attune-ai 2.0.0__py3-none-any.whl

attune/workflows/document_gen/workflow.py

@@ -0,0 +1,1426 @@
+"""Document Generation Workflow.
+
+Main workflow orchestration for documentation generation.
+
+Copyright 2025 Smart-AI-Memory
+Licensed under Fair Source License 0.9
+"""
+
+import logging
+from datetime import datetime
+from pathlib import Path
+from typing import Any
+
+from attune.config import _validate_file_path
+
+from ..base import BaseWorkflow, ModelTier
+from .config import DOC_GEN_STEPS, TOKEN_COSTS
+from .report_formatter import format_doc_gen_report
+
+logger = logging.getLogger(__name__)
+
+
+class DocumentGenerationWorkflow(BaseWorkflow):
+    """Multi-tier document generation workflow.
+
+    Uses cheap models for outlining, capable models for content
+    generation, and premium models for final polish and consistency
+    review.
+
+    Usage:
+        workflow = DocumentGenerationWorkflow()
+        result = await workflow.execute(
+            source_code="...",
+            doc_type="api_reference",
+            audience="developers"
+        )
+    """
+
+    name = "doc-gen"
+    description = "Cost-optimized documentation generation pipeline"
+    stages = ["outline", "write", "polish"]
+    tier_map = {
+        "outline": ModelTier.CHEAP,
+        "write": ModelTier.CAPABLE,
+        "polish": ModelTier.PREMIUM,
+    }
+
+    def __init__(
+        self,
+        skip_polish_threshold: int = 1000,
+        max_sections: int = 10,
+        max_write_tokens: int | None = None,  # Auto-scaled if None
+        section_focus: list[str] | None = None,
+        chunked_generation: bool = True,
+        sections_per_chunk: int = 3,
+        max_cost: float = 5.0,  # Cost guardrail in USD
+        cost_warning_threshold: float = 0.8,  # Warn at 80% of max_cost
+        graceful_degradation: bool = True,  # Return partial results on error
+        export_path: str | Path | None = None,  # Export docs to file (e.g., "docs/generated")
+        max_display_chars: int = 45000,  # Max chars before chunking output
+        enable_auth_strategy: bool = True,  # Enable intelligent auth routing
+        **kwargs: Any,
+    ):
+        """Initialize workflow with enterprise-safe defaults.
+
+        Args:
+            skip_polish_threshold: Skip premium polish for docs under this
+                token count (they're already good enough).
+            max_sections: Maximum number of sections to generate.
+            max_write_tokens: Maximum tokens for content generation.
+                If None, auto-scales based on section count (recommended).
+            section_focus: Optional list of specific sections to generate
+                (e.g., ["Testing Guide", "API Reference"]).
+            chunked_generation: If True, generates large docs in chunks to avoid
+                truncation (default True).
+            sections_per_chunk: Number of sections to generate per chunk (default 3).
+            max_cost: Maximum cost in USD before stopping (default $5).
+                Set to 0 to disable cost limits.
+            cost_warning_threshold: Percentage of max_cost to trigger warning (default 0.8).
+            graceful_degradation: If True, return partial results on errors
+                instead of failing completely (default True).
+            export_path: Optional directory to export generated docs (e.g., "docs/generated").
+                If provided, documentation will be saved to a file automatically.
+            max_display_chars: Maximum characters before splitting output into chunks
+                for display (default 45000). Helps avoid terminal/UI truncation.
+            enable_auth_strategy: If True, use intelligent subscription vs API routing
+                based on module size (default True).
+
+        """
+        super().__init__(**kwargs)
+        self.skip_polish_threshold = skip_polish_threshold
+        self.max_sections = max_sections
+        self._user_max_write_tokens = max_write_tokens  # Store user preference
+        self.max_write_tokens = max_write_tokens or 16000  # Will be auto-scaled
+        self.section_focus = section_focus
+        self.chunked_generation = chunked_generation
+        self.sections_per_chunk = sections_per_chunk
+        self.max_cost = max_cost
+        self.cost_warning_threshold = cost_warning_threshold
+        self.graceful_degradation = graceful_degradation
+        self.export_path = Path(export_path) if export_path else None
+        self.max_display_chars = max_display_chars
+        self.enable_auth_strategy = enable_auth_strategy
+        self._total_content_tokens: int = 0
+        self._accumulated_cost: float = 0.0
+        self._cost_warning_issued: bool = False
+        self._partial_results: dict = {}
+        self._auth_mode_used: str | None = None  # Track which auth was recommended
+
+    def _estimate_cost(self, tier: ModelTier, input_tokens: int, output_tokens: int) -> float:
+        """Estimate cost for a given tier and token counts."""
+        costs = TOKEN_COSTS.get(tier, TOKEN_COSTS[ModelTier.CAPABLE])
+        input_cost = (input_tokens / 1000) * costs["input"]
+        output_cost = (output_tokens / 1000) * costs["output"]
+        return input_cost + output_cost
+
+    def _track_cost(
+        self,
+        tier: ModelTier,
+        input_tokens: int,
+        output_tokens: int,
+    ) -> tuple[float, bool]:
+        """Track accumulated cost and check against limits.
+
+        Returns:
+            Tuple of (cost_for_this_call, should_stop)
+
+        """
+        cost = self._estimate_cost(tier, input_tokens, output_tokens)
+        self._accumulated_cost += cost
+
+        # Check warning threshold
+        if (
+            self.max_cost > 0
+            and not self._cost_warning_issued
+            and self._accumulated_cost >= self.max_cost * self.cost_warning_threshold
+        ):
+            self._cost_warning_issued = True
+            logger.warning(
+                f"Doc-gen cost approaching limit: ${self._accumulated_cost:.2f} "
+                f"of ${self.max_cost:.2f} ({self.cost_warning_threshold * 100:.0f}% threshold)",
+            )
+
+        # Check if we should stop
+        should_stop = self.max_cost > 0 and self._accumulated_cost >= self.max_cost
+        if should_stop:
+            logger.warning(
+                f"Doc-gen cost limit reached: ${self._accumulated_cost:.2f} >= ${self.max_cost:.2f}",
+            )
+
+        return cost, should_stop
+
+    def _auto_scale_tokens(self, section_count: int) -> int:
+        """Auto-scale max_write_tokens based on section count.
