claude-code-tools 0.2.0__py3-none-any.whl → 1.4.1__py3-none-any.whl

claude_code_tools/search_index.py

@@ -0,0 +1,1321 @@
+"""Tantivy-based search index for session exports."""
+
+import json
+import math
+import shutil
+import sys
+from dataclasses import dataclass
+from datetime import datetime
+from importlib.metadata import version as get_pkg_version
+from pathlib import Path
+from typing import Any, Optional
+
+from claude_code_tools.session_utils import is_valid_session
+
+
+def _get_package_version() -> str:
+    """Get installed package version for automatic index rebuilding."""
+    try:
+        return get_pkg_version("claude-code-tools")
+    except Exception:
+        return "unknown"
+
+# Lazy imports to allow module to load even if deps not installed
+try:
+    import yaml
+    YAML_AVAILABLE = True
+except ImportError:
+    yaml = None  # type: ignore
+    YAML_AVAILABLE = False
+
+try:
+    import tantivy
+    TANTIVY_AVAILABLE = True
+except ImportError:
+    tantivy = None  # type: ignore
+    TANTIVY_AVAILABLE = False
+
+
+def _require_deps():
+    """Raise helpful error if required dependencies are not installed."""
+    missing = []
+    if not YAML_AVAILABLE:
+        missing.append("pyyaml")
+    if not TANTIVY_AVAILABLE:
+        missing.append("tantivy")
+
+    if missing:
+        raise ImportError(
+            f"Missing dependencies for search indexing: {', '.join(missing)}\n"
+            f"Install with: pip install {' '.join(missing)}\n"
+            "Or reinstall claude-code-tools: uv tool install claude-code-tools"
+        )
+
+
+def _iso_to_epoch_ms(iso_str: str) -> int:
+    """Convert ISO timestamp string to epoch milliseconds.
+
+    Args:
+        iso_str: ISO format timestamp (e.g., "2025-12-08T18:22:59.582Z")
+
+    Returns:
+        Epoch time in milliseconds, or 0 if parsing fails
+    """
+    if not iso_str:
+        return 0
+    try:
+        # Handle both Z suffix and +00:00 timezone formats
+        if "T" in iso_str:
+            dt = datetime.fromisoformat(iso_str.replace("Z", "+00:00"))
+        else:
+            dt = datetime.fromisoformat(iso_str)
+        return int(dt.timestamp() * 1000)
+    except (ValueError, TypeError):
+        return 0
+
+
+@dataclass
+class SearchResult:
+    """A search result with metadata and snippet."""
+
+    session_id: str
+    agent: str
+    project: str
+    branch: Optional[str]
+    cwd: Optional[str]
+    modified: str
+    lines: int
+    snippet: str
+    score: float
+    export_path: str
+
+
+class IndexState:
+    """Track which files have been indexed to enable incremental updates."""
+
+    def __init__(self, state_path: Path):
+        self.state_path = state_path
+        self.indexed_files: dict[str, dict[str, Any]] = {}
+        self._load()
+
+    def _load(self):
+        """Load state from disk."""
+        if self.state_path.exists():
+            try:
+                with open(self.state_path, "r") as f:
+                    self.indexed_files = json.load(f)
+            except (json.JSONDecodeError, IOError):
+                self.indexed_files = {}
+
+    def save(self):
+        """Save state to disk."""
+        self.state_path.parent.mkdir(parents=True, exist_ok=True)
+        with open(self.state_path, "w") as f:
+            json.dump(self.indexed_files, f)
+
+    def needs_reindex(self, file_path: Path) -> bool:
+        """Check if file needs to be (re-)indexed."""
+        key = str(file_path)
+        if key not in self.indexed_files:
+            return True
+
+        stat = file_path.stat()
+        stored = self.indexed_files[key]
+
+        # Reindex if mtime or size changed
+        return (
+            stat.st_mtime != stored.get("mtime")
+            or stat.st_size != stored.get("size")
+        )
+
+    def mark_indexed(self, file_path: Path):
+        """Mark file as indexed with current metadata."""
+        stat = file_path.stat()
+        self.indexed_files[str(file_path)] = {
+            "mtime": stat.st_mtime,
+            "size": stat.st_size,
+        }
+
+
+class SessionIndex:
+    """Tantivy-based full-text search index for sessions."""
+
+    def __init__(self, index_path: Path):
+        """
+        Initialize or open a session index.
