htmlgraph 0.9.3__py3-none-any.whl → 0.27.5__py3-none-any.whl

htmlgraph/cigs/cost.py

@@ -0,0 +1,475 @@
+"""
+CostCalculator for CIGS (Computational Imperative Guidance System).
+
+Provides token cost prediction and actual cost tracking for tool operations.
+Implements cost estimation heuristics based on tool type and operation complexity.
+
+Design Reference:
+- CIGS Design Doc: .htmlgraph/spikes/computational-imperative-guidance-system-design.md
+- Part 5.2: Cost Efficiency Score
+- Part 5.3: Cost estimation rules
+"""
+
+from .models import (
+    CostMetrics,
+    OperationClassification,
+    TokenCost,
+)
+
+
+class CostCalculator:
+    """Calculate token costs for tool operations and delegations."""
+
+    # Cost estimation heuristics (tokens per operation)
+    # Based on typical token consumption patterns
+    COST_ESTIMATES = {
+        "Read": 5000,  # Per file read
+        "Grep": 3000,  # Per search result batch
+        "Glob": 2000,  # Per glob pattern
+        "Edit": 4000,  # Per edit operation
+        "Write": 4000,  # Per write operation
+        "NotebookEdit": 4500,  # Notebook cells are complex
+        "Delete": 2000,  # File deletion
+        "Bash": {  # Variable based on command type
+            "default": 2000,
+            "git": 1500,
+            "pytest": 5000,
+            "npm": 5000,
+        },
+        "Task": 500,  # Orchestrator Task call (minimal context)
+        "AskUserQuestion": 1000,  # User interaction
+        "TodoWrite": 500,  # Tracking operations
+    }
+
+    # Subagent delegation costs (optimal path)
+    SUBAGENT_COSTS = {
+        "spawn_gemini": 500,  # Explorer subagent cost
+        "spawn_codex": 800,  # Coder subagent cost
+        "spawn_copilot": 600,  # Git subagent cost
+        "Task": 500,  # Orchestrator Task
+    }
+
+    def __init__(self) -> None:
+        """Initialize CostCalculator."""
+        self.tool_history: list[str] = []  # Track recent tool usage for patterns
+
+    def predict_cost(self, tool: str, params: dict) -> int:
+        """
+        Predict token cost for direct tool execution.
+
+        Args:
+            tool: Tool name (Read, Grep, Bash, etc.)
+            params: Tool parameters
+
+        Returns:
+            Estimated token cost
+        """
+        if tool not in self.COST_ESTIMATES:
+            return 2000  # Default estimate for unknown tools
+
+        estimate = self.COST_ESTIMATES[tool]
+
+        # Handle variable estimates
+        if isinstance(estimate, dict):
+            return self._estimate_variable_cost(tool, params, estimate)
+
+        # Apply modifiers based on operation complexity
+        assert isinstance(estimate, int)
+        return self._apply_complexity_modifiers(tool, params, estimate)
+
+    def _estimate_variable_cost(
+        self, tool: str, params: dict, estimates: dict[str, int]
+    ) -> int:
+        """Estimate cost for tools with variable pricing."""
+        if tool == "Bash":
+            command = params.get("command", "")
+
+            # Git operations
+            if any(cmd in command for cmd in ["git add", "git commit", "git push"]):
+                return int(estimates.get("git", estimates.get("default", 2000)))
+
+            # Test operations
+            if any(cmd in command for cmd in ["pytest", "uv run pytest"]):
+                return int(estimates.get("pytest", estimates.get("default", 5000)))
+
+            if "npm test" in command or "yarn test" in command:
+                return int(estimates.get("npm", estimates.get("default", 5000)))
+
+            # Default bash
+            return int(estimates.get("default", 2000))
+
+        return int(estimates.get("default", 2000))
+
+    def _apply_complexity_modifiers(
+        self, tool: str, params: dict, base_cost: int | dict[str, int]
+    ) -> int:
+        """Apply complexity modifiers to base cost estimate."""