+
+        Enterprise projects may have 20+ sections requiring more tokens.
+        """
+        if self._user_max_write_tokens is not None:
+            return self._user_max_write_tokens  # User override
+
+        # Base: 2000 tokens per section, minimum 16000, maximum 64000
+        scaled = max(16000, min(64000, section_count * 2000))
+        logger.info(f"Auto-scaled max_write_tokens to {scaled} for {section_count} sections")
+        return scaled
+
+    def _export_document(
+        self,
+        document: str,
+        doc_type: str,
+        report: str | None = None,
+    ) -> tuple[Path | None, Path | None]:
+        """Export generated documentation to file.
+
+        Args:
+            document: The generated documentation content
+            doc_type: Document type for naming
+            report: Optional report to save alongside document
+
+        Returns:
+            Tuple of (doc_path, report_path) or (None, None) if export disabled
+
+        """
+        if not self.export_path:
+            return None, None
+
+        # Create export directory
+        self.export_path.mkdir(parents=True, exist_ok=True)
+
+        # Generate filename with timestamp
+        timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+        safe_doc_type = doc_type.replace(" ", "_").replace("/", "-").lower()
+        doc_filename = f"{safe_doc_type}_{timestamp}.md"
+        report_filename = f"{safe_doc_type}_{timestamp}_report.txt"
+
+        doc_path = self.export_path / doc_filename
+        report_path = self.export_path / report_filename if report else None
+
+        # Write document
+        try:
+            validated_doc_path = _validate_file_path(str(doc_path))
+            validated_doc_path.write_text(document, encoding="utf-8")
+            logger.info(f"Documentation exported to: {validated_doc_path}")
+
+            # Write report if provided
+            if report and report_path:
+                validated_report_path = _validate_file_path(str(report_path))
+                validated_report_path.write_text(report, encoding="utf-8")
+                logger.info(f"Report exported to: {validated_report_path}")
+
+            return validated_doc_path, validated_report_path if report else None
+        except (OSError, ValueError) as e:
+            logger.error(f"Failed to export documentation: {e}")
+            return None, None
+
+    def _chunk_output_for_display(self, content: str, chunk_prefix: str = "PART") -> list[str]:
+        """Split large output into displayable chunks.
+
+        Args:
+            content: The content to chunk
+            chunk_prefix: Prefix for chunk headers
+
+        Returns:
+            List of content chunks, each under max_display_chars
+
+        """
+        if len(content) <= self.max_display_chars:
+            return [content]
+
+        chunks = []
+        # Try to split on section boundaries (## headers)
+        import re
+
+        sections = re.split(r"(?=^## )", content, flags=re.MULTILINE)
+
+        current_chunk = ""
+        chunk_num = 1
+
+        for section in sections:
+            # If adding this section would exceed limit, save current chunk
+            if current_chunk and len(current_chunk) + len(section) > self.max_display_chars:
+                chunks.append(
+                    f"{'=' * 60}\n{chunk_prefix} {chunk_num} of {{total}}\n{'=' * 60}\n\n"
+                    + current_chunk,
+                )
+                chunk_num += 1
+                current_chunk = section
+            else:
+                current_chunk += section
+
+        # Add final chunk
+        if current_chunk:
+            chunks.append(
+                f"{'=' * 60}\n{chunk_prefix} {chunk_num} of {{total}}\n{'=' * 60}\n\n"
+                + current_chunk,
+            )
+
+        # Update total count in all chunks
+        total = len(chunks)
+        chunks = [chunk.format(total=total) for chunk in chunks]
+
+        return chunks
+
+    def should_skip_stage(self, stage_name: str, input_data: Any) -> tuple[bool, str | None]:
+        """Skip polish for short documents."""
+        if stage_name == "polish":
+            if self._total_content_tokens < self.skip_polish_threshold:
+                self.tier_map["polish"] = ModelTier.CAPABLE
+                return False, None
+        return False, None
+
+    async def run_stage(
+        self,
+        stage_name: str,
+        tier: ModelTier,
+        input_data: Any,
+    ) -> tuple[Any, int, int]:
+        """Execute a document generation stage."""
+        if stage_name == "outline":
+            return await self._outline(input_data, tier)
+        if stage_name == "write":
+            return await self._write(input_data, tier)
+        if stage_name == "polish":
+            return await self._polish(input_data, tier)
+        raise ValueError(f"Unknown stage: {stage_name}")
+
+    async def _outline(self, input_data: dict, tier: ModelTier) -> tuple[dict, int, int]:
+        """Generate document outline from source."""