+
+        Args:
+            index_path: Directory for the Tantivy index
+
+        Raises:
+            ImportError: If tantivy or pyyaml is not installed
+        """
+        _require_deps()
+
+        self.index_path = Path(index_path)
+        self.state = IndexState(self.index_path / "index_state.json")
+
+        # Define schema
+        self.schema_builder = tantivy.SchemaBuilder()
+
+        # Stored fields (retrieved but not searched)
+        self.schema_builder.add_text_field("session_id", stored=True)
+        self.schema_builder.add_text_field("agent", stored=True)
+        self.schema_builder.add_text_field("project", stored=True)
+        self.schema_builder.add_text_field("branch", stored=True)
+        self.schema_builder.add_text_field("cwd", stored=True)
+        self.schema_builder.add_text_field("created", stored=True)
+        self.schema_builder.add_text_field("modified", stored=True)
+        # Fast field for sorting by modification time (epoch milliseconds)
+        self.schema_builder.add_unsigned_field("modified_ts", stored=True, fast=True)
+        self.schema_builder.add_integer_field("lines", stored=True)
+        # export_path needs "raw" tokenizer for exact match deletion
+        self.schema_builder.add_text_field("export_path", stored=True, tokenizer_name="raw")
+
+        # First and last message fields (for preview in TUI)
+        self.schema_builder.add_text_field("first_msg_role", stored=True)
+        self.schema_builder.add_text_field("first_msg_content", stored=True)
+        self.schema_builder.add_text_field("last_msg_role", stored=True)
+        self.schema_builder.add_text_field("last_msg_content", stored=True)
+
+        # Session type fields (for filtering in TUI)
+        self.schema_builder.add_text_field("derivation_type", stored=True)
+        self.schema_builder.add_text_field("is_sidechain", stored=True)  # "true"/"false"
+
+        # Claude home field (for filtering by source Claude home directory)
+        # Use "raw" tokenizer so paths are indexed as single tokens for exact matching
+        self.schema_builder.add_text_field("claude_home", stored=True, tokenizer_name="raw")
+
+        # Custom title field (from /rename command)
+        self.schema_builder.add_text_field("custom_title", stored=True)
+
+        # Searchable content field
+        self.schema_builder.add_text_field("content", stored=True)
+
+        self.schema = self.schema_builder.build()
+
+        # Create or open index
+        self.index_path.mkdir(parents=True, exist_ok=True)
+
+        # Check index version - rebuild if package version changed
+        version_file = self.index_path / "VERSION"
+        current_version = _get_package_version()
+        needs_rebuild = False
+
+        if version_file.exists():
+            try:
+                stored_version = version_file.read_text().strip()
+                if stored_version != current_version:
+                    needs_rebuild = True
+                    print(
+                        f"Package version changed ({stored_version} -> {current_version}), "
+                        "rebuilding index..."
+                    )
+            except IOError:
+                needs_rebuild = True
+                print("Index version file corrupted, rebuilding...")
+        elif (self.index_path / "meta.json").exists():
+            # Index exists but no VERSION file - legacy index, rebuild
+            needs_rebuild = True
+            print("Upgrading index to versioned format, rebuilding...")
+
+        if needs_rebuild:
+            # Clear index directory contents (but keep the directory itself)
+            for item in self.index_path.iterdir():
+                if item.is_file():
+                    item.unlink()
+                elif item.is_dir():
+                    shutil.rmtree(item)
+            # Reset state tracker
+            self.state = IndexState(self.index_path / "index_state.json")
+
+        # Create or open index
+        self.index = tantivy.Index(self.schema, path=str(self.index_path))
+
+        # Write current version
+        version_file.write_text(current_version)
+
+    def _parse_export_file(self, export_path: Path) -> Optional[dict[str, Any]]:
+        """
+        Parse an exported session file with YAML front matter.
+
+        Returns:
+            Dict with metadata and content, or None if parsing fails
+        """
+        try:
+            content = export_path.read_text(encoding="utf-8")
+
+            if not content.startswith("---\n"):
+                return None
+
+            end_idx = content.find("\n---\n", 4)
+            if end_idx == -1:
+                return None
+
+            yaml_str = content[4:end_idx]
+            metadata = yaml.safe_load(yaml_str)
+
+            conversation = content[end_idx + 5:]
+
+            return {
+                "metadata": metadata,
+                "content": conversation,
+                "export_path": str(export_path),
+            }
+        except Exception:
+            return None
+
+    def build_from_exports(
+        self, exports_dir: Path, incremental: bool = True
+    ) -> dict[str, int]:
+        """
+        Build or update index from exported session files.
+
+        Args:
+            exports_dir: Directory containing exported .txt files
+                (expects subdirs: claude/, codex/)
+            incremental: If True, only index new/modified files
+
+        Returns:
+            Stats dict: {indexed, skipped, failed}
+        """
+        stats = {"indexed": 0, "skipped": 0, "failed": 0}
+
+        writer = self.index.writer()
+
+        # Find all export files
+        export_files = []
+        for subdir in ["claude", "codex"]:
+            subdir_path = exports_dir / subdir
+            if subdir_path.exists():
+                export_files.extend(subdir_path.glob("*.txt"))
+
+        for export_path in export_files:
+            # Check if needs indexing
+            if incremental and not self.state.needs_reindex(export_path):
+                stats["skipped"] += 1
+                continue
+
+            # Parse export file
+            parsed = self._parse_export_file(export_path)
+            if parsed is None:
+                stats["failed"] += 1
+                continue
+
+            metadata = parsed["metadata"]
+
+            # Create document
+            doc = tantivy.Document()
+            doc.add_text("session_id", metadata.get("session_id", ""))
+            doc.add_text("agent", metadata.get("agent", ""))
+            doc.add_text("project", metadata.get("project", ""))
+            doc.add_text("branch", metadata.get("branch", "") or "")
+            doc.add_text("cwd", metadata.get("cwd", "") or "")
+            doc.add_text("created", metadata.get("created", "") or "")
+            modified_str = metadata.get("modified", "")
+            doc.add_text("modified", modified_str)
+            doc.add_unsigned("modified_ts", _iso_to_epoch_ms(modified_str))
+            doc.add_integer("lines", metadata.get("lines", 0))
+            doc.add_text("export_path", parsed["export_path"])
+
+            # First and last message fields
+            first_msg = metadata.get("first_msg", {}) or {}
+            last_msg = metadata.get("last_msg", {}) or {}
+            doc.add_text("first_msg_role", first_msg.get("role", ""))
+            doc.add_text("first_msg_content", first_msg.get("content", ""))
+            doc.add_text("last_msg_role", last_msg.get("role", ""))
+            doc.add_text("last_msg_content", last_msg.get("content", ""))
+
+            # Session type fields
+            doc.add_text("derivation_type", metadata.get("derivation_type", "") or "")
+            doc.add_text(
+                "is_sidechain",
+                "true" if metadata.get("is_sidechain") else "false"
+            )
+
+            # Custom title (from /rename command)
+            doc.add_text("custom_title", metadata.get("customTitle", "") or "")
+
+            doc.add_text("content", parsed["content"])
+
+            writer.add_document(doc)
+            self.state.mark_indexed(export_path)
+            stats["indexed"] += 1
+
+        writer.commit()
+        self.state.save()
+        self.index.reload()
+
+        return stats
+
+    def index_single_file(
+        self, export_path: Path, writer, incremental: bool = True
+    ) -> dict:
+        """
+        Index a single export file.
