chore: import upstream snapshot with attribution

application/api/answer/services/compression/__init__.py

@@ -0,0 +1,20 @@
+"""
+Compression module for managing conversation context compression.
+
+"""
+
+from application.api.answer.services.compression.orchestrator import (
+    CompressionOrchestrator,
+)
+from application.api.answer.services.compression.service import CompressionService
+from application.api.answer.services.compression.types import (
+    CompressionResult,
+    CompressionMetadata,
+)
+
+__all__ = [
+    "CompressionOrchestrator",
+    "CompressionService",
+    "CompressionResult",
+    "CompressionMetadata",
+]

application/api/answer/services/compression/message_builder.py

@@ -0,0 +1,249 @@
+"""Message reconstruction utilities for compression."""
+
+import json
+import logging
+import uuid
+from typing import Dict, List, Optional
+
+logger = logging.getLogger(__name__)
+
+
+class MessageBuilder:
+    """Builds message arrays from compressed context."""
+
+    @staticmethod
+    def build_from_compressed_context(
+        system_prompt: str,
+        compressed_summary: Optional[str],
+        recent_queries: List[Dict],
+        include_tool_calls: bool = False,
+        context_type: str = "pre_request",
+    ) -> List[Dict]:
+        """
+        Build messages from compressed context.
+
+        Args:
+            system_prompt: Original system prompt
+            compressed_summary: Compressed summary (if any)
+            recent_queries: Recent uncompressed queries
+            include_tool_calls: Whether to include tool calls from history
+            context_type: Type of context ('pre_request' or 'mid_execution')
+
+        Returns:
+            List of message dicts ready for LLM
+        """
+        # Append compression summary to system prompt if present
+        if compressed_summary:
+            system_prompt = MessageBuilder._append_compression_context(
+                system_prompt, compressed_summary, context_type
+            )
+
+        messages = [{"role": "system", "content": system_prompt}]
+
+        # Add recent history
+        for query in recent_queries:
+            if "prompt" in query and "response" in query:
+                messages.append({"role": "user", "content": query["prompt"]})
+                messages.append({"role": "assistant", "content": query["response"]})
+
+            # Add tool calls from history if present
+            if include_tool_calls and "tool_calls" in query:
+                for tool_call in query["tool_calls"]:
+                    call_id = tool_call.get("call_id") or str(uuid.uuid4())
+                    args = tool_call.get("arguments")
+                    args_str = (
+                        json.dumps(args)
+                        if isinstance(args, dict)
+                        else (args or "{}")
+                    )
+                    messages.append({
+                        "role": "assistant",
+                        "content": None,
+                        "tool_calls": [{
+                            "id": call_id,
+                            "type": "function",
+                            "function": {
+                                "name": tool_call.get("action_name", ""),
+                                "arguments": args_str,
+                            },
+                        }],
+                    })
+                    result = tool_call.get("result")
+                    result_str = (
+                        json.dumps(result)
+                        if not isinstance(result, str)
+                        else (result or "")
+                    )
+                    messages.append({
+                        "role": "tool",
+                        "tool_call_id": call_id,
+                        "content": result_str,
+                    })
+
+        # If no recent queries (everything was compressed), add a continuation user message
+        if len(recent_queries) == 0 and compressed_summary:
+            messages.append({
+                "role": "user",
+                "content": "Please continue with the remaining tasks based on the context above."
+            })
+            logger.info("Added continuation user message to maintain proper turn-taking after full compression")
+
+        return messages
+
+    @staticmethod
+    def _append_compression_context(
+        system_prompt: str, compressed_summary: str, context_type: str = "pre_request"
+    ) -> str:
+        """
+        Append compression context to system prompt.
+
+        Args:
+            system_prompt: Original system prompt
+            compressed_summary: Summary to append
+            context_type: Type of compression context
+
+        Returns:
+            Updated system prompt
+        """
+        # Remove existing compression context if present
+        if "This session is being continued" in system_prompt or "Context window limit reached" in system_prompt:
+            parts = system_prompt.split("\n\n---\n\n")
+            system_prompt = parts[0]
+
+        # Build appropriate context message based on type
+        if context_type == "mid_execution":
+            context_message = (
+                "\n\n---\n\n"
+                "Context window limit reached during execution. "
+                "Previous conversation has been compressed to fit within limits. "
+                "The conversation is summarized below:\n\n"
+                f"{compressed_summary}"
+            )
+        else:  # pre_request
+            context_message = (
+                "\n\n---\n\n"
+                "This session is being continued from a previous conversation that "
+                "has been compressed to fit within context limits. "
+                "The conversation is summarized below:\n\n"
+                f"{compressed_summary}"
+            )
+
+        return system_prompt + context_message
+
+    @staticmethod
+    def rebuild_messages_after_compression(
+        messages: List[Dict],
+        compressed_summary: Optional[str],
+        recent_queries: List[Dict],
+        include_current_execution: bool = False,
+        include_tool_calls: bool = False,
+    ) -> Optional[List[Dict]]:
+        """
+        Rebuild the message list after compression so tool execution can continue.
+
+        Args:
+            messages: Original message list
+            compressed_summary: Compressed summary
+            recent_queries: Recent uncompressed queries
+            include_current_execution: Whether to preserve current execution messages
+            include_tool_calls: Whether to include tool calls from history
+
+        Returns:
+            Rebuilt message list or None if failed
+        """
+        # Find the system message
+        system_message = next(
+            (msg for msg in messages if msg.get("role") == "system"), None
+        )
+        if not system_message:
+            logger.warning("No system message found in messages list")
+            return None
+
+        # Update system message with compressed summary
+        if compressed_summary:
+            content = system_message.get("content", "")
+            system_message["content"] = MessageBuilder._append_compression_context(
+                content, compressed_summary, "mid_execution"
+            )
+            logger.info(
+                "Appended compression summary to system prompt (truncated): %s",
+                (
+                    compressed_summary[:500] + "..."
+                    if len(compressed_summary) > 500
+                    else compressed_summary
+                ),
+            )
+
+        rebuilt_messages = [system_message]
+
+        # Add recent history from compressed context
+        for query in recent_queries:
+            if "prompt" in query and "response" in query:
+                rebuilt_messages.append({"role": "user", "content": query["prompt"]})
+                rebuilt_messages.append(
+                    {"role": "assistant", "content": query["response"]}
+                )
+
+            # Add tool calls from history if present
+            if include_tool_calls and "tool_calls" in query:
+                for tool_call in query["tool_calls"]:
+                    call_id = tool_call.get("call_id") or str(uuid.uuid4())
+                    args = tool_call.get("arguments")
+                    args_str = (
+                        json.dumps(args)
+                        if isinstance(args, dict)
+                        else (args or "{}")
+                    )
+                    rebuilt_messages.append({
+                        "role": "assistant",
+                        "content": None,
+                        "tool_calls": [{
+                            "id": call_id,
+                            "type": "function",
+                            "function": {
+                                "name": tool_call.get("action_name", ""),
+                                "arguments": args_str,
+                            },
+                        }],
+                    })
+                    result = tool_call.get("result")
+                    result_str = (
+                        json.dumps(result)
+                        if not isinstance(result, str)
+                        else (result or "")
+                    )
+                    rebuilt_messages.append({
+                        "role": "tool",
+                        "tool_call_id": call_id,
+                        "content": result_str,
+                    })
+
+        # If no recent queries (everything was compressed), add a continuation user message
+        if len(recent_queries) == 0 and compressed_summary:
+            rebuilt_messages.append({
+                "role": "user",
+                "content": "Please continue with the remaining tasks based on the context above."
+            })
+            logger.info("Added continuation user message to maintain proper turn-taking after full compression")
+
+        if include_current_execution:
+            # Preserve any messages that were added during the current execution cycle
+            recent_msg_count = 1  # system message
+            for query in recent_queries:
+                if "prompt" in query and "response" in query:
+                    recent_msg_count += 2
+                if "tool_calls" in query:
+                    recent_msg_count += len(query["tool_calls"]) * 2
+
+            if len(messages) > recent_msg_count:
+                current_execution_messages = messages[recent_msg_count:]
+                rebuilt_messages.extend(current_execution_messages)
+                logger.info(
+                    f"Preserved {len(current_execution_messages)} messages from current execution cycle"
+                )
+
+        logger.info(
+            f"Messages rebuilt: {len(messages)} → {len(rebuilt_messages)} messages. "
+            f"Ready to continue tool execution."
+        )
+        return rebuilt_messages