+        from pathlib import Path
+
+        source_code = input_data.get("source_code", "")
+        target = input_data.get("target", "")
+        doc_type = input_data.get("doc_type", "general")
+        audience = input_data.get("audience", "developers")
+
+        # Use target if source_code not provided
+        content_to_document = source_code or target
+
+        # If target looks like a file path and source_code wasn't provided, read the file
+        if not source_code and target:
+            target_path = Path(target)
+            if target_path.exists() and target_path.is_file():
+                try:
+                    content_to_document = target_path.read_text(encoding="utf-8")
+                    # Prepend file info for context
+                    content_to_document = f"# File: {target}\n\n{content_to_document}"
+                except Exception as e:
+                    # If we can't read the file, log and use the path as-is
+                    import logging
+
+                    logging.getLogger(__name__).warning(f"Could not read file {target}: {e}")
+            elif target_path.suffix in (
+                ".py",
+                ".js",
+                ".ts",
+                ".tsx",
+                ".java",
+                ".go",
+                ".rs",
+                ".md",
+                ".txt",
+            ):
+                # Looks like a file path but doesn't exist - warn
+                import logging
+
+                logging.getLogger(__name__).warning(
+                    f"Target appears to be a file path but doesn't exist: {target}",
+                )
+
+        # === AUTH STRATEGY INTEGRATION ===
+        # Detect module size and recommend auth mode (first stage only)
+        if self.enable_auth_strategy:
+            try:
+                from attune.models import (
+                    count_lines_of_code,
+                    get_auth_strategy,
+                    get_module_size_category,
+                )
+
+                # Calculate module size
+                module_lines = 0
+                if target and Path(target).exists():
+                    module_lines = count_lines_of_code(target)
+                elif content_to_document:
+                    # Count from source code content
+                    module_lines = len(
+                        [
+                            line
+                            for line in content_to_document.split("\n")
+                            if line.strip() and not line.strip().startswith("#")
+                        ]
+                    )
+
+                if module_lines > 0:
+                    # Get auth strategy (first-time setup if needed)
+                    strategy = get_auth_strategy()
+
+                    # Get recommended auth mode
+                    recommended_mode = strategy.get_recommended_mode(module_lines)
+                    self._auth_mode_used = recommended_mode.value
+
+                    # Get size category
+                    size_category = get_module_size_category(module_lines)
+
+                    # Log recommendation
+                    logger.info(
+                        f"Module: {target or 'source'} ({module_lines} LOC, {size_category})"
+                    )
+                    logger.info(f"Recommended auth mode: {recommended_mode.value}")
+
+                    # Get cost estimate
+                    cost_estimate = strategy.estimate_cost(module_lines, recommended_mode)
+
+                    if recommended_mode.value == "subscription":
+                        logger.info(
+                            f"Cost: {cost_estimate['quota_cost']} "
+                            f"(fits in {cost_estimate['fits_in_context']} context)"
+                        )
+                    else:  # API
+                        logger.info(
+                            f"Cost: ~${cost_estimate['monetary_cost']:.4f} "
+                            f"(1M context window)"
+                        )
+
+            except Exception as e:
+                # Don't fail workflow if auth strategy fails
+                logger.warning(f"Auth strategy detection failed: {e}")
+
+        system = """You are an expert technical writer specializing in API Reference documentation.
+
+IMPORTANT: This is API REFERENCE documentation, not a tutorial. Focus on documenting EVERY function/class with structured Args/Returns/Raises format.
+
+Create a detailed, structured outline for API Reference documentation:
+
+1. **Logical Section Structure** (emphasize API reference sections):
+   - Overview/Introduction (brief)
+   - Quick Start (1 complete example)
+   - API Reference - Functions (one subsection per function with Args/Returns/Raises)
+   - API Reference - Classes (one subsection per class with Args/Returns/Raises for methods)
+   - Usage Examples (showing how to combine multiple functions)
+   - Additional reference sections as needed
+
+2. **For Each Section**:
+   - Clear purpose and what readers will learn
+   - Specific topics to cover
+   - Types of examples to include (with actual code)
+
+3. **Key Requirements**:
+   - Include sections for real, copy-paste ready code examples
+   - Plan for comprehensive API documentation with all parameters
+   - Include edge cases and error handling examples
+   - Add best practices and common patterns
+
+Format as a numbered list with section titles and detailed descriptions."""
+
+        user_message = f"""Create a comprehensive documentation outline:
+
+Document Type: {doc_type}
+Target Audience: {audience}
+
+IMPORTANT: This documentation should be production-ready with:
+- Real, executable code examples (not placeholders)
+- Complete API reference with parameter types and descriptions
+- Usage guides showing common patterns
+- Edge case handling and error scenarios
+- Best practices for the target audience
+
+Content to document:
+{content_to_document[:4000]}
+
+Generate an outline that covers all these aspects comprehensively."""
+
+        response, input_tokens, output_tokens = await self._call_llm(
+            tier,
+            system,
+            user_message,
+            max_tokens=1000,
+        )
+
+        return (
+            {
+                "outline": response,
+                "doc_type": doc_type,
+                "audience": audience,
+                "content_to_document": content_to_document,
+            },
+            input_tokens,
+            output_tokens,
+        )
+
+    def _parse_outline_sections(self, outline: str) -> list[str]:
+        """Parse top-level section titles from the outline.
+
+        Only matches main sections like "1. Introduction", "2. Setup", etc.
+        Ignores sub-sections like "2.1 Prerequisites" or nested items.
+        """
+        import re
+
+        sections = []
+        # Match only top-level sections: digit followed by period and space/letter
+        # e.g., "1. Introduction" but NOT "1.1 Sub-section" or "2.1.3 Deep"
+        top_level_pattern = re.compile(r"^(\d+)\.\s+([A-Za-z].*)")
+
+        for line in outline.split("\n"):
+            stripped = line.strip()
+            match = top_level_pattern.match(stripped)
+            if match:
+                # section_num = match.group(1) - not needed, only extracting title
+                title = match.group(2).strip()
+                # Remove any trailing description after " - "
+                if " - " in title:
+                    title = title.split(" - ")[0].strip()
+                sections.append(title)
+
+        return sections
+
+    async def _write(self, input_data: dict, tier: ModelTier) -> tuple[dict, int, int]:
+        """Write content based on the outline."""
+        outline = input_data.get("outline", "")
+        doc_type = input_data.get("doc_type", "general")
+        audience = input_data.get("audience", "developers")
+        content_to_document = input_data.get("content_to_document", "")
+
+        # Parse sections from outline
+        sections = self._parse_outline_sections(outline)
+
+        # Auto-scale tokens based on section count
+        self.max_write_tokens = self._auto_scale_tokens(len(sections))
+
+        # Use chunked generation for large outlines (more than sections_per_chunk * 2)
+        use_chunking = (
+            self.chunked_generation
+            and len(sections) > self.sections_per_chunk * 2
+            and not self.section_focus  # Don't chunk if already focused
+        )
+
+        if use_chunking:
+            return await self._write_chunked(
+                sections,
+                outline,
+                doc_type,
+                audience,
+                content_to_document,
+                tier,
+            )
+
+        # Handle section_focus for targeted generation
+        section_instruction = ""
+        if self.section_focus:
+            sections_list = ", ".join(self.section_focus)
+            section_instruction = f"""
+IMPORTANT: Focus ONLY on generating these specific sections:
+{sections_list}
+
+Generate comprehensive, detailed content for each of these sections."""