+
+        Args:
+            export_path: Path to exported .txt file
+            writer: Tantivy index writer
+            incremental: If True, skip if already indexed
+
+        Returns:
+            Dict with result: {status, error}
+            status: 'indexed', 'skipped', 'failed'
+        """
+        result: dict = {"status": "failed", "error": None}
+
+        if not export_path.exists():
+            result["error"] = "file not found"
+            return result
+
+        # Check if needs indexing
+        if incremental and not self.state.needs_reindex(export_path):
+            result["status"] = "skipped"
+            return result
+
+        # Parse export file
+        try:
+            parsed = self._parse_export_file(export_path)
+        except Exception as e:
+            result["error"] = f"parse error: {e}"
+            return result
+
+        if parsed is None:
+            result["error"] = "failed to parse YAML frontmatter"
+            return result
+
+        metadata = parsed["metadata"]
+
+        try:
+            # Create document
+            doc = tantivy.Document()
+            doc.add_text("session_id", metadata.get("session_id", ""))
+            doc.add_text("agent", metadata.get("agent", ""))
+            doc.add_text("project", metadata.get("project", ""))
+            doc.add_text("branch", metadata.get("branch", "") or "")
+            doc.add_text("cwd", metadata.get("cwd", "") or "")
+            doc.add_text("created", metadata.get("created", "") or "")
+            modified_str = metadata.get("modified", "")
+            doc.add_text("modified", modified_str)
+            doc.add_unsigned("modified_ts", _iso_to_epoch_ms(modified_str))
+            doc.add_integer("lines", metadata.get("lines", 0))
+            doc.add_text("export_path", parsed["export_path"])
+
+            # First and last message fields
+            first_msg = metadata.get("first_msg", {}) or {}
+            last_msg = metadata.get("last_msg", {}) or {}
+            doc.add_text("first_msg_role", first_msg.get("role", ""))
+            doc.add_text("first_msg_content", first_msg.get("content", ""))
+            doc.add_text("last_msg_role", last_msg.get("role", ""))
+            doc.add_text("last_msg_content", last_msg.get("content", ""))
+
+            # Session type fields
+            doc.add_text(
+                "derivation_type", metadata.get("derivation_type", "") or ""
+            )
+            doc.add_text(
+                "is_sidechain",
+                "true" if metadata.get("is_sidechain") else "false"
+            )
+
+            # Custom title (from /rename command)
+            doc.add_text("custom_title", metadata.get("customTitle", "") or "")
+
+            doc.add_text("content", parsed["content"])
+
+            writer.add_document(doc)
+            self.state.mark_indexed(export_path)
+            result["status"] = "indexed"
+        except Exception as e:
+            result["error"] = f"index error: {e}"
+
+        return result
+
+    def get_writer(self):
+        """Get a writer for batch indexing."""
+        return self.index.writer()
+
+    def commit_and_reload(self, writer):
+        """Commit writer changes and reload index."""
+        writer.commit()
+        self.state.save()
+        self.index.reload()
+
+    def build_from_files(
+        self, export_files: list[Path], incremental: bool = True
+    ) -> dict[str, int]:
+        """
+        Build or update index from a list of exported session files.
+
+        This is the per-project alternative to build_from_exports().
+        Instead of expecting a single directory, it accepts a list of
+        export file paths from multiple project directories.
+
+        Args:
+            export_files: List of paths to exported .txt files
+            incremental: If True, only index new/modified files
+
+        Returns:
+            Stats dict: {indexed, skipped, failed}
+        """
+        stats = {"indexed": 0, "skipped": 0, "failed": 0}
+
+        writer = self.get_writer()
+
+        for export_path in export_files:
+            result = self.index_single_file(export_path, writer, incremental)
+            if result["status"] == "indexed":
+                stats["indexed"] += 1
+            elif result["status"] == "skipped":
+                stats["skipped"] += 1
+            else:
+                stats["failed"] += 1
+
+        self.commit_and_reload(writer)
+
+        return stats
+
+    def _extract_session_content(
+        self, jsonl_path: Path, agent: str
+    ) -> tuple[str, int, str]:
+        """
+        Extract searchable content from a session file.
+
+        Handles both Claude and Codex JSONL formats.