application/api/answer/services/compression/orchestrator.py

@@ -0,0 +1,273 @@
+"""High-level compression orchestration."""
+
+import logging
+from typing import Any, Dict, Optional
+
+from application.api.answer.services.compression.service import CompressionService
+from application.api.answer.services.compression.threshold_checker import (
+    CompressionThresholdChecker,
+)
+from application.api.answer.services.compression.types import CompressionResult
+from application.api.answer.services.conversation_service import ConversationService
+from application.core.model_utils import (
+    get_api_key_for_provider,
+    get_provider_from_model_id,
+)
+from application.core.settings import settings
+from application.llm.llm_creator import LLMCreator
+
+logger = logging.getLogger(__name__)
+
+
+class CompressionOrchestrator:
+    """
+    Facade for compression operations.
+
+    Coordinates between all compression components and provides
+    a simple interface for callers.
+    """
+
+    def __init__(
+        self,
+        conversation_service: ConversationService,
+        threshold_checker: Optional[CompressionThresholdChecker] = None,
+    ):
+        """
+        Initialize orchestrator.
+
+        Args:
+            conversation_service: Service for DB operations
+            threshold_checker: Custom threshold checker (optional)
+        """
+        self.conversation_service = conversation_service
+        self.threshold_checker = threshold_checker or CompressionThresholdChecker()
+
+    def compress_if_needed(
+        self,
+        conversation_id: str,
+        user_id: str,
+        model_id: str,
+        decoded_token: Dict[str, Any],
+        current_query_tokens: int = 500,
+        model_user_id: Optional[str] = None,
+    ) -> CompressionResult:
+        """
+        Check if compression is needed and perform it if so.
+
+        This is the main entry point for compression operations.
+
+        Args:
+            conversation_id: Conversation ID
+            user_id: Caller's user id — used for conversation access checks
+            model_id: Model being used for conversation
+            decoded_token: User's decoded JWT token
+            current_query_tokens: Estimated tokens for current query
+            model_user_id: BYOM-resolution scope (model owner); defaults
+                to ``user_id`` for built-in / caller-owned models.
+
+        Returns:
+            CompressionResult with summary and recent queries
+        """
+        try:
+            # Conversation row is owned by the caller, not the model owner.
+            conversation = self.conversation_service.get_conversation(
+                conversation_id, user_id
+            )
+
+            if not conversation:
+                logger.warning(
+                    f"Conversation {conversation_id} not found for user {user_id}"
+                )
+                return CompressionResult.failure("Conversation not found")
+
+            # Use model-owner scope so per-user BYOM context windows
+            # (e.g. 8k) compute the threshold against the right limit.
+            registry_user_id = model_user_id or user_id
+            if not self.threshold_checker.should_compress(
+                conversation,
+                model_id,
+                current_query_tokens,
+                user_id=registry_user_id,
+            ):
+                # No compression needed, return full history
+                queries = conversation.get("queries", [])
+                return CompressionResult.success_no_compression(queries)
+
+            # Perform compression
+            return self._perform_compression(
+                conversation_id,
+                conversation,
+                model_id,
+                decoded_token,
+                user_id=user_id,
+                model_user_id=model_user_id,
+            )
+
+        except Exception as e:
+            logger.error(
+                f"Error in compress_if_needed: {str(e)}", exc_info=True
+            )
+            return CompressionResult.failure(str(e))
+
+    def _perform_compression(
+        self,
+        conversation_id: str,
+        conversation: Dict[str, Any],
+        model_id: str,
+        decoded_token: Dict[str, Any],
+        user_id: Optional[str] = None,
+        model_user_id: Optional[str] = None,
+    ) -> CompressionResult:
+        """
+        Perform the actual compression operation.
+
+        Args:
+            conversation_id: Conversation ID
+            conversation: Conversation document
+            model_id: Model ID for conversation
+            decoded_token: User token
+            user_id: Caller's id (for conversation reload after compression)
+            model_user_id: BYOM-resolution scope (model owner)
+
+        Returns:
+            CompressionResult
+        """
+        try:
+            # Determine which model to use for compression
+            compression_model = (
+                settings.COMPRESSION_MODEL_OVERRIDE
+                if settings.COMPRESSION_MODEL_OVERRIDE
+                else model_id
+            )
+
+            # Use model-owner scope so provider/api_key resolves to the
+            # owner's BYOM record (shared-agent dispatch).
+            caller_user_id = user_id
+            if caller_user_id is None and isinstance(decoded_token, dict):
+                caller_user_id = decoded_token.get("sub")
+            registry_user_id = model_user_id or caller_user_id
+            provider = get_provider_from_model_id(
+                compression_model, user_id=registry_user_id
+            )
+            api_key = get_api_key_for_provider(provider)
+
+            compression_llm = LLMCreator.create_llm(
+                provider,
+                api_key=api_key,
+                user_api_key=None,
+                decoded_token=decoded_token,
+                model_id=compression_model,
+                agent_id=conversation.get("agent_id"),
+                model_user_id=registry_user_id,
+            )
+            # Side-channel LLM tag — distinguishes compression rows
+            # from primary stream rows for cost-attribution dashboards.
+            compression_llm._token_usage_source = "compression"
+
+            # Create compression service with DB update capability
+            compression_service = CompressionService(
+                llm=compression_llm,
+                model_id=compression_model,
+                conversation_service=self.conversation_service,
+            )
+
+            # Compress all queries up to the latest
+            queries_count = len(conversation.get("queries", []))
+            compress_up_to = queries_count - 1
+
+            if compress_up_to < 0:
+                logger.warning("No queries to compress")
+                return CompressionResult.success_no_compression([])
+
+            logger.info(
+                f"Initiating compression for conversation {conversation_id}: "
+                f"compressing all {queries_count} queries (0-{compress_up_to})"
+            )
+
+            # Perform compression and save to DB
+            metadata = compression_service.compress_and_save(
+                conversation_id, conversation, compress_up_to
+            )
+
+            logger.info(
+                f"Compression successful - ratio: {metadata.compression_ratio:.1f}x, "
+                f"saved {metadata.original_token_count - metadata.compressed_token_count} tokens"
+            )
+
+            # Reload under caller (conversation is owned by caller).
+            reload_user_id = caller_user_id
+            if reload_user_id is None and isinstance(decoded_token, dict):
+                reload_user_id = decoded_token.get("sub")
+            conversation = self.conversation_service.get_conversation(
+                conversation_id, user_id=reload_user_id
+            )
+
+            # Get compressed context
+            compressed_summary, recent_queries = (
+                compression_service.get_compressed_context(conversation)
+            )
+
+            return CompressionResult.success_with_compression(
+                compressed_summary, recent_queries, metadata
+            )
+
+        except Exception as e:
+            logger.error(f"Error performing compression: {str(e)}", exc_info=True)
+            return CompressionResult.failure(str(e))
+
+    def compress_mid_execution(
+        self,
+        conversation_id: str,
+        user_id: str,
+        model_id: str,
+        decoded_token: Dict[str, Any],
+        current_conversation: Optional[Dict[str, Any]] = None,
+        model_user_id: Optional[str] = None,
+    ) -> CompressionResult:
+        """
+        Perform compression during tool execution.
+
+        Args:
+            conversation_id: Conversation ID
+            user_id: Caller's user id — used for conversation access checks
+            model_id: Model ID
+            decoded_token: User token
+            current_conversation: Pre-loaded conversation (optional)
+            model_user_id: BYOM-resolution scope (model owner). For
+                shared-agent dispatch this is the agent owner; defaults
+                to ``user_id`` so built-in / caller-owned models are
+                unaffected.
+
+        Returns:
+            CompressionResult
+        """
+        try:
+            # Load conversation if not provided
+            if current_conversation:
+                conversation = current_conversation
+            else:
+                conversation = self.conversation_service.get_conversation(
+                    conversation_id, user_id
+                )
+
+            if not conversation:
+                logger.warning(
+                    f"Could not load conversation {conversation_id} for mid-execution compression"
+                )
+                return CompressionResult.failure("Conversation not found")
+
+            # Perform compression
+            return self._perform_compression(
+                conversation_id,
+                conversation,
+                model_id,
+                decoded_token,
+                user_id=user_id,
+                model_user_id=model_user_id,
+            )
+
+        except Exception as e:
+            logger.error(
+                f"Error in mid-execution compression: {str(e)}", exc_info=True
+            )
+            return CompressionResult.failure(str(e))