+
+        system = f"""You are an expert technical writer creating comprehensive developer documentation.
+
+YOUR TASK HAS TWO CRITICAL PHASES - YOU MUST COMPLETE BOTH:
+
+═══════════════════════════════════════════════════════════════
+PHASE 1: Write Comprehensive Documentation
+═══════════════════════════════════════════════════════════════
+
+Write clear, helpful documentation with:
+- Overview and introduction explaining what this code does
+- Real, executable code examples (NOT placeholders - use actual code from source)
+- Usage guides showing how to use the code in real scenarios
+- Best practices and common patterns
+- Step-by-step instructions where helpful
+- Tables, diagrams, and visual aids as appropriate
+- Clear explanations appropriate for {audience}
+
+Do this naturally - write the kind of documentation that helps developers understand and use the code effectively.
+
+═══════════════════════════════════════════════════════════════
+PHASE 2: Add Structured API Reference Sections (MANDATORY)
+═══════════════════════════════════════════════════════════════
+
+After writing the comprehensive documentation above, you MUST add structured API reference sections for EVERY function and class method.
+
+For EACH function/method in the source code, add this EXACT structure:
+
+---
+### `function_name()`
+
+**Function Signature:**
+```python
+def function_name(param1: type, param2: type = default) -> return_type
+```
+
+**Description:**
+[Brief description of what the function does - 1-2 sentences]
+
+**Args:**
+- `param1` (`type`): Clear description of this parameter
+- `param2` (`type`, optional): Description. Defaults to `default`.
+
+**Returns:**
+- `return_type`: Description of the return value
+
+**Raises:**
+- `ExceptionType`: Description of when and why this exception occurs
+- `AnotherException`: Another exception case
+
+**Example:**
+```python
+from module import function_name
+
+# Show real usage with actual code
+result = function_name(actual_value, param2=123)
+print(result)
+```
+---
+
+CRITICAL RULES FOR PHASE 2:
+- Include **Args:** header for ALL functions (write "None" if no parameters)
+- Include **Returns:** header for ALL functions (write "None" if void/no return)
+- Include **Raises:** header for ALL functions (write "None" if no exceptions)
+- Use backticks for code: `param_name` (`type`)
+- Document EVERY public function and method you see in the source code
+
+{section_instruction}
+
+═══════════════════════════════════════════════════════════════
+REMINDER: BOTH PHASES ARE MANDATORY
+═══════════════════════════════════════════════════════════════
+
+1. Write comprehensive documentation (Phase 1) - what you do naturally
+2. Add structured API reference sections (Phase 2) - for every function/method
+
+Do NOT skip Phase 2 after completing Phase 1. Both phases are required for complete documentation."""
+
+        user_message = f"""Write comprehensive, production-ready documentation in TWO PHASES:
+
+Document Type: {doc_type}
+Target Audience: {audience}
+
+Outline to follow:
+{outline}
+
+Source code to document (extract actual class names, function signatures, parameters):
+{content_to_document[:5000]}
+
+═══════════════════════════════════════════════════════════════
+YOUR TASK:
+═══════════════════════════════════════════════════════════════
+
+PHASE 1: Write comprehensive documentation
+- Use the outline above as your guide
+- Include real, executable code examples from the source
+- Show usage patterns, best practices, common workflows
+- Write clear explanations that help developers understand the code
+
+PHASE 2: Add structured API reference sections
+- For EACH function/method in the source code, add:
+  - Function signature
+  - Description
+  - **Args:** section (every parameter with type and description)
+  - **Returns:** section (return type and description)
+  - **Raises:** section (exceptions that can occur)
+  - Example code snippet
+
+═══════════════════════════════════════════════════════════════
+IMPORTANT: Complete BOTH phases. Don't stop after Phase 1.
+═══════════════════════════════════════════════════════════════
+
+Generate the complete documentation now, ensuring both comprehensive content AND structured API reference sections."""
+
+        response, input_tokens, output_tokens = await self._call_llm(
+            tier,
+            system,
+            user_message,
+            max_tokens=self.max_write_tokens,
+        )
+
+        self._total_content_tokens = output_tokens
+
+        return (
+            {
+                "draft_document": response,
+                "doc_type": doc_type,
+                "audience": audience,
+                "outline": outline,
+                "chunked": False,
+                "source_code": content_to_document,  # Pass through for API reference generation
+            },
+            input_tokens,
+            output_tokens,
+        )
+
+    async def _write_chunked(
+        self,
+        sections: list[str],
+        outline: str,
+        doc_type: str,
+        audience: str,
+        content_to_document: str,
+        tier: ModelTier,
+    ) -> tuple[dict, int, int]:
+        """Generate documentation in chunks to avoid truncation.
+
+        Enterprise-safe: includes cost tracking and graceful degradation.
+        """
+        all_content: list[str] = []
+        total_input_tokens: int = 0
+        total_output_tokens: int = 0
+        stopped_early: bool = False
+        error_message: str | None = None
+
+        # Split sections into chunks
+        chunks = []
+        for i in range(0, len(sections), self.sections_per_chunk):
+            chunks.append(sections[i : i + self.sections_per_chunk])
+
+        logger.info(f"Generating documentation in {len(chunks)} chunks")
+
+        for chunk_idx, chunk_sections in enumerate(chunks):
+            sections_list = ", ".join(chunk_sections)
+
+            # Build context about what came before
+            previous_context = ""
+            if chunk_idx > 0 and all_content:
+                # Include last 500 chars of previous content for continuity
+                previous_context = f"""
+Previous sections already written (for context/continuity):
+...{all_content[-1][-500:]}
+
+Continue with the next sections, maintaining consistent style and terminology."""