+        # Handle dict base_cost (shouldn't happen but for safety)
+        if isinstance(base_cost, dict):
+            base_cost = base_cost.get("default", 2000)
+
+        modified_cost: float = float(base_cost)
+
+        if tool == "Read":
+            # Multiple files increase cost
+            if isinstance(params.get("file_path"), list):
+                modified_cost = base_cost * len(params["file_path"])
+            # Large offset/limits indicate large reads
+            elif params.get("limit", 2000) > 5000:
+                modified_cost = base_cost * 2
+
+        elif tool == "Grep":
+            # Complex regex patterns increase cost
+            pattern = params.get("pattern", "")
+            if len(pattern) > 100:  # Complex pattern
+                modified_cost = base_cost * 1.5
+            # Multiline matching is more expensive
+            if params.get("multiline", False):
+                modified_cost = base_cost * 1.3
+
+        elif tool in ["Edit", "Write"]:
+            # Multiple edits or large content
+            if isinstance(params.get("file_path"), list):
+                modified_cost = base_cost * len(params["file_path"])
+            # Large content increases cost
+            content = params.get("content", params.get("new_string", ""))
+            if len(content) > 10000:
+                modified_cost = base_cost * 1.5
+
+        return int(modified_cost)
+
+    def optimal_cost(self, classification: OperationClassification) -> int:
+        """
+        Calculate optimal cost with proper delegation.
+
+        Args:
+            classification: OperationClassification with tool and category info
+
+        Returns:
+            Estimated token cost with optimal delegation strategy
+        """
+        tool = classification.tool
+
+        # Orchestrator tools already optimal
+        if tool in ["Task", "AskUserQuestion", "TodoWrite"]:
+            estimate = self.COST_ESTIMATES.get(tool, 500)
+            return int(estimate) if isinstance(estimate, int) else 500
+
+        # Map to subagent based on category
+        if classification.category == "exploration":
+            return int(self.SUBAGENT_COSTS["spawn_gemini"])
+        elif classification.category == "implementation":
+            return int(self.SUBAGENT_COSTS["spawn_codex"])
+        elif classification.category == "git":
+            return int(self.SUBAGENT_COSTS["spawn_copilot"])
+        elif classification.category == "testing":
+            return int(self.SUBAGENT_COSTS["Task"])
+
+        # Default to Task for unknown categories
+        return int(self.SUBAGENT_COSTS["Task"])
+
+    def calculate_actual_cost(self, tool: str, result: dict) -> TokenCost:
+        """
+        Calculate actual cost after tool execution.
+
+        Args:
+            tool: Tool that was executed
+            result: Result dictionary from tool execution
+
+        Returns:
+            TokenCost with actual metrics
+        """
+        # Get predicted cost for this tool
+        predicted_tokens = self._extract_predicted_cost(tool, result)
+
+        # Extract actual cost if available in result
+        actual_tokens = self._extract_actual_cost(tool, result)
+
+        # If no actual cost in result, use predicted
+        if actual_tokens is None:
+            actual_tokens = predicted_tokens
+
+        # Determine subagent cost based on tool type
+        subagent_tokens = self._get_subagent_cost(tool)
+
+        # Calculate orchestrator overhead
+        orchestrator_tokens = self._estimate_orchestrator_overhead(tool, result)
+
+        # Calculate savings
+        estimated_savings = max(
+            0, actual_tokens - subagent_tokens - orchestrator_tokens
+        )
+
+        return TokenCost(
+            total_tokens=actual_tokens,
+            orchestrator_tokens=orchestrator_tokens,
+            subagent_tokens=subagent_tokens,
+            estimated_savings=estimated_savings,
+        )
+
+    def _extract_predicted_cost(self, tool: str, result: dict) -> int:
+        """Extract or estimate predicted cost from result."""
+        # Check if result contains cost metadata
+        if "predicted_cost" in result:
+            return int(result["predicted_cost"])
+
+        if "metadata" in result and "predicted_cost" in result["metadata"]:
+            return int(result["metadata"]["predicted_cost"])
+
+        # Fall back to default estimate
+        estimate = self.COST_ESTIMATES.get(tool, 2000)
+        return int(estimate) if isinstance(estimate, int) else 2000
+
+    def _extract_actual_cost(self, tool: str, result: dict) -> int | None:
+        """Extract actual cost from execution result if available."""