application/api/answer/services/compression/prompt_builder.py

@@ -0,0 +1,149 @@
+"""Compression prompt building logic."""
+
+import logging
+from pathlib import Path
+from typing import Any, Dict, List, Optional
+
+logger = logging.getLogger(__name__)
+
+
+class CompressionPromptBuilder:
+    """Builds prompts for LLM compression calls."""
+
+    def __init__(self, version: str = "v1.0"):
+        """
+        Initialize prompt builder.
+
+        Args:
+            version: Prompt template version to use
+        """
+        self.version = version
+        self.system_prompt = self._load_prompt(version)
+
+    def _load_prompt(self, version: str) -> str:
+        """
+        Load prompt template from file.
+
+        Args:
+            version: Version string (e.g., 'v1.0')
+
+        Returns:
+            Prompt template content
+
+        Raises:
+            FileNotFoundError: If prompt template file doesn't exist
+        """
+        current_dir = Path(__file__).resolve().parents[4]
+        prompt_path = current_dir / "prompts" / "compression" / f"{version}.txt"
+
+        try:
+            with open(prompt_path, "r") as f:
+                return f.read()
+        except FileNotFoundError:
+            logger.error(f"Compression prompt template not found: {prompt_path}")
+            raise FileNotFoundError(
+                f"Compression prompt template '{version}' not found at {prompt_path}. "
+                f"Please ensure the template file exists."
+            )
+
+    def build_prompt(
+        self,
+        queries: List[Dict[str, Any]],
+        existing_compressions: Optional[List[Dict[str, Any]]] = None,
+    ) -> List[Dict[str, str]]:
+        """
+        Build messages for compression LLM call.
+
+        Args:
+            queries: List of query objects to compress
+            existing_compressions: List of previous compression points
+
+        Returns:
+            List of message dicts for LLM
+        """
+        # Build conversation text
+        conversation_text = self._format_conversation(queries)
+
+        # Add existing compression context if present
+        existing_compression_context = ""
+        if existing_compressions and len(existing_compressions) > 0:
+            existing_compression_context = (
+                "\n\nIMPORTANT: This conversation has been compressed before. "
+                "Previous compression summaries:\n\n"
+            )
+            for i, comp in enumerate(existing_compressions):
+                existing_compression_context += (
+                    f"--- Compression {i + 1} (up to message {comp.get('query_index', 'unknown')}) ---\n"
+                    f"{comp.get('compressed_summary', '')}\n\n"
+                )
+            existing_compression_context += (
+                "Your task is to create a NEW summary that incorporates the context from "
+                "previous compressions AND the new messages below. The final summary should "
+                "be comprehensive and include all important information from both previous "
+                "compressions and new messages.\n\n"
+            )
+
+        user_prompt = (
+            f"{existing_compression_context}"
+            f"Here is the conversation to summarize:\n\n"
+            f"{conversation_text}"
+        )
+
+        messages = [
+            {"role": "system", "content": self.system_prompt},
+            {"role": "user", "content": user_prompt},
+        ]
+
+        return messages
+
+    def _format_conversation(self, queries: List[Dict[str, Any]]) -> str:
+        """
+        Format conversation queries into readable text for compression.
+
+        Args:
+            queries: List of query objects
+
+        Returns:
+            Formatted conversation text
+        """
+        conversation_lines = []
+
+        for i, query in enumerate(queries):
+            conversation_lines.append(f"--- Message {i + 1} ---")
+            conversation_lines.append(f"User: {query.get('prompt', '')}")
+
+            # Add tool calls if present
+            tool_calls = query.get("tool_calls", [])
+            if tool_calls:
+                conversation_lines.append("\nTool Calls:")
+                for tc in tool_calls:
+                    tool_name = tc.get("tool_name", "unknown")
+                    action_name = tc.get("action_name", "unknown")
+                    arguments = tc.get("arguments", {})
+                    result = tc.get("result", "")
+                    if result is None:
+                        result = ""
+                    status = tc.get("status", "unknown")
+
+                    # Include full tool result for complete compression context
+                    conversation_lines.append(
+                        f"  - {tool_name}.{action_name}({arguments}) "
+                        f"[{status}] → {result}"
+                    )
+
+            # Add agent thought if present
+            thought = query.get("thought", "")
+            if thought:
+                conversation_lines.append(f"\nAgent Thought: {thought}")
+
+            # Add assistant response
+            conversation_lines.append(f"\nAssistant: {query.get('response', '')}")
+
+            # Add sources if present
+            sources = query.get("sources", [])
+            if sources:
+                conversation_lines.append(f"\nSources Used: {len(sources)} documents")
+
+            conversation_lines.append("")  # Empty line between messages
+
+        return "\n".join(conversation_lines)

application/api/answer/services/compression/service.py

@@ -0,0 +1,316 @@
+"""Core compression service with simplified responsibilities."""
+
+import logging
+import re
+from datetime import datetime, timezone
+from typing import Any, Dict, List, Optional
+
+from application.api.answer.services.compression.prompt_builder import (
+    CompressionPromptBuilder,
+)
+from application.api.answer.services.compression.token_counter import TokenCounter
+from application.api.answer.services.compression.types import (
+    CompressionMetadata,
+)
+from application.core.settings import settings
+
+logger = logging.getLogger(__name__)
+
+
+class CompressionService:
+    """
+    Service for compressing conversation history.
+
+    Handles DB updates.
+    """
+
+    def __init__(
+        self,
+        llm,
+        model_id: str,
+        conversation_service=None,
+        prompt_builder: Optional[CompressionPromptBuilder] = None,
+    ):
+        """
+        Initialize compression service.
+
+        Args:
+            llm: LLM instance to use for compression
+            model_id: Model ID for compression
+            conversation_service: Service for DB operations (optional, for DB updates)
+            prompt_builder: Custom prompt builder (optional)
+        """
+        self.llm = llm
+        self.model_id = model_id
+        self.conversation_service = conversation_service
+        self.prompt_builder = prompt_builder or CompressionPromptBuilder(
+            version=settings.COMPRESSION_PROMPT_VERSION
+        )
+
+    def compress_conversation(
+        self,
+        conversation: Dict[str, Any],
+        compress_up_to_index: int,
+    ) -> CompressionMetadata:
+        """
+        Compress conversation history up to specified index.
+
+        Args:
+            conversation: Full conversation document
+            compress_up_to_index: Last query index to include in compression
+
+        Returns:
+            CompressionMetadata with compression details
+
+        Raises:
+            ValueError: If compress_up_to_index is invalid
+        """
+        try:
+            queries = conversation.get("queries", [])
+
+            if compress_up_to_index < 0 or compress_up_to_index >= len(queries):
+                raise ValueError(
+                    f"Invalid compress_up_to_index: {compress_up_to_index} "
+                    f"(conversation has {len(queries)} queries)"
+                )
+
+            # Get queries to compress
+            queries_to_compress = queries[: compress_up_to_index + 1]
+
+            # Check if there are existing compressions. ``compression_metadata``
+            # is a nullable JSONB column, so a never-compressed conversation
+            # reads back as None; ``get(key, {})`` would return that None (the
+            # default only applies to absent keys), so coalesce with ``or {}``.
+            existing_compressions = (conversation.get("compression_metadata") or {}).get(
+                "compression_points", []
+            )
+
+            if existing_compressions:
+                logger.info(
+                    f"Found {len(existing_compressions)} previous compression(s) - "
+                    f"will incorporate into new summary"
+                )
+
+            # Calculate original token count
+            original_tokens = TokenCounter.count_query_tokens(queries_to_compress)
+
+            # Log tool call stats
+            self._log_tool_call_stats(queries_to_compress)
+
+            # Build compression prompt
+            messages = self.prompt_builder.build_prompt(
+                queries_to_compress, existing_compressions
+            )
+
+            # Call LLM to generate compression
+            logger.info(
+                f"Starting compression: {len(queries_to_compress)} queries "
+                f"(messages 0-{compress_up_to_index}, {original_tokens} tokens) "
+                f"using model {self.model_id}"
+            )
+
+            # See note in conversation_service.py: ``self.model_id`` is
+            # the registry id (UUID for BYOM); the LLM's own model_id is
+            # what the provider's API actually expects.
+            response = self.llm.gen(
+                model=getattr(self.llm, "model_id", None) or self.model_id,
+                messages=messages,
+                max_tokens=4000,
+            )
+
+            # Extract summary from response
+            compressed_summary = self._extract_summary(response)
+
+            # Calculate compressed token count
+            compressed_tokens = TokenCounter.count_message_tokens(
+                [{"content": compressed_summary}]
+            )
+
+            # Calculate compression ratio
+            compression_ratio = (
+                original_tokens / compressed_tokens if compressed_tokens > 0 else 0
+            )
+
+            logger.info(
+                f"Compression complete: {original_tokens} → {compressed_tokens} tokens "
+                f"({compression_ratio:.1f}x compression)"
+            )
+
+            # Build compression metadata
+            compression_metadata = CompressionMetadata(
+                timestamp=datetime.now(timezone.utc),
+                query_index=compress_up_to_index,
+                compressed_summary=compressed_summary,
+                original_token_count=original_tokens,
+                compressed_token_count=compressed_tokens,
+                compression_ratio=compression_ratio,
+                model_used=self.model_id,
+                compression_prompt_version=self.prompt_builder.version,
+            )
+
+            return compression_metadata
+
+        except Exception as e:
+            logger.error(f"Error compressing conversation: {str(e)}", exc_info=True)
+            raise
+
+    def compress_and_save(
+        self,
+        conversation_id: str,
+        conversation: Dict[str, Any],
+        compress_up_to_index: int,
+    ) -> CompressionMetadata:
+        """
+        Compress conversation and save to database.
+
+        Args:
+            conversation_id: Conversation ID
+            conversation: Full conversation document
+            compress_up_to_index: Last query index to include
+
+        Returns:
+            CompressionMetadata
+
+        Raises:
+            ValueError: If conversation_service not provided or invalid index
+        """
+        if not self.conversation_service:
+            raise ValueError(
+                "conversation_service required for compress_and_save operation"
+            )
+
+        # Perform compression
+        metadata = self.compress_conversation(conversation, compress_up_to_index)
+
+        # Save to database
+        self.conversation_service.update_compression_metadata(
+            conversation_id, metadata.to_dict()
+        )
+
+        logger.info(f"Compression metadata saved to database for {conversation_id}")
+
+        return metadata
+
+    def get_compressed_context(
+        self, conversation: Dict[str, Any]
+    ) -> tuple[Optional[str], List[Dict[str, Any]]]:
+        """
+        Get compressed summary + recent uncompressed messages.
+
+        Args:
+            conversation: Full conversation document
+
+        Returns:
+            (compressed_summary, recent_messages)
+        """
+        try:
+            # ``or {}`` guards against a NULL ``compression_metadata`` column
+            # (reads back as None), which would crash the ``.get`` calls below.
+            compression_metadata = conversation.get("compression_metadata") or {}
+
+            if not compression_metadata.get("is_compressed"):
+                logger.debug("No compression metadata found - using full history")
+                queries = conversation.get("queries", [])
+                if queries is None:
+                    logger.error("Conversation queries is None - returning empty list")
+                    return None, []
+                return None, queries
+
+            compression_points = compression_metadata.get("compression_points", [])
+
+            if not compression_points:
+                logger.debug("No compression points found - using full history")
+                queries = conversation.get("queries", [])
+                if queries is None:
+                    logger.error("Conversation queries is None - returning empty list")
+                    return None, []
+                return None, queries
+
+            # Get the most recent compression point
+            latest_compression = compression_points[-1]
+            compressed_summary = latest_compression.get("compressed_summary")
+            last_compressed_index = latest_compression.get("query_index")
+            compressed_tokens = latest_compression.get("compressed_token_count", 0)
+            original_tokens = latest_compression.get("original_token_count", 0)
+
+            # Get only messages after compression point
+            queries = conversation.get("queries", [])
+            total_queries = len(queries)
+            recent_queries = queries[last_compressed_index + 1 :]
+
+            logger.info(
+                f"Using compressed context: summary ({compressed_tokens} tokens, "
+                f"compressed from {original_tokens}) + {len(recent_queries)} recent messages "
+                f"(messages {last_compressed_index + 1}-{total_queries - 1})"
+            )
+
+            return compressed_summary, recent_queries
+
+        except Exception as e:
+            logger.error(
+                f"Error getting compressed context: {str(e)}", exc_info=True
+            )
+            queries = conversation.get("queries", [])
+            if queries is None:
+                return None, []
+            return None, queries
+
+    def _extract_summary(self, llm_response: str) -> str:
+        """
+        Extract clean summary from LLM response.
+
+        Args:
+            llm_response: Raw LLM response
+
+        Returns:
+            Cleaned summary text
+        """
+        try:
+            # Try to extract content within <summary> tags
+            summary_match = re.search(
+                r"<summary>(.*?)</summary>", llm_response, re.DOTALL
+            )
+
+            if summary_match:
+                summary = summary_match.group(1).strip()
+            else:
+                # If no summary tags, remove analysis tags and use the rest
+                summary = re.sub(
+                    r"<analysis>.*?</analysis>", "", llm_response, flags=re.DOTALL
+                ).strip()
+
+            return summary
+
+        except Exception as e:
+            logger.warning(f"Error extracting summary: {str(e)}, using full response")
+            return llm_response
+
+    def _log_tool_call_stats(self, queries: List[Dict[str, Any]]) -> None:
+        """Log statistics about tool calls in queries."""
+        total_tool_calls = 0
+        total_tool_result_chars = 0
+        tool_call_breakdown = {}
+
+        for q in queries:
+            for tc in q.get("tool_calls", []):
+                total_tool_calls += 1
+                tool_name = tc.get("tool_name", "unknown")
+                action_name = tc.get("action_name", "unknown")
+                key = f"{tool_name}.{action_name}"
+                tool_call_breakdown[key] = tool_call_breakdown.get(key, 0) + 1
+
+                # Track total tool result size
+                result = tc.get("result", "")
+                if result:
+                    total_tool_result_chars += len(str(result))
+
+        if total_tool_calls > 0:
+            tool_breakdown_str = ", ".join(
+                f"{tool}({count})"
+                for tool, count in sorted(tool_call_breakdown.items())
+            )
+            tool_result_kb = total_tool_result_chars / 1024
+            logger.info(
+                f"Tool call breakdown: {tool_breakdown_str} "
+                f"(total result size: {tool_result_kb:.1f} KB, {total_tool_result_chars:,} chars)"
+            )