+
+            system = f"""You are an expert technical writer creating comprehensive developer documentation.
+
+Write ONLY these sections (part {chunk_idx + 1} of {len(chunks)}): {sections_list}
+
+YOUR TASK FOR THESE SECTIONS (TWO PHASES):
+
+═══════════════════════════════════════════════════════════════
+PHASE 1: Comprehensive Content
+═══════════════════════════════════════════════════════════════
+- Write clear explanations and overviews
+- Include real, executable code examples (extract from source)
+- Show usage patterns and workflows
+- Add best practices and common patterns
+- Professional language for {audience}
+
+═══════════════════════════════════════════════════════════════
+PHASE 2: Structured API Reference
+═══════════════════════════════════════════════════════════════
+For EACH function/method in these sections, add:
+
+### `function_name()`
+
+**Function Signature:**
+```python
+def function_name(params) -> return_type
+```
+
+**Description:**
+[Brief description]
+
+**Args:**
+- `param` (`type`): Description
+
+**Returns:**
+- `type`: Description
+
+**Raises:**
+- `Exception`: When it occurs
+
+**Example:**
+```python
+# Real usage example
+```
+
+═══════════════════════════════════════════════════════════════
+Complete BOTH phases for these sections.
+═══════════════════════════════════════════════════════════════"""
+
+            user_message = f"""Write comprehensive documentation for these sections in TWO PHASES:
+
+Sections to write: {sections_list}
+
+Document Type: {doc_type}
+Target Audience: {audience}
+
+Source code (extract actual functions/classes from here):
+{content_to_document[:3000]}
+
+Full outline (for context):
+{outline}
+{previous_context}
+
+PHASE 1: Write comprehensive content with real code examples
+PHASE 2: Add structured API reference sections with **Args:**, **Returns:**, **Raises:**
+
+Generate complete sections now, ensuring both phases are complete."""
+
+            try:
+                response, input_tokens, output_tokens = await self._call_llm(
+                    tier,
+                    system,
+                    user_message,
+                    max_tokens=self.max_write_tokens // len(chunks) + 2000,
+                )
+
+                # Track cost and check limits
+                _, should_stop = self._track_cost(tier, input_tokens, output_tokens)
+
+                all_content.append(response)
+                total_input_tokens += input_tokens
+                total_output_tokens += output_tokens
+
+                logger.info(
+                    f"Chunk {chunk_idx + 1}/{len(chunks)} complete: "
+                    f"{len(response)} chars, {output_tokens} tokens, "
+                    f"cost so far: ${self._accumulated_cost:.2f}",
+                )
+
+                # Check cost limit
+                if should_stop:
+                    stopped_early = True
+                    remaining = len(chunks) - chunk_idx - 1
+                    error_message = (
+                        f"Cost limit reached (${self._accumulated_cost:.2f}). "
+                        f"Stopped after {chunk_idx + 1}/{len(chunks)} chunks. "
+                        f"{remaining} chunks not generated."
+                    )
+                    logger.warning(error_message)
+                    break
+
+            except Exception as e:
+                error_message = f"Error generating chunk {chunk_idx + 1}: {e}"
+                logger.error(error_message)
+                if not self.graceful_degradation:
+                    raise
+                stopped_early = True
+                break
+
+        # Combine all chunks
+        combined_document = "\n\n".join(all_content)
+        self._total_content_tokens = total_output_tokens
+
+        # Store partial results for graceful degradation
+        self._partial_results = {
+            "draft_document": combined_document,
+            "sections_completed": len(all_content),
+            "sections_total": len(chunks),
+        }
+
+        result = {
+            "draft_document": combined_document,
+            "doc_type": doc_type,
+            "audience": audience,
+            "outline": outline,
+            "chunked": True,
+            "chunk_count": len(chunks),
+            "chunks_completed": len(all_content),
+            "stopped_early": stopped_early,
+            "accumulated_cost": self._accumulated_cost,
+            "source_code": content_to_document,  # Pass through for API reference generation
+        }
+
+        if error_message:
+            result["warning"] = error_message
+
+        return (result, total_input_tokens, total_output_tokens)
+
+    async def _polish(self, input_data: dict, tier: ModelTier) -> tuple[dict, int, int]:
+        """Final review and consistency polish using LLM.
+
+        Enterprise-safe: chunks large documents to avoid truncation.
+        Supports XML-enhanced prompts when enabled in workflow config.
+        """
+        draft_document = input_data.get("draft_document", "")
+        doc_type = input_data.get("doc_type", "general")
+        audience = input_data.get("audience", "developers")
+
+        # Check if document is too large and needs chunked polishing
+        # Rough estimate: 4 chars per token, 10k tokens threshold for chunking
+        estimated_tokens = len(draft_document) // 4
+        needs_chunked_polish = estimated_tokens > 10000
+
+        if needs_chunked_polish:
+            logger.info(
+                f"Large document detected (~{estimated_tokens} tokens). "
+                "Using chunked polish for enterprise safety.",
+            )
+            return await self._polish_chunked(input_data, tier)
+
+        # Build input payload for prompt
+        input_payload = f"""Document Type: {doc_type}
+Target Audience: {audience}
+
+Draft:
+{draft_document}"""
+
+        # Check if XML prompts are enabled
+        if self._is_xml_enabled():
+            # Use XML-enhanced prompt
+            user_message = self._render_xml_prompt(
+                role="senior technical editor",
+                goal="Polish and improve the documentation for consistency and quality",
+                instructions=[
+                    "Standardize terminology and formatting",
+                    "Improve clarity and flow",
+                    "Add missing cross-references",
+                    "Fix grammatical issues",
+                    "Identify gaps and add helpful notes",
+                    "Ensure examples are complete and accurate",
+                ],
+                constraints=[
+                    "Maintain the original structure and intent",
+                    "Keep content appropriate for the target audience",
+                    "Preserve code examples while improving explanations",
+                ],
+                input_type="documentation_draft",
+                input_payload=input_payload,
+                extra={
+                    "doc_type": doc_type,
+                    "audience": audience,
+                },
+            )
+            system = None  # XML prompt includes all context
+        else:
+            # Use legacy plain text prompts
+            system = """You are a senior technical editor specializing in developer documentation.