+
+        Args:
+            jsonl_path: Path to session JSONL file
+            agent: Agent type ('claude' or 'codex')
+
+        Returns:
+            Tuple of (content_string, user_message_count, custom_title)
+        """
+        messages = []
+        user_count = 0  # Count only user messages for the "lines" metric
+        custom_title = ""  # Session name from /rename command
+
+        try:
+            with open(jsonl_path, "r", encoding="utf-8") as f:
+                for line in f:
+                    line = line.strip()
+                    if not line:
+                        continue
+
+                    try:
+                        data = json.loads(line)
+                    except json.JSONDecodeError:
+                        continue
+
+                    # Check for custom-title (from /rename command)
+                    if data.get("type") == "custom-title":
+                        custom_title = data.get("customTitle", "")
+                        continue
+
+                    role: Optional[str] = None
+                    text: str = ""
+
+                    if agent == "claude":
+                        # Claude format: type is "user" or "assistant"
+                        msg_type = data.get("type")
+                        if msg_type not in ("user", "assistant"):
+                            continue
+
+                        role = msg_type
+                        message = data.get("message", {})
+                        content = message.get("content")
+
+                        # Count user messages, but exclude tool results
+                        # Tool results have content as list with {"type": "tool_result"}
+                        # Real user messages have content as string or list with text
+                        if role == "user":
+                            is_tool_result = (
+                                isinstance(content, list)
+                                and len(content) > 0
+                                and isinstance(content[0], dict)
+                                and content[0].get("type") == "tool_result"
+                            )
+                            if not is_tool_result:
+                                user_count += 1
+
+                        if not content:
+                            continue
+
+                        # Extract text content
+                        if isinstance(content, str):
+                            text = content
+                        elif isinstance(content, list):
+                            for block in content:
+                                if isinstance(block, str):
+                                    text += block + "\n"
+                                elif isinstance(block, dict):
+                                    if block.get("type") == "text":
+                                        text += block.get("text", "") + "\n"
+                                    elif block.get("type") == "tool_use":
+                                        tool_name = block.get("name", "")
+                                        text += f"[Tool: {tool_name}]\n"
+                                        # Index tool input content
+                                        tool_input = block.get("input", {})
+                                        if isinstance(tool_input, dict):
+                                            for value in tool_input.values():
+                                                if isinstance(value, str) and value:
+                                                    text += f"{value}\n"
+                                    elif block.get("type") == "tool_result":
+                                        result = block.get("content", "")
+                                        if isinstance(result, str):
+                                            text += result[:500] + "\n"
+
+                    elif agent == "codex":
+                        # Codex format: type is "response_item" with payload
+                        if data.get("type") != "response_item":
+                            continue
+
+                        payload = data.get("payload", {})
+                        if payload.get("type") != "message":
+                            continue
+
+                        role = payload.get("role")
+                        if role == "user":
+                            user_count += 1
+                        content = payload.get("content", [])
+
+                        if not isinstance(content, list):
+                            continue
+
+                        for block in content:
+                            if not isinstance(block, dict):
+                                continue
+                            block_type = block.get("type")
+                            # input_text (user) and output_text (assistant)
+                            if block_type in ("input_text", "output_text"):
+                                text += block.get("text", "") + "\n"
+
+                    if role and text.strip():
+                        messages.append(f"[{role}] {text.strip()}")
+
+        except (OSError, IOError):
+            pass
+
+        return "\n\n".join(messages), user_count, custom_title
+
+    def _parse_jsonl_session(self, jsonl_path: Path) -> Optional[dict[str, Any]]:
+        """
+        Parse a JSONL session file directly (no export step).
+
+        Uses extract_session_metadata() from export_session.py for metadata,
+        and handles both Claude and Codex JSONL formats for content.
+
+        Args:
+            jsonl_path: Path to the session JSONL file
+
+        Returns:
+            Dict with metadata and content suitable for indexing, or None on failure
+        """
+        try:
+            # Detect agent from path
+            path_str = str(jsonl_path)
+            agent = "codex" if ".codex" in path_str else "claude"
+
+            # Use existing helper for metadata extraction
+            from claude_code_tools.export_session import extract_session_metadata
+            metadata = extract_session_metadata(jsonl_path, agent)
+
+            # Extract content for full-text search (also extracts custom_title)
+            content, msg_count, custom_title = self._extract_session_content(
+                jsonl_path, agent
+            )
+
+            # Add custom_title to metadata (extracted during content scan)
+            if custom_title:
+                metadata["customTitle"] = custom_title
+
+            if msg_count == 0:
+                return {"_skip_reason": "empty"}
+
+            # Map metadata fields to expected format
+            first_msg = metadata.get("first_msg") or {"role": "", "content": ""}
+            last_msg = metadata.get("last_msg") or {"role": "", "content": ""}
+
+            # Always use filename-derived session_id (the canonical identifier)
+            # Internal sessionId field can be stale in forked sessions
+            # Extract UUID from filename: last 36 chars of stem (handles both
+            # "uuid.jsonl" and "rollout-timestamp-uuid.jsonl" formats)
+            stem = jsonl_path.stem
+            session_id = stem[-36:] if len(stem) >= 36 else stem
+
+            return {
+                "metadata": {
+                    "session_id": session_id,
+                    "agent": agent,
+                    "project": metadata.get("project", ""),
+                    "branch": metadata.get("branch", "") or "",
+                    "cwd": metadata.get("cwd", "") or "",
+                    "created": metadata.get("created", "") or "",
+                    "modified": metadata.get("modified", "") or "",
+                    "is_sidechain": metadata.get("is_sidechain", False),
+                    "derivation_type": metadata.get("derivation_type", "") or "",
+                    "session_type": metadata.get("session_type"),
+                    "customTitle": metadata.get("customTitle", "") or "",
+                },
+                "content": content,
+                "first_msg": first_msg,
+                "last_msg": last_msg,
+                "lines": msg_count,
+                "file_path": str(jsonl_path),
+            }
+        except Exception as e:
+            return {"_skip_reason": "parse_error", "_error": str(e)}
+
+    def index_from_jsonl(
+        self,
+        jsonl_files: list[Path],
+        incremental: bool = True,
+        claude_home: Optional[Path] = None,
+        codex_home: Optional[Path] = None,
+        show_progress: bool = False,
+    ) -> dict[str, int]:
+        """
+        Build or update index directly from JSONL session files.