application/api/answer/services/compression/threshold_checker.py

@@ -0,0 +1,110 @@
+"""Compression threshold checking logic."""
+
+import logging
+from typing import Any, Dict
+
+from application.core.model_utils import get_token_limit
+from application.core.settings import settings
+from application.api.answer.services.compression.token_counter import TokenCounter
+
+logger = logging.getLogger(__name__)
+
+
+class CompressionThresholdChecker:
+    """Determines if compression is needed based on token thresholds."""
+
+    def __init__(self, threshold_percentage: float = None):
+        """
+        Initialize threshold checker.
+
+        Args:
+            threshold_percentage: Percentage of context to use as threshold
+                                 (defaults to settings.COMPRESSION_THRESHOLD_PERCENTAGE)
+        """
+        self.threshold_percentage = (
+            threshold_percentage or settings.COMPRESSION_THRESHOLD_PERCENTAGE
+        )
+
+    def should_compress(
+        self,
+        conversation: Dict[str, Any],
+        model_id: str,
+        current_query_tokens: int = 500,
+        user_id: str | None = None,
+    ) -> bool:
+        """
+        Determine if compression is needed.
+
+        Args:
+            conversation: Full conversation document
+            model_id: Target model for this request
+            current_query_tokens: Estimated tokens for current query
+            user_id: Owner — needed so per-user BYOM custom-model UUIDs
+                resolve when looking up the context window.
+
+        Returns:
+            True if tokens >= threshold% of context window
+        """
+        try:
+            # Calculate total tokens in conversation
+            total_tokens = TokenCounter.count_conversation_tokens(conversation)
+            total_tokens += current_query_tokens
+
+            # Get context window limit for model
+            context_limit = get_token_limit(model_id, user_id=user_id)
+
+            # Calculate threshold
+            threshold = int(context_limit * self.threshold_percentage)
+
+            compression_needed = total_tokens >= threshold
+            percentage_used = (total_tokens / context_limit) * 100
+
+            if compression_needed:
+                logger.warning(
+                    f"COMPRESSION TRIGGERED: {total_tokens} tokens / {context_limit} limit "
+                    f"({percentage_used:.1f}% used, threshold: {self.threshold_percentage * 100:.0f}%)"
+                )
+            else:
+                logger.info(
+                    f"Compression check: {total_tokens}/{context_limit} tokens "
+                    f"({percentage_used:.1f}% used, threshold: {self.threshold_percentage * 100:.0f}%) - No compression needed"
+                )
+
+            return compression_needed
+
+        except Exception as e:
+            logger.error(f"Error checking compression need: {str(e)}", exc_info=True)
+            return False
+
+    def check_message_tokens(
+        self, messages: list, model_id: str, user_id: str | None = None
+    ) -> bool:
+        """
+        Check if message list exceeds threshold.
+
+        Args:
+            messages: List of message dicts
+            model_id: Target model
+            user_id: Owner — needed so per-user BYOM custom-model UUIDs
+                resolve when looking up the context window.
+
+        Returns:
+            True if at or above threshold
+        """
+        try:
+            current_tokens = TokenCounter.count_message_tokens(messages)
+            context_limit = get_token_limit(model_id, user_id=user_id)
+            threshold = int(context_limit * self.threshold_percentage)
+
+            if current_tokens >= threshold:
+                logger.warning(
+                    f"Message context limit approaching: {current_tokens}/{context_limit} tokens "
+                    f"({(current_tokens/context_limit)*100:.1f}%)"
+                )
+                return True
+
+            return False
+
+        except Exception as e:
+            logger.error(f"Error checking message tokens: {str(e)}", exc_info=True)
+            return False