+
+Polish and improve this documentation. The writer was asked to complete TWO PHASES:
+- Phase 1: Comprehensive content with real examples
+- Phase 2: Structured API reference sections with **Args:**, **Returns:**, **Raises:**
+
+Your job is to verify BOTH phases are complete and polish to production quality.
+
+═══════════════════════════════════════════════════════════════
+CRITICAL: Verify Phase 2 Completion
+═══════════════════════════════════════════════════════════════
+
+1. **Check for Missing API Reference Sections**:
+   - Scan the entire document for all functions and methods
+   - EVERY function MUST have these sections:
+     - **Args:** (write "None" if no parameters)
+     - **Returns:** (write "None" if void)
+     - **Raises:** (write "None" if no exceptions)
+   - If ANY function is missing these sections, ADD them now
+   - Format: **Args:**, **Returns:**, **Raises:** (bold headers with colons)
+
+2. **Polish API Reference Sections**:
+   - Verify all parameters have types in backticks: `param` (`type`)
+   - Ensure return values are clearly described
+   - Check exception documentation is complete
+   - Validate code examples in each function section
+
+3. **Polish General Content**:
+   - Verify code examples are complete and runnable
+   - Ensure proper imports and setup code
+   - Replace any placeholders with real code
+   - Standardize terminology throughout
+   - Fix formatting inconsistencies
+   - Improve clarity and flow
+   - Add cross-references between sections
+
+4. **Production Readiness**:
+   - Remove any TODO or placeholder comments
+   - Ensure professional tone
+   - Add helpful notes, tips, and warnings
+   - Verify edge cases are covered
+
+═══════════════════════════════════════════════════════════════
+Return the complete, polished document. Add a brief "## Polish Notes" section at the end summarizing improvements made."""
+
+            user_message = f"""Polish this documentation to production quality.
+
+The writer was asked to complete TWO PHASES:
+1. Comprehensive content with real examples
+2. Structured API reference with **Args:**, **Returns:**, **Raises:** for every function
+
+Verify BOTH phases are complete, then polish:
+
+{input_payload}
+
+═══════════════════════════════════════════════════════════════
+YOUR TASKS:
+═══════════════════════════════════════════════════════════════
+
+1. SCAN for missing API reference sections
+   - Find every function/method in the document
+   - Check if it has **Args:**, **Returns:**, **Raises:** sections
+   - ADD these sections if missing (use "None" if no parameters/returns/exceptions)
+
+2. POLISH existing content
+   - Verify code examples are complete and runnable
+   - Ensure terminology is consistent
+   - Fix formatting issues
+   - Improve clarity and flow
+
+3. VALIDATE production readiness
+   - Remove TODOs and placeholders
+   - Add warnings and best practices
+   - Ensure professional tone
+
+Return the complete, polished documentation with all API reference sections present."""
+
+        # Calculate polish tokens based on draft size (at least as much as write stage)
+        polish_max_tokens = max(self.max_write_tokens, 20000)
+
+        # Try executor-based execution first (Phase 3 pattern)
+        if self._executor is not None or self._api_key:
+            try:
+                step = DOC_GEN_STEPS["polish"]
+                # Override step max_tokens with dynamic value
+                step.max_tokens = polish_max_tokens
+                response, input_tokens, output_tokens, cost = await self.run_step_with_executor(
+                    step=step,
+                    prompt=user_message,
+                    system=system,
+                )
+            except Exception:
+                # Fall back to legacy _call_llm if executor fails
+                response, input_tokens, output_tokens = await self._call_llm(
+                    tier,
+                    system or "",
+                    user_message,
+                    max_tokens=polish_max_tokens,
+                )
+        else:
+            # Legacy path for backward compatibility
+            response, input_tokens, output_tokens = await self._call_llm(
+                tier,
+                system or "",
+                user_message,
+                max_tokens=polish_max_tokens,
+            )
+
+        # Parse XML response if enforcement is enabled
+        parsed_data = self._parse_xml_response(response)
+
+        # Add structured API reference sections (Step 4: Post-processing)
+        source_code = input_data.get("source_code", "")
+        if source_code:
+            logger.info("Adding structured API reference sections to polished document...")
+            response = await self._add_api_reference_sections(
+                narrative_doc=response,
+                source_code=source_code,
+                tier=ModelTier.CHEAP,  # Use cheap tier for structured extraction
+            )
+        else:
+            logger.warning("No source code available for API reference generation")
+
+        result = {
+            "document": response,
+            "doc_type": doc_type,
+            "audience": audience,
+            "model_tier_used": tier.value,
+            "accumulated_cost": self._accumulated_cost,  # Track total cost
+            "auth_mode_used": self._auth_mode_used,  # Track recommended auth mode
+        }
+
+        # Merge parsed XML data if available
+        if parsed_data.get("xml_parsed"):
+            result.update(
+                {
+                    "xml_parsed": True,
+                    "summary": parsed_data.get("summary"),
+                    "findings": parsed_data.get("findings", []),
+                    "checklist": parsed_data.get("checklist", []),
+                },
+            )
+
+        # Add formatted report for human readability
+        result["formatted_report"] = format_doc_gen_report(result, input_data)
+
+        # Export documentation if export_path is configured
+        doc_path, report_path = self._export_document(
+            document=response,
+            doc_type=doc_type,
+            report=result["formatted_report"],
+        )
+        if doc_path:
+            result["export_path"] = str(doc_path)
+            result["report_path"] = str(report_path) if report_path else None
+            logger.info(f"Documentation saved to: {doc_path}")
+
+        # Chunk output for display if needed
+        output_chunks = self._chunk_output_for_display(
+            result["formatted_report"],
+            chunk_prefix="DOC OUTPUT",
+        )
+        if len(output_chunks) > 1:
+            result["output_chunks"] = output_chunks
+            result["output_chunk_count"] = len(output_chunks)
+            logger.info(
+                f"Report split into {len(output_chunks)} chunks for display "
+                f"(total {len(result['formatted_report'])} chars)",
+            )
+
+        return (result, input_tokens, output_tokens)
+
+    async def _polish_chunked(self, input_data: dict, tier: ModelTier) -> tuple[dict, int, int]:
+        """Polish large documents in chunks to avoid truncation.