+
+        This is the "zero-config" approach - no export step needed.
+        Parses JSONL directly and indexes to Tantivy.
+
+        Args:
+            jsonl_files: List of paths to session JSONL files
+            incremental: If True, only index new/modified files
+            claude_home: Claude home directory (stored for filtering)
+            codex_home: Codex home directory (stored for filtering)
+            show_progress: If True, show tqdm progress bar
+
+        Returns:
+            Stats dict: {indexed, skipped, failed}
+        """
+        claude_home_str = str(claude_home) if claude_home else ""
+        codex_home_str = str(codex_home) if codex_home else str(Path.home() / ".codex")
+        stats = {
+            "indexed": 0, "skipped": 0, "failed": 0,
+            "empty": 0, "parse_error": 0, "index_error": 0,
+        }
+
+        writer = self.get_writer()
+
+        # Wrap iterator with tqdm if progress requested
+        file_iter: Any = jsonl_files
+        if show_progress:
+            try:
+                from tqdm import tqdm
+                file_iter = tqdm(
+                    jsonl_files,
+                    desc="Indexing sessions",
+                    unit="file",
+                    file=sys.stderr,
+                )
+            except ImportError:
+                print("Note: Install tqdm for progress bar (pip install tqdm)",
+                      file=sys.stderr)
+
+        for jsonl_path in file_iter:
+            # Check if needs indexing
+            if incremental and not self.state.needs_reindex(jsonl_path):
+                stats["skipped"] += 1
+                continue
+
+            # Skip invalid sessions (metadata-only files like file-history-snapshot)
+            if not is_valid_session(jsonl_path):
+                stats["skipped"] += 1
+                # Mark as indexed so we don't re-check next time
+                self.state.mark_indexed(jsonl_path)
+                continue
+
+            # Parse JSONL file
+            parsed = self._parse_jsonl_session(jsonl_path)
+            if parsed is None or "_skip_reason" in parsed:
+                stats["failed"] += 1
+                if parsed:
+                    reason = parsed.get("_skip_reason", "unknown")
+                    if reason == "empty":
+                        stats["empty"] += 1
+                    elif reason == "parse_error":
+                        stats["parse_error"] += 1
+                continue
+
+            metadata = parsed["metadata"]
+            first_msg = parsed["first_msg"]
+            last_msg = parsed["last_msg"]
+
+            # Skip helper sessions (SDK/headless sessions used for analysis)
+            if metadata.get("session_type") == "helper":
+                stats["skipped"] += 1
+                # Mark as indexed so we don't re-check next time
+                self.state.mark_indexed(jsonl_path)
+                continue
+
+            # Skip sessions run from inside claude_home or codex_home directories
+            cwd = metadata.get("cwd", "") or ""
+            if cwd and (
+                (claude_home_str and cwd.startswith(claude_home_str))
+                or (codex_home_str and cwd.startswith(codex_home_str))
+            ):
+                stats["skipped"] += 1
+                continue
+
+            try:
+                session_id = metadata.get("session_id", "")
+                file_path_str = str(jsonl_path)
+
+                # Delete by file path (unique), not sessionId (shared by sub-agents)
+                writer.delete_documents("export_path", file_path_str)
+
+                # Create document
+                doc = tantivy.Document()
+                doc.add_text("session_id", session_id)
+                doc.add_text("agent", metadata.get("agent", ""))
+                doc.add_text("project", metadata.get("project", ""))
+                doc.add_text("branch", metadata.get("branch", "") or "")
+                doc.add_text("cwd", metadata.get("cwd", "") or "")
+                doc.add_text("created", metadata.get("created", "") or "")
+                modified_str = metadata.get("modified", "")
+                doc.add_text("modified", modified_str)
+                doc.add_unsigned("modified_ts", _iso_to_epoch_ms(modified_str))
+                doc.add_integer("lines", parsed.get("lines", 0))
+                doc.add_text("export_path", parsed["file_path"])  # Store JSONL path
+
+                # First and last message fields
+                doc.add_text("first_msg_role", first_msg.get("role", ""))
+                doc.add_text("first_msg_content", first_msg.get("content", ""))
+                doc.add_text("last_msg_role", last_msg.get("role", ""))
+                doc.add_text("last_msg_content", last_msg.get("content", ""))
+
+                # Session type fields
+                doc.add_text(
+                    "derivation_type", metadata.get("derivation_type", "") or ""
+                )
+                doc.add_text(
+                    "is_sidechain",
+                    "true" if metadata.get("is_sidechain") else "false"
+                )
+
+                # Custom title (from /rename command)
+                doc.add_text("custom_title", metadata.get("customTitle", "") or "")
+
+                # Source home (for filtering by source directory)
+                # Detect from path whether this is a Claude or Codex session
+                agent = metadata.get("agent", "")
+                if agent == "codex" or ".codex" in file_path_str:
+                    # Codex session - store codex home
+                    doc.add_text("claude_home", codex_home_str)
+                else:
+                    # Claude session - store claude home
+                    doc.add_text("claude_home", claude_home_str)
+
+                doc.add_text("content", parsed["content"])
+
+                writer.add_document(doc)
+                self.state.mark_indexed(jsonl_path)
+                stats["indexed"] += 1
+            except Exception:
+                stats["failed"] += 1
+                stats["index_error"] += 1
+
+        self.commit_and_reload(writer)
+
+        return stats
+
+    def prune_deleted(self) -> int:
+        """
+        Remove deleted files from the index.