application/api/answer/services/compression/token_counter.py

@@ -0,0 +1,133 @@
+"""Token counting utilities for compression."""
+
+import logging
+from typing import Any, Dict, List
+
+from application.utils import num_tokens_from_string
+from application.core.settings import settings
+
+logger = logging.getLogger(__name__)
+
+
+class TokenCounter:
+    """Centralized token counting for conversations and messages."""
+
+    # Per-image token estimate. Provider tokenizers vary widely
+    # (Gemini ~258, GPT-4o 85-1500, Claude ~1500) and the actual cost
+    # depends on resolution/detail we can't see here. Errs slightly high
+    # so the threshold check stays conservative.
+    _IMAGE_PART_TOKEN_ESTIMATE = 1500
+
+    @staticmethod
+    def count_message_tokens(messages: List[Dict]) -> int:
+        """
+        Calculate total tokens in a list of messages.
+
+        Args:
+            messages: List of message dicts with 'content' field
+
+        Returns:
+            Total token count
+        """
+        total_tokens = 0
+        for message in messages:
+            content = message.get("content", "")
+            if isinstance(content, str):
+                total_tokens += num_tokens_from_string(content)
+            elif isinstance(content, list):
+                # Handle structured content (tool calls, image parts, etc.)
+                for item in content:
+                    if isinstance(item, dict):
+                        total_tokens += TokenCounter._count_content_part(item)
+        return total_tokens
+
+    @staticmethod
+    def _count_content_part(item: Dict) -> int:
+        # Image/file attachments are billed by the provider per image,
+        # not proportional to the inline bytes/base64 string.
+        # ``str(item)`` on a 1MB image inflates the count by ~10000x,
+        # which trips spurious compression and overflows downstream
+        # input limits.
+        item_type = item.get("type")
+
+        if "files" in item:
+            files = item.get("files")
+            count = len(files) if isinstance(files, list) and files else 1
+            return TokenCounter._IMAGE_PART_TOKEN_ESTIMATE * count
+
+        if "image_url" in item or item_type in {
+            "image",
+            "image_url",
+            "input_image",
+            "file",
+        }:
+            return TokenCounter._IMAGE_PART_TOKEN_ESTIMATE
+
+        return num_tokens_from_string(str(item))
+
+    @staticmethod
+    def count_query_tokens(
+        queries: List[Dict[str, Any]], include_tool_calls: bool = True
+    ) -> int:
+        """
+        Count tokens across multiple query objects.
+
+        Args:
+            queries: List of query objects from conversation
+            include_tool_calls: Whether to count tool call tokens
+
+        Returns:
+            Total token count
+        """
+        total_tokens = 0
+
+        for query in queries:
+            # Count prompt and response tokens
+            if "prompt" in query:
+                total_tokens += num_tokens_from_string(query["prompt"])
+            if "response" in query:
+                total_tokens += num_tokens_from_string(query["response"])
+            if "thought" in query:
+                total_tokens += num_tokens_from_string(query.get("thought", ""))
+
+            # Count tool call tokens
+            if include_tool_calls and "tool_calls" in query:
+                for tool_call in query["tool_calls"]:
+                    tool_call_string = (
+                        f"Tool: {tool_call.get('tool_name')} | "
+                        f"Action: {tool_call.get('action_name')} | "
+                        f"Args: {tool_call.get('arguments')} | "
+                        f"Response: {tool_call.get('result')}"
+                    )
+                    total_tokens += num_tokens_from_string(tool_call_string)
+
+        return total_tokens
+
+    @staticmethod
+    def count_conversation_tokens(
+        conversation: Dict[str, Any], include_system_prompt: bool = False
+    ) -> int:
+        """
+        Calculate total tokens in a conversation.
+
+        Args:
+            conversation: Conversation document
+            include_system_prompt: Whether to include system prompt in count
+
+        Returns:
+            Total token count
+        """
+        try:
+            queries = conversation.get("queries", [])
+            total_tokens = TokenCounter.count_query_tokens(queries)
+
+            # Add system prompt tokens if requested
+            if include_system_prompt:
+                # Rough estimate for system prompt
+                total_tokens += settings.RESERVED_TOKENS.get("system_prompt", 500)
+
+            return total_tokens
+
+        except Exception as e:
+            logger.error(f"Error calculating conversation tokens: {str(e)}")
+            return 0

application/api/answer/services/compression/types.py

@@ -0,0 +1,91 @@
+"""Type definitions for compression module."""
+
+from dataclasses import dataclass, field
+from datetime import datetime
+from typing import Any, Dict, List, Optional
+
+
+@dataclass
+class CompressionMetadata:
+    """Metadata about a compression operation."""
+
+    timestamp: datetime
+    query_index: int
+    compressed_summary: str
+    original_token_count: int
+    compressed_token_count: int
+    compression_ratio: float
+    model_used: str
+    compression_prompt_version: str
+
+    def to_dict(self) -> Dict[str, Any]:
+        """Convert to dictionary for DB storage."""
+        return {
+            "timestamp": self.timestamp,
+            "query_index": self.query_index,
+            "compressed_summary": self.compressed_summary,
+            "original_token_count": self.original_token_count,
+            "compressed_token_count": self.compressed_token_count,
+            "compression_ratio": self.compression_ratio,
+            "model_used": self.model_used,
+            "compression_prompt_version": self.compression_prompt_version,
+        }
+
+
+@dataclass
+class CompressionResult:
+    """Result of a compression operation."""
+
+    success: bool
+    compressed_summary: Optional[str] = None
+    recent_queries: List[Dict[str, Any]] = field(default_factory=list)
+    metadata: Optional[CompressionMetadata] = None
+    error: Optional[str] = None
+    compression_performed: bool = False
+
+    @classmethod
+    def success_with_compression(
+        cls, summary: str, queries: List[Dict], metadata: CompressionMetadata
+    ) -> "CompressionResult":
+        """Create a successful result with compression."""
+        return cls(
+            success=True,
+            compressed_summary=summary,
+            recent_queries=queries,
+            metadata=metadata,
+            compression_performed=True,
+        )
+
+    @classmethod
+    def success_no_compression(cls, queries: List[Dict]) -> "CompressionResult":
+        """Create a successful result without compression needed."""
+        return cls(
+            success=True,
+            recent_queries=queries,
+            compression_performed=False,
+        )
+
+    @classmethod
+    def failure(cls, error: str) -> "CompressionResult":
+        """Create a failure result."""
+        return cls(success=False, error=error, compression_performed=False)
+
+    def as_history(self) -> List[Dict[str, str]]:
+        """
+        Convert recent queries to history format.
+
+        Returns:
+            List of prompt/response dicts (with thought when present so
+            DeepSeek-style providers can re-attach reasoning_content on
+            replay).
+        """
+        out: List[Dict[str, str]] = []
+        for q in self.recent_queries:
+            entry: Dict[str, str] = {
+                "prompt": q["prompt"],
+                "response": q["response"],
+            }
+            if q.get("thought"):
+                entry["thought"] = q["thought"]
+            out.append(entry)
+        return out

application/api/answer/services/continuation_service.py

@@ -0,0 +1,163 @@
+"""Service for saving and restoring tool-call continuation state.
+
+When a stream pauses (tool needs approval or client-side execution),
+the full execution state is persisted to Postgres so the client can
+resume later by sending tool_actions.
+"""
+
+import logging
+from typing import Any, Dict, List, Optional
+
+from application.storage.db.base_repository import looks_like_uuid
+from application.storage.db.repositories.conversations import ConversationsRepository
+from application.storage.db.repositories.pending_tool_state import (
+    PendingToolStateRepository,
+)
+from application.storage.db.serialization import coerce_pg_native as _make_serializable
+from application.storage.db.session import db_readonly, db_session
+
+logger = logging.getLogger(__name__)
+
+# TTL for pending states — auto-cleaned after this period
+PENDING_STATE_TTL_SECONDS = 30 * 60  # 30 minutes
+
+# Re-export so the existing tests at tests/api/answer/services/test_continuation_service_pg.py
+# can keep importing ``_make_serializable`` from here.
+__all__ = ["_make_serializable", "ContinuationService", "PENDING_STATE_TTL_SECONDS"]
+
+
+class ContinuationService:
+    """Manages pending tool-call state in Postgres."""
+
+    def __init__(self):
+        # No-op constructor retained for call-site compatibility. State
+        # lives in Postgres now; each operation opens its own short-lived
+        # session rather than holding a connection on the service.
+        pass
+
+    def save_state(
+        self,
+        conversation_id: str,
+        user: str,
+        messages: List[Dict],
+        pending_tool_calls: List[Dict],
+        tools_dict: Dict,
+        tool_schemas: List[Dict],
+        agent_config: Dict,
+        client_tools: Optional[List[Dict]] = None,
+    ) -> str:
+        """Save execution state for later continuation.
+
+        ``conversation_id`` may be a Postgres UUID or the legacy Mongo
+        ``ObjectId`` string — the latter is resolved via
+        ``conversations.legacy_mongo_id`` to find the matching row.
+
+        Args:
+            conversation_id: The conversation this state belongs to.
+            user: Owner user ID.
+            messages: Full messages array at the pause point.
+            pending_tool_calls: Tool calls awaiting client action.
+            tools_dict: Serializable tools configuration dict.
+            tool_schemas: LLM-formatted tool schemas (agent.tools).
+            agent_config: Config needed to recreate the agent on resume.
+            client_tools: Client-provided tool schemas for client-side execution.
+
+        Returns:
+            The string ID (conversation_id as provided) of the saved state.
+        """
+        with db_session() as conn:
+            conv = ConversationsRepository(conn).get_by_legacy_id(conversation_id)
+            if conv is not None:
+                pg_conv_id = conv["id"]
+            elif looks_like_uuid(conversation_id):
+                pg_conv_id = conversation_id
+            else:
+                # Unresolvable legacy ObjectId — downstream ``CAST AS uuid``
+                # would raise and poison the save. Surface the mismatch so
+                # the caller can decide (the stream loop in routes/base.py
+                # already wraps this in try/except).
+                raise ValueError(
+                    f"Cannot save continuation state: conversation_id "
+                    f"{conversation_id!r} is neither a PG UUID nor a "
+                    f"backfilled legacy Mongo id."
+                )
+            PendingToolStateRepository(conn).save_state(
+                pg_conv_id,
+                user,
+                messages=_make_serializable(messages),
+                pending_tool_calls=_make_serializable(pending_tool_calls),
+                tools_dict=_make_serializable(tools_dict),
+                tool_schemas=_make_serializable(tool_schemas),
+                agent_config=_make_serializable(agent_config),
+                client_tools=_make_serializable(client_tools) if client_tools else None,
+            )
+
+        logger.info(
+            f"Saved continuation state for conversation {conversation_id} "
+            f"with {len(pending_tool_calls)} pending tool call(s)"
+        )
+        return conversation_id
+
+    def load_state(
+        self, conversation_id: str, user: str
+    ) -> Optional[Dict[str, Any]]:
+        """Load pending continuation state.
+
+        Returns:
+            The state dict, or None if no pending state exists.
+        """
+        with db_readonly() as conn:
+            conv = ConversationsRepository(conn).get_by_legacy_id(conversation_id)
+            if conv is not None:
+                pg_conv_id = conv["id"]
+            elif looks_like_uuid(conversation_id):
+                pg_conv_id = conversation_id
+            else:
+                # Unresolvable legacy ObjectId → no state can exist for it.
+                return None
+            doc = PendingToolStateRepository(conn).load_state(pg_conv_id, user)
+        if not doc:
+            return None
+        return doc
+
+    def delete_state(self, conversation_id: str, user: str) -> bool:
+        """Delete pending state after successful resumption.
+
+        Returns:
+            True if a row was deleted.
+        """
+        with db_session() as conn:
+            conv = ConversationsRepository(conn).get_by_legacy_id(conversation_id)
+            if conv is not None:
+                pg_conv_id = conv["id"]
+            elif looks_like_uuid(conversation_id):
+                pg_conv_id = conversation_id
+            else:
+                # Unresolvable legacy ObjectId → nothing to delete.
+                return False
+            deleted = PendingToolStateRepository(conn).delete_state(pg_conv_id, user)
+        if deleted:
+            logger.info(
+                f"Deleted continuation state for conversation {conversation_id}"
+            )
+        return deleted
+
+    def mark_resuming(self, conversation_id: str, user: str) -> bool:
+        """Flip the pending row to ``resuming`` so a crashed resume can be retried."""
+        with db_session() as conn:
+            conv = ConversationsRepository(conn).get_by_legacy_id(conversation_id)
+            if conv is not None:
+                pg_conv_id = conv["id"]
+            elif looks_like_uuid(conversation_id):
+                pg_conv_id = conversation_id
+            else:
+                return False
+            flipped = PendingToolStateRepository(conn).mark_resuming(
+                pg_conv_id, user
+            )
+        if flipped:
+            logger.info(
+                f"Marked continuation state as resuming for conversation "
+                f"{conversation_id}"
+            )
+        return flipped