+
+        Splits the document by section headers and polishes each chunk separately,
+        then combines the results.
+        """
+        import re
+
+        draft_document = input_data.get("draft_document", "")
+        doc_type = input_data.get("doc_type", "general")
+        audience = input_data.get("audience", "developers")
+
+        # Split document by major section headers (## headers)
+        sections = re.split(r"(?=^## )", draft_document, flags=re.MULTILINE)
+        sections = [s.strip() for s in sections if s.strip()]
+
+        if len(sections) <= 1:
+            # If we can't split by sections, split by character count
+            chunk_size = 15000  # ~3750 tokens per chunk
+            sections = [
+                draft_document[i : i + chunk_size]
+                for i in range(0, len(draft_document), chunk_size)
+            ]
+
+        logger.info(f"Polishing document in {len(sections)} chunks")
+
+        polished_chunks: list[str] = []
+        total_input_tokens: int = 0
+        total_output_tokens: int = 0
+
+        for chunk_idx, section in enumerate(sections):
+            system = """You are a senior technical editor specializing in developer documentation.
+
+Polish this section to production quality. The writer was asked to complete TWO PHASES:
+1. Comprehensive content with real examples
+2. Structured API reference with **Args:**, **Returns:**, **Raises:** for every function
+
+Verify both phases are complete in this section:
+
+═══════════════════════════════════════════════════════════════
+CRITICAL: Check for Missing API Reference Format
+═══════════════════════════════════════════════════════════════
+
+1. **Scan for functions/methods in this section**
+   - If any function is missing **Args:**, **Returns:**, **Raises:** sections, ADD them
+   - Format: **Args:**, **Returns:**, **Raises:** (bold headers with colons)
+   - Write "None" if no parameters/returns/exceptions
+
+2. **Polish API Documentation**:
+   - Verify parameters documented with types in backticks
+   - Ensure return values and exceptions are clear
+   - Validate code examples are complete
+
+3. **Polish General Content**:
+   - Ensure all examples are runnable with proper imports
+   - Standardize terminology and formatting
+   - Fix grammatical issues
+   - Remove TODOs and placeholders
+
+Return ONLY the polished section. Do not add commentary about changes."""
+
+            user_message = f"""Polish this section to production quality (part {chunk_idx + 1} of {len(sections)}):
+
+Document Type: {doc_type}
+Target Audience: {audience}
+
+Section to polish:
+{section}
+
+Check if all functions have **Args:**, **Returns:**, **Raises:** sections - add if missing.
+Make all code examples complete and executable."""
+
+            try:
+                response, input_tokens, output_tokens = await self._call_llm(
+                    tier,
+                    system,
+                    user_message,
+                    max_tokens=8000,
+                )
+
+                # Track cost
+                _, should_stop = self._track_cost(tier, input_tokens, output_tokens)
+
+                polished_chunks.append(response)
+                total_input_tokens += input_tokens
+                total_output_tokens += output_tokens
+
+                logger.info(
+                    f"Polish chunk {chunk_idx + 1}/{len(sections)} complete, "
+                    f"cost so far: ${self._accumulated_cost:.2f}",
+                )
+
+                if should_stop:
+                    logger.warning(
+                        f"Cost limit reached during polish. "
+                        f"Returning {len(polished_chunks)}/{len(sections)} polished chunks.",
+                    )
+                    # Add remaining sections unpolished
+                    polished_chunks.extend(sections[chunk_idx + 1 :])
+                    break
+
+            except Exception as e:
+                logger.error(f"Error polishing chunk {chunk_idx + 1}: {e}")
+                if self.graceful_degradation:
+                    # Keep original section on error
+                    polished_chunks.append(section)
+                else:
+                    raise
+
+        # Combine polished chunks
+        polished_document = "\n\n".join(polished_chunks)
+
+        # Add structured API reference sections (Step 4: Post-processing)
+        source_code = input_data.get("source_code", "")
+        if source_code:
+            logger.info("Adding structured API reference sections to chunked polished document...")
+            polished_document = await self._add_api_reference_sections(
+                narrative_doc=polished_document,
+                source_code=source_code,
+                tier=ModelTier.CHEAP,  # Use cheap tier for structured extraction
+            )
+        else:
+            logger.warning("No source code available for API reference generation")
+
+        result = {
+            "document": polished_document,
+            "doc_type": doc_type,
+            "audience": audience,
+            "model_tier_used": tier.value,
+            "polish_chunked": True,
+            "polish_chunks": len(sections),
+            "accumulated_cost": self._accumulated_cost,
+        }
+
+        # Add formatted report
+        result["formatted_report"] = format_doc_gen_report(result, input_data)
+
+        # Export documentation if export_path is configured
+        doc_path, report_path = self._export_document(
+            document=polished_document,
+            doc_type=doc_type,
+            report=result["formatted_report"],
+        )
+        if doc_path:
+            result["export_path"] = str(doc_path)
+            result["report_path"] = str(report_path) if report_path else None
+            logger.info(f"Documentation saved to: {doc_path}")
+
+        # Chunk output for display if needed
+        output_chunks = self._chunk_output_for_display(
+            result["formatted_report"],
+            chunk_prefix="DOC OUTPUT",
+        )
+        if len(output_chunks) > 1:
+            result["output_chunks"] = output_chunks
+            result["output_chunk_count"] = len(output_chunks)
+            logger.info(
+                f"Report split into {len(output_chunks)} chunks for display "
+                f"(total {len(result['formatted_report'])} chars)",
+            )
+
+        return (result, total_input_tokens, total_output_tokens)
+
+    def _extract_functions_from_source(self, source_code: str) -> list[dict]:
+        """Extract function information from source code using AST.