+
+        Checks all files tracked in the index state and removes any that
+        no longer exist on disk. This handles the case where sessions are
+        deleted outside of the search tool.
+
+        Returns:
+            Number of entries pruned from the index.
+        """
+        if not self.state.indexed_files:
+            return 0
+
+        # First pass: identify deleted files
+        files_to_remove = [
+            file_path for file_path in self.state.indexed_files
+            if not Path(file_path).exists()
+        ]
+
+        if not files_to_remove:
+            return 0
+
+        # Only create writer if we have deletions
+        writer = self.get_writer()
+
+        for file_path in files_to_remove:
+            # Delete from Tantivy index
+            writer.delete_documents("export_path", file_path)
+            # Remove from state tracking
+            del self.state.indexed_files[file_path]
+
+        self.commit_and_reload(writer)
+
+        return len(files_to_remove)
+
+    def _generate_snippet(
+        self, content: str, query: str, max_len: int = 200
+    ) -> str:
+        """
+        Generate a snippet centered around the first match.
+
+        Args:
+            content: Full content text
+            query: Search query
+            max_len: Maximum snippet length
+
+        Returns:
+            Snippet with match context
+        """
+        if not query:
+            # For empty query, return start of content
+            return content[:max_len].strip() + "..." if len(content) > max_len else content.strip()
+
+        lower_content = content.lower()
+        lower_query = query.lower()
+
+        # Find first occurrence
+        pos = lower_content.find(lower_query)
+        if pos == -1:
+            # Try individual words
+            for word in query.split():
+                pos = lower_content.find(word.lower())
+                if pos != -1:
+                    break
+
+        if pos == -1:
+            return content[:max_len].strip() + "..."
+
+        # Center snippet around match
+        half_len = max_len // 2
+        start = max(0, pos - half_len)
+        end = min(len(content), pos + len(query) + half_len)
+
+        snippet = content[start:end].strip()
+
+        # Add ellipses
+        if start > 0:
+            snippet = "..." + snippet
+        if end < len(content):
+            snippet = snippet + "..."
+
+        # Clean up newlines
+        snippet = " ".join(snippet.split())
+
+        return snippet
+
+    def _calculate_recency_score(
+        self, modified: str, base_score: float, half_life_days: float = 7.0
+    ) -> float:
+        """
+        Apply recency boost using exponential decay.
+
+        Args:
+            modified: ISO format timestamp
+            base_score: Original Tantivy score
+            half_life_days: Days for score to decay by half
+
+        Returns:
+            Score with recency boost (up to 2x for very recent)
+        """
+        try:
+            # Parse modified timestamp
+            if "T" in modified:
+                mod_time = datetime.fromisoformat(modified.replace("Z", "+00:00"))
+            else:
+                mod_time = datetime.fromisoformat(modified)
+
+            now = datetime.now(mod_time.tzinfo) if mod_time.tzinfo else datetime.now()
+            age_seconds = (now - mod_time).total_seconds()
+
+            half_life_secs = half_life_days * 24 * 3600
+            recency_boost = 1.0 + math.exp(-age_seconds / half_life_secs)
+
+            return base_score * recency_boost
+        except Exception:
+            return base_score
+
+    def search(
+        self,
+        query: str,
+        limit: int = 50,
+        project: Optional[str] = None,
+    ) -> list[dict[str, Any]]:
+        """
+        Search the index.
+
+        Args:
+            query: Search query (empty returns recent sessions)
+            limit: Maximum results
+            project: Filter to specific project
+
+        Returns:
+            List of result dicts with metadata and snippets
+        """
+        if not query and not project:
+            return self.get_recent(limit=limit)
+
+        searcher = self.index.searcher()
+
+        # Build query
+        if query:
+            # Use Index.parse_query with default fields
+            parsed_query = self.index.parse_query(query, ["content"])
+        else:
+            # Match all for project filter
+            parsed_query = self.index.parse_query("*", ["content"])
+
+        # Execute search
+        results = []
+        top_docs = searcher.search(parsed_query, limit * 2)  # Fetch extra for filtering
+
+        for score, doc_address in top_docs.hits:
+            doc = searcher.doc(doc_address)
+
+            # Extract fields
+            session_id = doc.get_first("session_id")
+            agent = doc.get_first("agent")
+            doc_project = doc.get_first("project")
+            branch = doc.get_first("branch")
+            cwd = doc.get_first("cwd")
+            created = doc.get_first("created")
+            modified = doc.get_first("modified")
+            lines = doc.get_first("lines")
+            export_path = doc.get_first("export_path")
+            content = doc.get_first("content")
+
+            # Apply project filter
+            if project and doc_project != project:
+                continue
+
+            # Skip if session file no longer exists (deleted outside of search)
+            if export_path and not Path(export_path).exists():
+                continue
+
+            # Generate snippet
+            snippet = self._generate_snippet(content, query)
+
+            # Apply recency scoring
+            final_score = self._calculate_recency_score(modified, score)
+
+            # Extract first/last message fields
+            first_msg_role = doc.get_first("first_msg_role") or ""
+            first_msg_content = doc.get_first("first_msg_content") or ""
+            last_msg_role = doc.get_first("last_msg_role") or ""
+            last_msg_content = doc.get_first("last_msg_content") or ""
+
+            results.append({
+                "session_id": session_id,
+                "agent": agent,
+                "project": doc_project,
+                "branch": branch,
+                "cwd": cwd,
+                "created": created,
+                "modified": modified,
+                "lines": lines,
+                "export_path": export_path,
+                "snippet": snippet,
+                "score": final_score,
+                "first_msg_role": first_msg_role,
+                "first_msg_content": first_msg_content,
+                "last_msg_role": last_msg_role,
+                "last_msg_content": last_msg_content,
+            })
+
+            if len(results) >= limit:
+                break
+
+        # Sort by score (recency-adjusted)
+        results.sort(key=lambda x: x["score"], reverse=True)
+
+        return results
+
+    def get_recent(
+        self, limit: int = 20, project: Optional[str] = None
+    ) -> list[dict[str, Any]]:
+        """
+        Get recent sessions sorted by modification time.