application/api/answer/services/conversation_service.py

@@ -0,0 +1,546 @@
+"""Conversation persistence service backed by Postgres.
+
+Handles create / append / update / compression for conversations during
+the answer-streaming path. Connections are opened per-operation rather
+than held for the duration of a stream.
+"""
+
+import logging
+import uuid
+from datetime import datetime, timezone
+from typing import Any, Dict, List, Optional
+
+from sqlalchemy import text as sql_text
+
+from application.core.settings import settings
+from application.storage.db.base_repository import looks_like_uuid
+from application.storage.db.repositories.agents import AgentsRepository
+from application.storage.db.repositories.conversations import (
+    ConversationsRepository,
+    MessageUpdateOutcome,
+)
+from application.storage.db.session import db_readonly, db_session
+
+
+logger = logging.getLogger(__name__)
+
+
+# Shown to the user if the worker dies mid-stream and the response is never finalised.
+TERMINATED_RESPONSE_PLACEHOLDER = (
+    "Response was terminated prior to completion, try regenerating."
+)
+
+
+class ConversationService:
+    def get_conversation(
+        self, conversation_id: str, user_id: str
+    ) -> Optional[Dict[str, Any]]:
+        """Retrieve a conversation with owner-or-shared access control.
+
+        Returns a dict in the legacy Mongo shape — ``queries`` is a list
+        of message dicts (prompt/response/...) — for compatibility with
+        the streaming pipeline that consumes this shape.
+        """
+        if not conversation_id or not user_id:
+            return None
+        try:
+            with db_readonly() as conn:
+                repo = ConversationsRepository(conn)
+                conv = repo.get_any(conversation_id, user_id)
+                if conv is None:
+                    logger.warning(
+                        f"Conversation not found or unauthorized - ID: {conversation_id}, User: {user_id}"
+                    )
+                    return None
+                messages = repo.get_messages(str(conv["id"]))
+            conv["queries"] = messages
+            conv["_id"] = str(conv["id"])
+            return conv
+        except Exception as e:
+            logger.error(f"Error fetching conversation: {str(e)}", exc_info=True)
+            return None
+
+    def save_conversation(
+        self,
+        conversation_id: Optional[str],
+        question: str,
+        response: str,
+        thought: str,
+        sources: List[Dict[str, Any]],
+        tool_calls: List[Dict[str, Any]],
+        llm: Any,
+        model_id: str,
+        decoded_token: Dict[str, Any],
+        index: Optional[int] = None,
+        api_key: Optional[str] = None,
+        agent_id: Optional[str] = None,
+        is_shared_usage: bool = False,
+        shared_token: Optional[str] = None,
+        attachment_ids: Optional[List[str]] = None,
+        metadata: Optional[Dict[str, Any]] = None,
+        visibility: str = "hidden",
+    ) -> str:
+        """Save or update a conversation in Postgres.
+
+        Returns the string conversation id (PG UUID as string, or the
+        caller-provided id if it was already a UUID).
+        """
+        if decoded_token is None:
+            raise ValueError("Invalid or missing authentication token")
+        user_id = decoded_token.get("sub")
+        if not user_id:
+            raise ValueError("User ID not found in token")
+        current_time = datetime.now(timezone.utc)
+
+        # Trim huge inline source text to a reasonable max before persist.
+        for source in sources:
+            if "text" in source and isinstance(source["text"], str):
+                source["text"] = source["text"][:1000]
+
+        message_payload = {
+            "prompt": question,
+            "response": response,
+            "thought": thought,
+            "sources": sources,
+            "tool_calls": tool_calls,
+            "attachments": attachment_ids,
+            "model_id": model_id,
+            "timestamp": current_time,
+        }
+        if metadata:
+            message_payload["metadata"] = metadata
+
+        if conversation_id is not None and index is not None:
+            with db_session() as conn:
+                repo = ConversationsRepository(conn)
+                conv = repo.get_any(conversation_id, user_id)
+                if conv is None:
+                    raise ValueError("Conversation not found or unauthorized")
+                conv_pg_id = str(conv["id"])
+                repo.update_message_at(conv_pg_id, index, message_payload)
+                repo.truncate_after(conv_pg_id, index)
+            return conversation_id
+        elif conversation_id:
+            with db_session() as conn:
+                repo = ConversationsRepository(conn)
+                conv = repo.get_any(conversation_id, user_id)
+                if conv is None:
+                    raise ValueError("Conversation not found or unauthorized")
+                conv_pg_id = str(conv["id"])
+                # append_message expects 'metadata' key either way; normalise.
+                append_payload = dict(message_payload)
+                append_payload.setdefault("metadata", metadata or {})
+                repo.append_message(conv_pg_id, append_payload)
+            return conversation_id
+        else:
+            messages_summary = [
+                {
+                    "role": "system",
+                    "content": "You are a helpful assistant that creates concise conversation titles. "
+                    "Summarize conversations in 3 words or less using the same language as the user.",
+                },
+                {
+                    "role": "user",
+                    "content": "Summarise following conversation in no more than 3 words, "
+                    "respond ONLY with the summary, use the same language as the "
+                    "user query \n\nUser: " + question + "\n\n" + "AI: " + response,
+                },
+            ]
+
+            # ``model_id`` here is the registry id (a UUID for BYOM
+            # records). The LLM's own ``model_id`` is the upstream name
+            # LLMCreator resolved at construction time — that's what
+            # the provider's API expects. Built-ins are unaffected.
+            completion = llm.gen(
+                model=getattr(llm, "model_id", None) or model_id,
+                messages=messages_summary,
+                # Reasoning-capable default models spend the whole budget inside
+                # reasoning_content before emitting any title, so 500 came back
+                # empty (finish_reason=length). Give enough room to finish
+                # thinking and still produce the 3-word title; non-reasoning
+                # models stop far short of this cap.
+                max_tokens=2000,
+            )
+
+            if not completion or not completion.strip():
+                completion = question[:50] if question else "New Conversation"
+
+            resolved_api_key: Optional[str] = None
+            resolved_agent_id: Optional[str] = None
+            if api_key:
+                with db_readonly() as conn:
+                    agent = AgentsRepository(conn).find_by_key(api_key)
+                if agent:
+                    resolved_api_key = agent.get("key")
+                if agent_id:
+                    resolved_agent_id = agent_id
+
+            with db_session() as conn:
+                repo = ConversationsRepository(conn)
+                conv = repo.create(
+                    user_id,
+                    completion,
+                    agent_id=resolved_agent_id,
+                    api_key=resolved_api_key,
+                    is_shared_usage=bool(resolved_agent_id and is_shared_usage),
+                    shared_token=(
+                        shared_token
+                        if (resolved_agent_id and is_shared_usage)
+                        else None
+                    ),
+                    visibility=visibility,
+                )
+                conv_pg_id = str(conv["id"])
+                append_payload = dict(message_payload)
+                append_payload.setdefault("metadata", metadata or {})
+                repo.append_message(conv_pg_id, append_payload)
+            return conv_pg_id
+
+    def save_user_question(
+        self,
+        conversation_id: Optional[str],
+        question: str,
+        decoded_token: Dict[str, Any],
+        *,
+        attachment_ids: Optional[List[str]] = None,
+        api_key: Optional[str] = None,
+        agent_id: Optional[str] = None,
+        is_shared_usage: bool = False,
+        shared_token: Optional[str] = None,
+        model_id: Optional[str] = None,
+        request_id: Optional[str] = None,
+        visibility: str = "hidden",
+        status: str = "pending",
+        index: Optional[int] = None,
+    ) -> Dict[str, str]:
+        """Reserve the placeholder message row before the LLM call.
+
+        ``index`` triggers regenerate semantics: messages at
+        ``position >= index`` are truncated so the new placeholder
+        lands at ``position = index`` rather than appending.
+
+        Returns ``{"conversation_id", "message_id", "request_id"}``.
+        """
+        if decoded_token is None:
+            raise ValueError("Invalid or missing authentication token")
+        user_id = decoded_token.get("sub")
+        if not user_id:
+            raise ValueError("User ID not found in token")
+
+        request_id = request_id or str(uuid.uuid4())
+
+        resolved_api_key: Optional[str] = None
+        resolved_agent_id: Optional[str] = None
+        if api_key and not conversation_id:
+            with db_readonly() as conn:
+                agent = AgentsRepository(conn).find_by_key(api_key)
+            if agent:
+                resolved_api_key = agent.get("key")
+            if agent_id:
+                resolved_agent_id = agent_id
+
+        with db_session() as conn:
+            repo = ConversationsRepository(conn)
+            if conversation_id:
+                conv = repo.get_any(conversation_id, user_id)
+                if conv is None:
+                    raise ValueError("Conversation not found or unauthorized")
+                conv_pg_id = str(conv["id"])
+                # Regenerate / edit-prior-question: drop the message at
+                # ``index`` and everything after it so the new
+                # ``reserve_message`` lands at ``position=index`` rather
+                # than appending at the end of the conversation.
+                if isinstance(index, int) and index >= 0:
+                    repo.truncate_after(conv_pg_id, keep_up_to=index - 1)
+            else:
+                fallback_name = (question[:50] if question else "New Conversation")
+                conv = repo.create(
+                    user_id,
+                    fallback_name,
+                    agent_id=resolved_agent_id,
+                    api_key=resolved_api_key,
+                    is_shared_usage=bool(resolved_agent_id and is_shared_usage),
+                    shared_token=(
+                        shared_token
+                        if (resolved_agent_id and is_shared_usage)
+                        else None
+                    ),
+                    visibility=visibility,
+                )
+                conv_pg_id = str(conv["id"])
+
+            row = repo.reserve_message(
