zwarm 2.3.5__py3-none-any.whl → 3.6.0__py3-none-any.whl

zwarm/sessions/claude.py

@@ -0,0 +1,481 @@
+"""
+Claude Session Manager - Background process management for Claude Code agents.
+
+This module implements the ClaudeSessionManager, which handles:
+- Spawning `claude -p --output-format stream-json` subprocesses
+- Parsing Claude Code's stream-json output format
+- Loading config from .zwarm/claude.toml
+
+Inherits shared functionality from BaseSessionManager.
+"""
+
+from __future__ import annotations
+
+import json
+import subprocess
+import tomllib
+from datetime import datetime
+from pathlib import Path
+from typing import Any
+from uuid import uuid4
+
+from .base import (
+    BaseSessionManager,
+    Session,
+    SessionMessage,
+    SessionStatus,
+)
+
+
+class ClaudeSessionManager(BaseSessionManager):
+    """
+    Manages background Claude Code sessions.
+
+    Sessions are stored in:
+    .zwarm/sessions/<session_id>/
+        meta.json      - Session metadata
+        turns/
+            turn_1.jsonl
+            turn_2.jsonl
+            ...
+
+    Uses claude -p --output-format stream-json --verbose for JSON output.
+    """
+
+    adapter_name = "claude"
+    default_model = "sonnet"  # Claude uses aliases like sonnet, opus, haiku
+
+    # =========================================================================
+    # Claude-specific config handling
+    # =========================================================================
+
+    def _load_claude_config(self) -> dict[str, Any]:
+        """
+        Load claude.toml from state_dir.
+
+        Returns parsed TOML as dict, or empty dict if not found.
+        """
+        claude_toml = self.state_dir / "claude.toml"
+        if not claude_toml.exists():
+            return {}
+        try:
+            with open(claude_toml, "rb") as f:
+                return tomllib.load(f)
+        except Exception:
+            return {}
+
+    # =========================================================================
+    # Session lifecycle (Claude-specific implementation)
+    # =========================================================================
+
+    def start_session(
+        self,
+        task: str,
+        working_dir: Path | None = None,
+        model: str | None = None,
+        sandbox: str = "workspace-write",
+        source: str = "user",
+    ) -> Session:
+        """
+        Start a new Claude Code session in the background.
+
+        Args:
+            task: The task description
+            working_dir: Working directory for claude (default: cwd)
+            model: Model override (default: from claude.toml or sonnet)
+            sandbox: Sandbox mode - maps to permission modes
+            source: Who spawned this session ("user" or "orchestrator:<id>")
+
+        Returns:
+            The created session
+
+        Note:
+            Settings are read from .zwarm/claude.toml.
+            Run `zwarm init` to set up the config.
+        """
+        session_id = str(uuid4())
+        working_dir = working_dir or Path.cwd()
+        now = datetime.now().isoformat()
+
+        # Load claude config from .zwarm/claude.toml
+        claude_config = self._load_claude_config()
+
+        # Get model from config or use default
+        effective_model = model or claude_config.get("model", self.default_model)
+
+        # Check if full_danger mode is enabled
+        full_danger = claude_config.get("full_danger", False)
+
+        session = Session(
+            id=session_id,
+            task=task,
+            status=SessionStatus.PENDING,
+            working_dir=working_dir,
+            created_at=now,
+            updated_at=now,
+            model=effective_model,
+            turn=1,
+            messages=[SessionMessage(role="user", content=task, timestamp=now)],
+            source=source,
+            adapter=self.adapter_name,
+        )
+
+        # Create session directory
+        session_dir = self._session_dir(session_id)
+        session_dir.mkdir(parents=True, exist_ok=True)
+
+        # Build command
+        cmd = [
+            "claude",
+            "-p",  # Print mode (non-interactive)
+            "--output-format", "stream-json",
+            "--verbose",  # Required for stream-json
+            "--model", effective_model,
+        ]
+
+        # Add working directory access
+        if working_dir != Path.cwd():
+            cmd.extend(["--add-dir", str(working_dir.absolute())])
+
+        # Full danger mode bypasses all permission checks
+        if full_danger:
+            cmd.append("--dangerously-skip-permissions")
+
+        # Add the task as the prompt
+        cmd.append(task)
+
+        # Start process with output redirected to file
+        output_path = self._output_path(session_id, 1)
+        output_file = open(output_path, "w")
+
+        proc = subprocess.Popen(
+            cmd,
+            cwd=working_dir,
+            stdout=output_file,
+            stderr=subprocess.STDOUT,
+            start_new_session=True,  # Detach from parent process group
+        )
+
+        session.pid = proc.pid
+        session.status = SessionStatus.RUNNING
+        self._save_session(session)
+
+        return session
+
+    def inject_message(
+        self,
+        session_id: str,
+        message: str,
+    ) -> Session | None:
+        """
+        Inject a follow-up message into a completed session.