+
+        Args:
+            limit: Maximum results
+            project: Filter to specific project
+
+        Returns:
+            List of result dicts sorted by recency
+        """
+        searcher = self.index.searcher()
+
+        # Match all documents
+        all_query = tantivy.Query.all_query()
+        top_docs = searcher.search(all_query, limit * 3)  # Fetch extra for filtering
+
+        results = []
+        for _, doc_address in top_docs.hits:
+            doc = searcher.doc(doc_address)
+
+            doc_project = doc.get_first("project")
+            export_path = doc.get_first("export_path")
+
+            # Apply project filter
+            if project and doc_project != project:
+                continue
+
+            # Skip if session file no longer exists (deleted outside of search)
+            if export_path and not Path(export_path).exists():
+                continue
+
+            modified = doc.get_first("modified")
+            content = doc.get_first("content")
+
+            results.append({
+                "session_id": doc.get_first("session_id"),
+                "agent": doc.get_first("agent"),
+                "project": doc_project,
+                "branch": doc.get_first("branch"),
+                "cwd": doc.get_first("cwd"),
+                "created": doc.get_first("created"),
+                "modified": modified,
+                "lines": doc.get_first("lines"),
+                "export_path": export_path,
+                "snippet": self._generate_snippet(content, ""),
+                "score": 0.0,
+                "first_msg_role": doc.get_first("first_msg_role") or "",
+                "first_msg_content": doc.get_first("first_msg_content") or "",
+                "last_msg_role": doc.get_first("last_msg_role") or "",
+                "last_msg_content": doc.get_first("last_msg_content") or "",
+                "claude_home": doc.get_first("claude_home") or "",
+            })
+
+        # Sort by modified timestamp (most recent first)
+        results.sort(key=lambda x: x["modified"] or "", reverse=True)
+
+        return results[:limit]
+
+    def get_latest_session(
+        self,
+        cwd: Optional[str] = None,
+        branch: Optional[str] = None,
+        agent: Optional[str] = None,
+        include_sub_agents: bool = False,
+    ) -> Optional[dict[str, Any]]:
+        """
+        Get the most recent session matching the given criteria.
+
+        Uses Tantivy's native sorting by modified_ts fast field for efficiency.
+
+        Args:
+            cwd: Filter to sessions from this working directory
+            branch: Filter to sessions on this git branch
+            agent: Filter to specific agent ("claude" or "codex")
+            include_sub_agents: If False (default), exclude sub-agent sessions
+
+        Returns:
+            Dict with session info, or None if no matching session found
+        """
+        searcher = self.index.searcher()
+
+        # Build query string for filtering
+        # Use parse_query with field:value syntax for exact matches
+        query_terms = []
+
+        if agent:
+            query_terms.append(f'agent:"{agent}"')
+
+        if cwd:
+            # Escape special chars in cwd path
+            escaped_cwd = cwd.replace('"', '\\"')
+            query_terms.append(f'cwd:"{escaped_cwd}"')
+
+        if branch:
+            query_terms.append(f'branch:"{branch}"')
+
+        # Combine with AND, or use match-all if no filters
+        if query_terms:
+            query_str = " AND ".join(query_terms)
+            query = self.index.parse_query(query_str, ["content"])
+        else:
+            query = tantivy.Query.all_query()
+
+        # Search with native sorting by modified_ts descending (most recent first)
+        # Fetch enough to find valid ones after filtering deleted files and sub-agents
+        top_docs = searcher.search(
+            query,
+            limit=100,
+            order_by_field="modified_ts",
+            order=tantivy.Order.Desc,
+        )
+
+        # Iterate through results (already sorted by recency) and return first valid
+        for _, doc_address in top_docs.hits:
+            doc = searcher.doc(doc_address)
+            export_path = doc.get_first("export_path")
+
+            # Skip if session file no longer exists
+            if export_path and not Path(export_path).exists():
+                continue
+
+            # Filter out sub-agents unless requested
+            if not include_sub_agents:
+                if export_path and (
+                    "/agent-" in export_path or
+                    export_path.split("/")[-1].startswith("agent-")
+                ):
+                    continue
+
+            # Found a valid session - return it
+            modified = doc.get_first("modified")
+            content = doc.get_first("content")
+
+            return {
+                "session_id": doc.get_first("session_id"),
+                "agent": doc.get_first("agent"),
+                "project": doc.get_first("project"),
+                "branch": doc.get_first("branch"),
+                "cwd": doc.get_first("cwd"),
+                "created": doc.get_first("created"),
+                "modified": modified,
+                "lines": doc.get_first("lines"),
+                "export_path": export_path,
+                "snippet": self._generate_snippet(content, "") if content else "",
+                "score": 0.0,
+                "first_msg_role": doc.get_first("first_msg_role") or "",
+                "first_msg_content": doc.get_first("first_msg_content") or "",
+                "last_msg_role": doc.get_first("last_msg_role") or "",
+                "last_msg_content": doc.get_first("last_msg_content") or "",
+                "derivation_type": doc.get_first("derivation_type") or "",
+                "is_sidechain": doc.get_first("is_sidechain") or "false",
+            }
+
+        return None
+
+
+def get_latest_session_from_index(
+    cwd: Optional[str] = None,
+    branch: Optional[str] = None,
+    agent: Optional[str] = None,
+    index_path: Optional[Path] = None,
+) -> Optional[dict[str, Any]]:
+    """
+    Convenience function to get latest session from the index.