+        # Check standard cost fields
+        if "actual_cost" in result:
+            return int(result["actual_cost"])
+
+        if "cost" in result:
+            return int(result["cost"])
+
+        if "metadata" in result and "cost" in result["metadata"]:
+            return int(result["metadata"]["cost"])
+
+        if "tokens" in result:
+            return int(result["tokens"])
+
+        # Try to estimate from output size for Read operations
+        if tool == "Read" and "output" in result:
+            output = result["output"]
+            if isinstance(output, str):
+                # Rough estimate: ~4 tokens per line
+                lines = len(output.split("\n"))
+                return int(lines * 4)
+
+        return None
+
+    def _get_subagent_cost(self, tool: str) -> int:
+        """Get cost if this operation were delegated to subagent."""
+        if tool in ["Task", "AskUserQuestion", "TodoWrite"]:
+            return 0  # Already delegated
+
+        if tool in ["Read", "Grep", "Glob"]:
+            return self.SUBAGENT_COSTS["spawn_gemini"]
+
+        if tool in ["Edit", "Write", "NotebookEdit", "Delete"]:
+            return self.SUBAGENT_COSTS["spawn_codex"]
+
+        if tool == "Bash":
+            # Might be git or test - estimate higher
+            return self.SUBAGENT_COSTS["spawn_copilot"]  # or Task
+
+        # Default
+        return self.SUBAGENT_COSTS["Task"]
+
+    def _estimate_orchestrator_overhead(self, tool: str, result: dict) -> int:
+        """Estimate orchestrator context overhead."""
+        # Orchestrator overhead is minimal for delegated operations
+        if tool in ["Task", "AskUserQuestion"]:
+            return 200  # Small overhead for delegation call
+
+        # Direct execution contributes full cost to orchestrator context
+        if tool == "Read":
+            # Context cost is proportional to file size
+            if "output" in result:
+                output = result["output"]
+                if isinstance(output, str):
+                    # ~4 tokens per line
+                    return int(len(output.split("\n")) * 4)
+            return 5000
+
+        # Other tools
+        return 1000  # Placeholder for other tools
+
+    def calculate_waste(self, actual_cost: int, optimal_cost: int) -> dict:
+        """
+        Calculate waste metrics comparing actual vs optimal cost.
+
+        Args:
+            actual_cost: Actual tokens consumed
+            optimal_cost: Tokens with optimal delegation
+
+        Returns:
+            Dictionary with waste metrics
+        """
+        waste_tokens = max(0, actual_cost - optimal_cost)
+
+        if actual_cost == 0:
+            waste_percentage = 0.0
+            efficiency_score = 100.0
+        else:
+            waste_percentage = (waste_tokens / actual_cost) * 100
+            efficiency_score = (optimal_cost / actual_cost) * 100
+
+        return {
+            "waste_tokens": waste_tokens,
+            "waste_percentage": waste_percentage,
+            "efficiency_score": efficiency_score,
+        }
+
+    def aggregate_session_costs(
+        self,
+        operations: list[tuple[str, dict, dict]],
+        violations_count: int = 0,
+    ) -> CostMetrics:
+        """
+        Aggregate cost metrics for a session.
+
+        Args:
+            operations: List of (tool, params, result) tuples
+            violations_count: Number of violations in session
+
+        Returns:
+            CostMetrics with aggregated session costs
+        """
+        total_tokens = 0
+        optimal_tokens = 0
+        orchestrator_tokens = 0
+        subagent_tokens = 0
+
+        for tool, params, result in operations:
+            # Predict cost
+            predicted = self.predict_cost(tool, params)
+            total_tokens += predicted
+
+            # Calculate actual if available
+            cost_record = self.calculate_actual_cost(tool, result)
+            total_tokens = max(total_tokens, cost_record.total_tokens)
+
+            # Accumulate subagent costs
+            subagent_tokens += cost_record.subagent_tokens
+            orchestrator_tokens += cost_record.orchestrator_tokens
+
+            # Accumulate optimal costs
+            # For delegated operations, optimal = subagent cost
+            # For direct operations that should be delegated, optimal = subagent cost
+            if tool in ["Task", "AskUserQuestion"]:
+                estimate = self.COST_ESTIMATES.get(tool, 500)
+                optimal_tokens += int(estimate) if isinstance(estimate, int) else 500
+            else:
+                optimal_tokens += self._get_subagent_cost(tool)
+
+        waste_tokens = max(0, total_tokens - optimal_tokens)
+
+        metrics = CostMetrics(
+            total_tokens=total_tokens,
+            optimal_tokens=optimal_tokens,
+            orchestrator_tokens=orchestrator_tokens,
+            subagent_tokens=subagent_tokens,
+            waste_tokens=waste_tokens,
+        )
+
+        return metrics
+
+    def classify_operation(
+        self,
+        tool: str,
+        params: dict,
+        is_exploration_sequence: bool = False,
+        tool_history: list[str] | None = None,
+    ) -> OperationClassification:
+        """
+        Classify a tool operation for cost and delegation analysis.
