint: Lock,
  create_study: FolderPlus,
  introspect_model: Search,
};

const TOOL_LABELS: Record<string, string> = {
  add_design_variable: 'Adding design variable',
  add_extractor: 'Adding extractor',
  add_objective: 'Adding objective',
  add_constraint: 'Adding constraint',
  create_study: 'Creating study',
  update_spec_field: 'Updating configuration',
  remove_node: 'Removing node',
};

export function ToolIndicator({ tool, status, result }: ToolIndicatorProps) {
  const Icon = TOOL_ICONS[tool] || Wrench;
  const label = TOOL_LABELS[tool] || tool;

  return (
    <div className={`flex items-center gap-2 px-3 py-2 rounded-lg text-sm ${
      status === 'running'
        ? 'bg-amber-500/10 text-amber-400 border border-amber-500/20'
        : 'bg-green-500/10 text-green-400 border border-green-500/20'
    }`}>
      {status === 'running' ? (
        <Loader2 className="w-4 h-4 animate-spin" />
      ) : (
        <Check className="w-4 h-4" />
      )}
      <Icon className="w-4 h-4" />
      <span className="font-medium">{label}</span>
      {status === 'complete' && result && (
        <span className="text-xs opacity-75 ml-2">{result}</span>
      )}
    </div>
  );
}
```

---

#### Task 5.2: Display Tool Calls in Chat
**File**: `frontend/src/components/chat/ChatMessage.tsx`

**Add tool call rendering**:
```typescript
import { ToolIndicator } from './ToolIndicator';

// In message rendering:
{message.toolCalls?.map((tc, idx) => (
  <ToolIndicator
    key={idx}
    tool={tc.tool}
    status={tc.status}
    result={tc.result}
  />
))}
```

---

## File Summary

### Files to Modify

| File | Changes |
|------|---------|
| `backend/api/services/claude_agent.py` | Add canvas context, streaming, create_study, interview tools |
| `backend/api/routes/claude.py` | Send full spec, handle canvas_edit, use streaming |
| `frontend/src/hooks/useChat.ts` | Add notifyCanvasEdit, handle streaming, onSpecUpdated |
| `frontend/src/pages/CanvasView.tsx` | Wire onSpecUpdated, pass notifyCanvasEdit |
| `frontend/src/components/canvas/SpecRenderer.tsx` | Call notifyCanvasEdit on user edits |

### Files to Create

| File | Purpose |
|------|---------|
| `backend/api/services/interview_engine.py` | Guided study creation |
| `frontend/src/components/chat/ToolIndicator.tsx` | Tool call UI |

---

## Testing Checklist

### Phase 1 Tests
- [ ] Claude mentions current canvas state in response
- [ ] Canvas updates without HTTP reload when Claude modifies spec
- [ ] User edits canvas → ask Claude about it → Claude knows the change

### Phase 2 Tests
- [ ] Text streams in as Claude types
- [ ] Tool calls show "running" then "complete" status

### Phase 3 Tests
- [ ] "Create a new study called bracket_v2" creates directory + spec
- [ ] Canvas shows new empty study

### Phase 4 Tests
- [ ] "Help me set up an optimization" starts interview
- [ ] Each answer updates canvas incrementally
- [ ] Interview completes with valid spec

---

## Ralph Loop Execution Notes

1. **Start with Phase 1** - It's the foundation for everything else
2. **Test each task individually** before moving to next
3. **The key insight**: `atomizer_spec.json` is the single source of truth
4. **Don't break existing functionality** - Power mode should still work

### Quick Wins (Do First)
1. Task 1.1 - Add canvas state to Claude context (easy, high value)
2. Task 1.2 - Send full spec in response (easy, removes HTTP reload)
3. Task 1.5 - Frontend use spec directly (easy, faster updates)

### Recommended Order
```
1.1 → 1.2 → 1.5 → 1.6 → 1.3 → 1.4 → 2.1 → 2.2 → 2.3 → 3.1 → 4.1 → 4.2 → 5.1 → 5.2
```

---

## Success Criteria

The project is complete when:

1. **Bi-directional sync works**: Claude modifies → Canvas updates instantly. User edits → Claude sees in next message.

2. **Streaming works**: Text appears as Claude types, tool calls show in real-time.

3. **Study creation works**: User can say "create bracket optimization with mass objective" and see it built on canvas.

4. **Interview works**: User can say "help me set up an optimization" and be guided through the process.

5. **No HTTP reloads**: Canvas updates purely through WebSocket.