+                conv_pg_id,
+                prompt=question,
+                placeholder_response=TERMINATED_RESPONSE_PLACEHOLDER,
+                request_id=request_id,
+                status=status,
+                attachments=attachment_ids,
+                model_id=model_id,
+            )
+            message_id = str(row["id"])
+
+        return {
+            "conversation_id": conv_pg_id,
+            "message_id": message_id,
+            "request_id": request_id,
+        }
+
+    def update_message_status(self, message_id: str, status: str) -> bool:
+        """Cheap status-only transition (e.g. ``pending → streaming``)."""
+        if not message_id:
+            return False
+        with db_session() as conn:
+            return ConversationsRepository(conn).update_message_status(
+                message_id, status,
+            )
+
+    def heartbeat_message(self, message_id: str) -> bool:
+        """Bump ``message_metadata.last_heartbeat_at`` so the reconciler's
+        staleness sweep counts the row as alive. No-ops on terminal rows.
+        """
+        if not message_id:
+            return False
+        with db_session() as conn:
+            return ConversationsRepository(conn).heartbeat_message(message_id)
+
+    def finalize_message(
+        self,
+        message_id: str,
+        response: str,
+        *,
+        thought: str = "",
+        sources: Optional[List[Dict[str, Any]]] = None,
+        tool_calls: Optional[List[Dict[str, Any]]] = None,
+        model_id: Optional[str] = None,
+        metadata: Optional[Dict[str, Any]] = None,
+        status: str = "complete",
+        error: Optional[BaseException] = None,
+        title_inputs: Optional[Dict[str, Any]] = None,
+    ) -> MessageUpdateOutcome:
+        """Commit the response and tool_call confirms in one transaction.
+
+        The outcome propagates directly from ``update_message_by_id`` so
+        callers (notably the SSE abort handler) can tell a fresh
+        finalize from "the row was already terminal" — the latter must
+        still be treated as success when the prior state was
+        ``complete``.
+        """
+        if not message_id:
+            return MessageUpdateOutcome.INVALID
+        sources = sources or []
+        for source in sources:
+            if "text" in source and isinstance(source["text"], str):
+                source["text"] = source["text"][:1000]
+
+        merged_metadata: Dict[str, Any] = dict(metadata or {})
+        if status == "failed" and error is not None:
+            merged_metadata.setdefault(
+                "error", f"{type(error).__name__}: {str(error)}"
+            )
+
+        update_fields: Dict[str, Any] = {
+            "response": response,
+            "status": status,
+            "thought": thought,
+            "sources": sources,
+            "tool_calls": tool_calls or [],
+            "metadata": merged_metadata,
+        }
+        if model_id is not None:
+            update_fields["model_id"] = model_id
+
+        # Atomic message update + tool_call_attempts confirm; the
+        # ``only_if_non_terminal`` guard prevents a late stream from
+        # retracting a row the reconciler already escalated.
+        with db_session() as conn:
+            repo = ConversationsRepository(conn)
+            outcome = repo.update_message_by_id(
+                message_id, update_fields,
+                only_if_non_terminal=True,
+            )
+            if outcome is not MessageUpdateOutcome.UPDATED:
+                logger.warning(
+                    f"finalize_message: no row updated for message_id={message_id} "
+                    f"(outcome={outcome.value} — possibly already terminal)"
+                )
+                return outcome
+            repo.confirm_executed_tool_calls(message_id)
+
+        # Outside the txn — title-gen is a multi-second LLM round trip.
+        if title_inputs and status == "complete":
+            try:
+                with db_session() as conn:
+                    self._maybe_generate_title(conn, message_id, title_inputs)
+            except Exception as e:
+                logger.error(
+                    f"finalize_message title generation failed: {e}",
+                    exc_info=True,
+                )
+        return MessageUpdateOutcome.UPDATED
+
+    def _maybe_generate_title(
+        self,
+        conn,
+        message_id: str,
+        title_inputs: Dict[str, Any],
+    ) -> None:
+        """Generate an LLM-summarised conversation name if one isn't set yet."""
+        llm = title_inputs.get("llm")
+        question = title_inputs.get("question") or ""
+        response = title_inputs.get("response") or ""
+        fallback_name = title_inputs.get("fallback_name") or question[:50]
+        if llm is None:
+            return
+
+        row = conn.execute(
+            sql_text(
+                "SELECT c.id, c.name FROM conversation_messages m "
+                "JOIN conversations c ON c.id = m.conversation_id "
+                "WHERE m.id = CAST(:mid AS uuid)"
+            ),
+            {"mid": message_id},
+        ).fetchone()
+        if row is None:
+            return
+        conv_id, current_name = str(row[0]), row[1]
+        if current_name and current_name != fallback_name:
+            return
+
+        messages_summary = [
+            {
+                "role": "system",
+                "content": "You are a helpful assistant that creates concise conversation titles. "
+                "Summarize conversations in 3 words or less using the same language as the user.",
+            },
+            {
+                "role": "user",
+                "content": "Summarise following conversation in no more than 3 words, "
+                "respond ONLY with the summary, use the same language as the "
+                "user query \n\nUser: " + question + "\n\n" + "AI: " + response,
+            },
+        ]
+        completion = llm.gen(
+            model=getattr(llm, "model_id", None) or title_inputs.get("model_id"),
+            messages=messages_summary,
+            # Reasoning-capable default models spend the whole budget inside
+            # reasoning_content before emitting any title, so 500 came back empty
+            # (finish_reason=length). Give room to finish and still produce the
+            # 3-word title; non-reasoning models stop far short of this cap.
+            max_tokens=2000,
+        )
+        if not completion or not completion.strip():
+            completion = fallback_name or "New Conversation"
+        conn.execute(
+            sql_text(
+                "UPDATE conversations SET name = :name, updated_at = now() "
+                "WHERE id = CAST(:id AS uuid)"
+            ),
+            {"id": conv_id, "name": completion.strip()},
+        )
+
+    def update_compression_metadata(
+        self, conversation_id: str, compression_metadata: Dict[str, Any]
+    ) -> None:
+        """Persist compression flags and append a compression point.
+
+        Mirrors the Mongo-era ``$set`` + ``$push $slice`` on
+        ``compression_metadata`` but goes through the PG repo API.
+        """
+        try:
+            with db_session() as conn:
+                repo = ConversationsRepository(conn)
+                # conversation_id here comes from the streaming pipeline
+                # which has already resolved it; accept either UUID or
+                # legacy id for safety.
+                conv = repo.get_by_legacy_id(conversation_id)
+                conv_pg_id = (
+                    str(conv["id"]) if conv is not None else conversation_id
+                )
+                repo.set_compression_flags(
+                    conv_pg_id,
+                    is_compressed=True,
+                    last_compression_at=compression_metadata.get("timestamp"),
+                )
+                repo.append_compression_point(
+                    conv_pg_id,
+                    compression_metadata,
+                    max_points=settings.COMPRESSION_MAX_HISTORY_POINTS,
+                )
+            logger.info(
+                f"Updated compression metadata for conversation {conversation_id}"
+            )
+        except Exception as e:
+            logger.error(
+                f"Error updating compression metadata: {str(e)}", exc_info=True
+            )
+            raise
+
+    def append_compression_message(
+        self, conversation_id: str, compression_metadata: Dict[str, Any]
+    ) -> None:
+        """Append a synthetic compression summary message to the conversation."""
+        try:
+            summary = compression_metadata.get("compressed_summary", "")
+            if not summary:
+                return
+            timestamp = compression_metadata.get(
+                "timestamp", datetime.now(timezone.utc)
+            )
+
+            with db_session() as conn:
+                repo = ConversationsRepository(conn)
+                conv = repo.get_by_legacy_id(conversation_id)
+                conv_pg_id = (
+                    str(conv["id"]) if conv is not None else conversation_id
+                )
+                repo.append_message(conv_pg_id, {
+                    "prompt": "[Context Compression Summary]",
+                    "response": summary,
+                    "thought": "",
+                    "sources": [],
+                    "tool_calls": [],
+                    "attachments": [],
+                    "model_id": compression_metadata.get("model_used"),
+                    "timestamp": timestamp,
+                })
+            logger.info(
+                f"Appended compression summary to conversation {conversation_id}"
+            )
+        except Exception as e:
+            logger.error(
+                f"Error appending compression summary: {str(e)}", exc_info=True
+            )
+
+    def get_compression_metadata(
+        self, conversation_id: str
+    ) -> Optional[Dict[str, Any]]:
+        """Fetch the stored compression metadata JSONB blob for a conversation."""
+        try:
+            with db_readonly() as conn:
+                repo = ConversationsRepository(conn)
+                conv = repo.get_by_legacy_id(conversation_id)
+                if conv is None:
+                    # Fallback to UUID lookup without user scoping — the
+                    # caller already holds an authenticated conversation
+                    # id from the streaming path. Gate on id shape so a
+                    # non-UUID (legacy ObjectId that wasn't backfilled)
+                    # doesn't reach CAST — the cast raises and spams the
+                    # logs with a stack trace on every call.
+                    if not looks_like_uuid(conversation_id):
+                        return None
+                    result = conn.execute(
+                        sql_text(
+                            "SELECT compression_metadata FROM conversations "
+                            "WHERE id = CAST(:id AS uuid)"
+                        ),
+                        {"id": conversation_id},
+                    )
+                    row = result.fetchone()
+                    return row[0] if row is not None else None
+            return conv.get("compression_metadata") if conv else None
+        except Exception as e:
+            logger.error(
+                f"Error getting compression metadata: {str(e)}", exc_info=True
+            )
+            return None