+
+        Args:
+            tool: Tool name
+            params: Tool parameters
+            is_exploration_sequence: Whether this is part of exploration sequence
+            tool_history: Recent tool usage history
+
+        Returns:
+            OperationClassification with cost and delegation info
+        """
+        # Categorize tool
+        category = self._categorize_tool(tool)
+
+        # Determine if delegation is recommended
+        should_delegate = self._should_delegate(tool, is_exploration_sequence)
+
+        # Get delegation suggestion
+        suggestion = self._get_delegation_suggestion(tool, category)
+
+        # Predict costs
+        predicted_cost = self.predict_cost(tool, params)
+
+        # Calculate optimal cost for this category
+        dummy_classification = OperationClassification(
+            tool=tool,
+            category="",  # Will be set below
+            should_delegate=should_delegate,
+            reason="",
+            predicted_cost=predicted_cost,
+            optimal_cost=0,
+            is_exploration_sequence=is_exploration_sequence,
+            suggested_delegation=suggestion,
+        )
+
+        optimal_cost = self.optimal_cost(dummy_classification)
+
+        # Calculate waste percentage
+        waste_tokens = max(0, predicted_cost - optimal_cost)
+        waste_pct = (waste_tokens / predicted_cost * 100) if predicted_cost > 0 else 0.0
+
+        return OperationClassification(
+            tool=tool,
+            category=category,
+            should_delegate=should_delegate,
+            reason=self._get_delegation_reason(tool, category, is_exploration_sequence),
+            predicted_cost=predicted_cost,
+            optimal_cost=optimal_cost,
+            is_exploration_sequence=is_exploration_sequence,
+            suggested_delegation=suggestion,
+            waste_percentage=waste_pct,
+        )
+
+    def _categorize_tool(self, tool: str) -> str:
+        """Categorize tool into operational type."""
+        if tool in ["Read", "Grep", "Glob"]:
+            return "exploration"
+        elif tool in ["Edit", "Write", "NotebookEdit", "Delete"]:
+            return "implementation"
+        elif tool in ["Task", "AskUserQuestion", "TodoWrite"]:
+            return "orchestration"
+        elif tool == "Bash":
+            return "execution"
+        else:
+            return "unknown"
+
+    def _should_delegate(self, tool: str, is_sequence: bool) -> bool:
+        """Determine if operation should be delegated."""
+        # Orchestrator tools are already delegated
+        if tool in ["Task", "AskUserQuestion", "TodoWrite"]:
+            return False
+
+        # Exploration sequences should be delegated
+        if is_sequence and tool in ["Read", "Grep", "Glob"]:
+            return True
+
+        # Single direct operations are allowed but not recommended
+        # CIGS messaging will recommend delegation
+        return False
+
+    def _get_delegation_suggestion(self, tool: str, category: str) -> str:
+        """Get suggested delegation for this tool."""
+        suggestions = {
+            "exploration": "spawn_gemini(prompt='Search and analyze codebase')",
+            "implementation": "spawn_codex(prompt='Implement with full testing')",
+            "execution": "Task(prompt='Execute and report results')",
+            "testing": "Task(prompt='Run tests and report')",
+        }
+        return suggestions.get(category, "Task(prompt='Delegate this operation')")
+
+    def _get_delegation_reason(
+        self, tool: str, category: str, is_sequence: bool
+    ) -> str:
+        """Get reason for delegation recommendation."""
+        if is_sequence:
+            return f"Multiple {category} operations detected (research work should be delegated)"
+
+        reasons = {
+            "exploration": "Exploration operations have unpredictable scope",
+            "implementation": "Implementation requires iteration and testing",
+            "execution": "Consider delegating execution for better isolation",
+        }
+        return reasons.get(category, "Delegation preserves your strategic context")