+
+    Args:
+        cwd: Filter to sessions from this working directory
+        branch: Filter to sessions on this git branch
+        agent: Filter to specific agent ("claude" or "codex")
+        index_path: Path to index (default: ~/.cctools/search-index)
+
+    Returns:
+        Dict with session info including 'session_id' and 'export_path',
+        or None if no matching session found
+    """
+    if index_path is None:
+        index_path = Path.home() / ".cctools" / "search-index"
+
+    if not index_path.exists():
+        return None
+
+    try:
+        idx = SessionIndex(index_path)
+        return idx.get_latest_session(cwd=cwd, branch=branch, agent=agent)
+    except Exception:
+        return None
+
+
+def auto_index(
+    index_path: Optional[Path] = None,
+    claude_home: Optional[Path] = None,
+    codex_home: Optional[Path] = None,
+    verbose: bool = False,
+    silent: bool = False,
+) -> dict[str, Any]:
+    """
+    Automatically index new/changed session files on launch.
+
+    This implements the "zero-config" Recall model approach:
+    - Scans session directories for JSONL files
+    - Indexes new/modified files incrementally
+    - Returns quickly if nothing changed
+
+    Args:
+        index_path: Path to Tantivy index (default: ~/.cctools/search-index)
+        claude_home: Claude home directory (default: ~/.claude)
+        codex_home: Codex home directory (default: ~/.codex)
+        verbose: If True, print progress messages
+        silent: If True, suppress ALL output including progress bars (for JSON mode)
+
+    Returns:
+        Dict with stats: {indexed, skipped, failed, total_files}
+    """
+    _require_deps()
+
+    # Default paths
+    if index_path is None:
+        index_path = Path.home() / ".cctools" / "search-index"
+    if claude_home is None:
+        claude_home = Path.home() / ".claude"
+    if codex_home is None:
+        codex_home = Path.home() / ".codex"
+
+    # Find all JSONL session files
+    jsonl_files: list[Path] = []
+    claude_file_count = 0
+    codex_file_count = 0
+
+    # Claude sessions: ~/.claude/projects/**/*.jsonl
+    claude_projects = claude_home / "projects"
+    if claude_projects.exists():
+        claude_files = list(claude_projects.glob("**/*.jsonl"))
+        claude_file_count = len(claude_files)
+        jsonl_files.extend(claude_files)
+
+    # Codex sessions: ~/.codex/sessions/YYYY/MM/DD/*.jsonl (fixed depth)
+    codex_sessions = codex_home / "sessions"
+    if codex_sessions.exists():
+        codex_files = list(codex_sessions.glob("????/??/??/*.jsonl"))
+        codex_file_count = len(codex_files)
+        jsonl_files.extend(codex_files)
+
+    if verbose and not silent:
+        print(f"Claude home: {claude_home} ({claude_file_count} files)")
+        print(f"Codex home:  {codex_home} ({codex_file_count} files)")
+        print(f"Total: {len(jsonl_files)} session files to check")
+
+    if not jsonl_files:
+        return {
+            "indexed": 0, "skipped": 0, "failed": 0,
+            "total_files": 0, "claude_files": 0, "codex_files": 0,
+        }
+
+    # Create/open index and run incremental indexing
+    # In silent mode, suppress tqdm progress bar for clean JSON output
+    index = SessionIndex(index_path)
+    stats = index.index_from_jsonl(
+        jsonl_files,
+        incremental=True,
+        claude_home=claude_home,
+        codex_home=codex_home,
+        show_progress=not silent,
+    )
+    stats["total_files"] = len(jsonl_files)
+    stats["claude_files"] = claude_file_count
+    stats["codex_files"] = codex_file_count
+
+    # Prune deleted sessions from index
+    pruned = index.prune_deleted()
+    stats["pruned"] = pruned
+
+    # Always show indexing/pruning activity (even non-verbose), but suppress in silent mode
+    if not silent:
+        if stats["indexed"] > 0:
+            print(f"Indexed {stats['indexed']} new/modified sessions")
+        if pruned > 0:
+            print(f"Pruned {pruned} deleted sessions from index")
+        # Only show "up to date" in verbose mode (avoid noise on every command)
+        if verbose and stats["indexed"] == 0 and pruned == 0:
+            print("Index up to date")
+
+    return stats