application/api/answer/services/persistence_policy.py

@@ -0,0 +1,44 @@
+"""Resolve whether an answer is persisted and whether it lists in the sidebar.
+
+Persistence (is a row written at all?) and visibility (does it show in the
+owner's sidebar?) are separate decisions. Conversations persist by default
+everywhere, and visibility defaults to ``hidden`` for every caller: only an
+explicit request-level ``visibility: "listed"`` — which the first-party UI
+sends on normal chats — puts a conversation in the owner's sidebar. The
+legacy ``save_conversation`` flag no longer affects either decision, so
+API/OpenAI-compatible clients that still send it (its old meaning was
+"persist this conversation") can't list rows into the agent owner's sidebar.
+"""
+
+from typing import Any, Optional, Tuple
+
+VISIBILITY_LISTED = "listed"
+VISIBILITY_HIDDEN = "hidden"
+
+
+def resolve_persistence(
+    *,
+    visibility_flag: Optional[Any] = None,
+    persist_flag: Optional[bool] = None,
+) -> Tuple[bool, str]:
+    """Resolve ``(should_persist, visibility)`` for an answer request.
+
+    Args:
+        visibility_flag: Request-level ``visibility`` value. Only the exact
+            string ``"listed"`` opts the conversation into the owner's
+            sidebar; anything else (including ``None``) stays hidden.
+        persist_flag: Explicit persistence opt-out (``False`` to skip writing
+            a row, e.g. stateless tool rounds that would orphan one). ``None``
+            keeps the always-persist default.
+
+    Returns:
+        ``(should_persist, visibility)`` where ``visibility`` is
+        ``"listed"`` or ``"hidden"``.
+    """
+    should_persist = True if persist_flag is None else bool(persist_flag)
+    visibility = (
+        VISIBILITY_LISTED
+        if visibility_flag == VISIBILITY_LISTED
+        else VISIBILITY_HIDDEN
+    )
+    return should_persist, visibility

application/api/answer/services/prompt_renderer.py

@@ -0,0 +1,118 @@
+import logging
+from typing import Any, Dict, Optional
+
+from application.templates.namespaces import NamespaceManager
+
+from application.templates.template_engine import TemplateEngine, TemplateRenderError
+
+logger = logging.getLogger(__name__)
+
+
+def format_docs_for_prompt(docs: Optional[list]) -> Optional[str]:
+    """Format retrieved chunks as XML-tagged documents for prompt injection.
+
+    Each chunk is wrapped in a ``<document index="n">`` block with a
+    ``<source>`` subtag (when a filename/title is known) so the model can
+    tell chunks apart and cite them by name.
+    """
+    if not docs:
+        return None
+    parts = []
+    for i, doc in enumerate(docs, start=1):
+        source = doc.get("filename") or doc.get("title") or doc.get("source")
+        lines = [f'<document index="{i}">']
+        if source:
+            lines.append(f"<source>{source}</source>")
+        lines.append(f"<content>\n{doc.get('text', '')}\n</content>")
+        lines.append("</document>")
+        parts.append("\n".join(lines))
+    return "\n\n".join(parts)
+
+
+class PromptRenderer:
+    """Service for rendering prompts with dynamic context using namespaces"""
+
+    def __init__(self):
+        self.template_engine = TemplateEngine()
+        self.namespace_manager = NamespaceManager()
+
+    def render_prompt(
+        self,
+        prompt_content: str,
+        user_id: Optional[str] = None,
+        request_id: Optional[str] = None,
+        passthrough_data: Optional[Dict[str, Any]] = None,
+        docs: Optional[list] = None,
+        docs_together: Optional[str] = None,
+        tools_data: Optional[Dict[str, Any]] = None,
+        **kwargs,
+    ) -> str:
+        """
+        Render prompt with full context from all namespaces.
+
+        Args:
+            prompt_content: Raw prompt template string
+            user_id: Current user identifier
+            request_id: Unique request identifier
+            passthrough_data: Parameters from web request
+            docs: RAG retrieved documents
+            docs_together: Concatenated document content
+            tools_data: Pre-fetched tool results organized by tool name
+            **kwargs: Additional parameters for namespace builders
+
+        Returns:
+            Rendered prompt string with all variables substituted
+
+        Raises:
+            TemplateRenderError: If template rendering fails
+        """
+        if not prompt_content:
+            return ""
+
+        uses_template = self._uses_template_syntax(prompt_content)
+
+        if not uses_template:
+            return self._apply_legacy_substitutions(prompt_content, docs_together)
+
+        try:
+            context = self.namespace_manager.build_context(
+                user_id=user_id,
+                request_id=request_id,
+                passthrough_data=passthrough_data,
+                docs=docs,
+                docs_together=docs_together,
+                tools_data=tools_data,
+                **kwargs,
+            )
+
+            return self.template_engine.render(prompt_content, context)
+        except TemplateRenderError:
+            raise
+        except Exception as e:
+            error_msg = f"Prompt rendering failed: {str(e)}"
+            logger.error(error_msg)
+            raise TemplateRenderError(error_msg) from e
+
+    def _uses_template_syntax(self, prompt_content: str) -> bool:
+        """Check if prompt uses Jinja2 template syntax"""
+        return "{{" in prompt_content and "}}" in prompt_content
+
+    def _apply_legacy_substitutions(
+        self, prompt_content: str, docs_together: Optional[str] = None
+    ) -> str:
+        """
+        Apply backward-compatible substitutions for old prompt format.
+
+        Handles the legacy {summaries} placeholder. When no documents were
+        retrieved the placeholder is removed so the model never sees the
+        raw template artifact.
+        """
+        return prompt_content.replace("{summaries}", docs_together or "")
+
+    def validate_template(self, prompt_content: str) -> bool:
+        """Validate prompt template syntax"""
+        return self.template_engine.validate_template(prompt_content)
+
+    def extract_variables(self, prompt_content: str) -> set[str]:
+        """Extract all variable names from prompt template"""
+        return self.template_engine.extract_variables(prompt_content)