+
+        Args:
+            source_code: Python source code to parse
+
+        Returns:
+            List of dicts with function information (name, args, returns, docstring)
+        """
+        import ast
+
+        functions = []
+
+        try:
+            tree = ast.parse(source_code)
+        except SyntaxError as e:
+            logger.warning(f"Failed to parse source code: {e}")
+            return functions
+
+        for node in ast.walk(tree):
+            # Extract top-level functions and class methods
+            if isinstance(node, ast.FunctionDef):
+                # Skip private functions (starting with _)
+                if node.name.startswith("_"):
+                    continue
+
+                # Extract function signature
+                args_list = []
+                for arg in node.args.args:
+                    arg_name = arg.arg
+                    # Get type annotation if available
+                    arg_type = ast.unparse(arg.annotation) if arg.annotation else "Any"
+                    args_list.append({"name": arg_name, "type": arg_type})
+
+                # Extract return type
+                return_type = ast.unparse(node.returns) if node.returns else "Any"
+
+                # Extract docstring
+                docstring = ast.get_docstring(node) or ""
+
+                functions.append({
+                    "name": node.name,
+                    "args": args_list,
+                    "return_type": return_type,
+                    "docstring": docstring,
+                    "lineno": node.lineno,
+                })
+
+        return functions
+
+    async def _generate_api_section_for_function(
+        self,
+        func_info: dict,
+        tier: ModelTier,
+    ) -> str:
+        """Generate structured API reference section for a single function.
+
+        This is a focused prompt that ONLY asks for Args/Returns/Raises format,
+        not narrative documentation.
+
+        Args:
+            func_info: Function information from AST extraction
+            tier: Model tier to use for generation
+
+        Returns:
+            Markdown formatted API reference section
+        """
+        func_name = func_info["name"]
+        args_list = func_info["args"]
+        return_type = func_info["return_type"]
+        docstring = func_info["docstring"]
+
+        # Build function signature
+        args_str = ", ".join([f"{arg['name']}: {arg['type']}" for arg in args_list])
+        signature = f"def {func_name}({args_str}) -> {return_type}"
+
+        system = """You are an API documentation generator. Output ONLY structured API reference sections in the EXACT format specified below.
+
+CRITICAL: Do NOT write explanatory text, questions, or narrative. Output ONLY the formatted section.
+
+REQUIRED FORMAT (copy this structure EXACTLY, replace bracketed content):
+
+### `function_name()`
+
+**Function Signature:**
+```python
+def function_name(param: type) -> return_type
+```
+
+**Description:**
+Brief 1-2 sentence description.
+
+**Args:**
+- `param_name` (`type`): Parameter description
+
+**Returns:**
+- `return_type`: Return value description
+
+**Raises:**
+- `ExceptionType`: When this exception occurs
+
+IMPORTANT:
+- Use "**Args:**" (NOT "Parameters" or "params")
+- Write "None" if no Args/Returns/Raises
+- NO conversational text - just the formatted section"""
+
+        user_message = f"""Generate API reference section using EXACT format specified in system prompt.
+
+Function:
+```python
+{signature}
+```
+
+Docstring:
+{docstring if docstring else "No docstring"}
+
+Output the formatted section EXACTLY as shown in system prompt. Use **Args:** (not Parameters). NO conversational text."""
+
+        try:
+            response, input_tokens, output_tokens = await self._call_llm(
+                tier,
+                system,
+                user_message,
+                max_tokens=1000,  # Small response - just the structured section
+            )
+
+            # Track cost
+            self._track_cost(tier, input_tokens, output_tokens)
+
+            return response
+
+        except Exception as e:
+            logger.error(f"Failed to generate API section for {func_name}: {e}")
+            # Return minimal fallback
+            return f"""### `{func_name}()`
+
+**Function Signature:**
+```python
+{signature}
+```
+
+**Description:**
+{docstring.split('.')[0] if docstring else "No description available."}
+
+**Args:**
+None
+
+**Returns:**
+- `{return_type}`: Return value
+
+**Raises:**
+None
+"""
+
+    async def _add_api_reference_sections(
+        self,
+        narrative_doc: str,
+        source_code: str,
+        tier: ModelTier,
+    ) -> str:
+        """Add structured API reference sections to narrative documentation.
+
+        This is Step 4 of the pipeline: after outline, write, and polish,
+        we add structured API reference sections extracted from source code.
+
+        Args:
+            narrative_doc: The polished narrative documentation
+            source_code: Original source code to extract functions from
+            tier: Model tier to use for API section generation
+
+        Returns:
+            Complete documentation with API reference appendix
+        """
+        logger.info("Adding structured API reference sections...")
+
+        # Extract functions from source code
+        functions = self._extract_functions_from_source(source_code)
+
+        if not functions:
+            logger.warning("No public functions found in source code")
+            return narrative_doc
+
+        logger.info(f"Found {len(functions)} public functions to document")
+
+        # Generate API section for each function
+        api_sections = []
+        for func_info in functions:
+            func_name = func_info["name"]
+            logger.debug(f"Generating API reference for {func_name}()")
+
+            api_section = await self._generate_api_section_for_function(
+                func_info, tier
+            )
+            api_sections.append(api_section)
+
+        # Append API reference section to narrative doc
+        full_doc = narrative_doc
+        full_doc += "\n\n---\n\n"
+        full_doc += "## API Reference\n\n"
+        full_doc += "Complete structured reference for all public functions:\n\n"
+        full_doc += "\n\n".join(api_sections)
+
+        logger.info(f"Added {len(api_sections)} API reference sections")
+
+        return full_doc
+
+