+
+        Claude Code supports --continue to continue a session, but for
+        simplicity we use the same context-injection pattern as Codex.
+
+        Args:
+            session_id: Session to continue
+            message: The follow-up message
+
+        Returns:
+            Updated session or None if not found/not ready
+        """
+        session = self.get_session(session_id)
+        if not session:
+            return None
+
+        if session.status == SessionStatus.RUNNING:
+            return None
+
+        # Build context from previous messages
+        context_parts = []
+        for msg in session.messages:
+            if msg.role == "user":
+                context_parts.append(f"USER: {msg.content}")
+            elif msg.role == "assistant":
+                context_parts.append(f"ASSISTANT: {msg.content}")
+
+        # Create augmented prompt with context
+        augmented_task = f"""Continue the following conversation:
+
+{chr(10).join(context_parts)}
+
+USER: {message}
+
+Continue from where you left off, addressing the user's new message."""
+
+        # Start new turn
+        session.turn += 1
+        now = datetime.now().isoformat()
+        session.messages.append(
+            SessionMessage(role="user", content=message, timestamp=now)
+        )
+
+        # Build command
+        claude_config = self._load_claude_config()
+        full_danger = claude_config.get("full_danger", False)
+
+        cmd = [
+            "claude",
+            "-p",
+            "--output-format", "stream-json",
+            "--verbose",
+            "--model", session.model,
+        ]
+
+        if session.working_dir != Path.cwd():
+            cmd.extend(["--add-dir", str(session.working_dir.absolute())])
+
+        if full_danger:
+            cmd.append("--dangerously-skip-permissions")
+
+        cmd.append(augmented_task)
+
+        # Start process
+        output_path = self._output_path(session.id, session.turn)
+        output_file = open(output_path, "w")
+
+        proc = subprocess.Popen(
+            cmd,
+            cwd=session.working_dir,
+            stdout=output_file,
+            stderr=subprocess.STDOUT,
+            start_new_session=True,
+        )
+
+        session.pid = proc.pid
+        session.status = SessionStatus.RUNNING
+        self._save_session(session)
+
+        return session
+
+    # =========================================================================
+    # Output parsing (Claude-specific stream-json format)
+    # =========================================================================
+
+    def _is_output_complete(self, session_id: str, turn: int) -> bool:
+        """
+        Check if output file indicates the task completed.
+
+        Looks for {"type":"result",...} in the stream-json output.
+        """
+        output_path = self._output_path(session_id, turn)
+        if not output_path.exists():
+            return False
+
+        try:
+            content = output_path.read_text()
+            for line in content.strip().split("\n"):
+                if not line.strip():
+                    continue
+                try:
+                    event = json.loads(line)
+                    event_type = event.get("type", "")
+                    # Claude uses "result" for completion
+                    if event_type == "result":
+                        return True
+                except json.JSONDecodeError:
+                    continue
+        except Exception:
+            pass
+
+        return False
+
+    def _parse_output(
+        self, output_path: Path
+    ) -> tuple[list[SessionMessage], dict[str, int], str | None]:
+        """
+        Parse stream-json output from claude -p.
+
+        Claude's stream-json format:
+        - {"type":"system","subtype":"init",...}
+        - {"type":"assistant","message":{"content":[{"type":"text","text":"..."}]}}
+        - {"type":"user","message":{"content":[{"type":"tool_result",...}]}}
+        - {"type":"result","usage":{...},...}
+
+        Returns:
+            (messages, token_usage, error)
+        """
+        messages: list[SessionMessage] = []
+        usage: dict[str, int] = {}
+        error: str | None = None
+
+        if not output_path.exists():
+            return messages, usage, "Output file not found"
+
+        content = output_path.read_text()
+
+        for line in content.strip().split("\n"):
+            if not line.strip():
+                continue
+
+            try:
+                event = json.loads(line)
+            except json.JSONDecodeError:
+                continue
+
+            event_type = event.get("type", "")
+
+            if event_type == "assistant":
+                # Assistant message with content array
+                msg = event.get("message", {})
+                content_blocks = msg.get("content", [])
+
+                for block in content_blocks:
+                    block_type = block.get("type", "")
+
+                    if block_type == "text":
+                        text = block.get("text", "")
+                        if text:
+                            messages.append(
+                                SessionMessage(
+                                    role="assistant",
+                                    content=text,
+                                    timestamp=datetime.now().isoformat(),
+                                )
+                            )
+
+                    elif block_type == "tool_use":
+                        # Track tool calls
+                        tool_name = block.get("name", "unknown")
+                        messages.append(
+                            SessionMessage(
+                                role="tool",
+                                content=f"[Calling: {tool_name}]",
+                                metadata={"function": tool_name, "tool_use_id": block.get("id")},
+                            )
+                        )
+
+            elif event_type == "user":
+                # Tool results
+                msg = event.get("message", {})
+                content_blocks = msg.get("content", [])
+
+                for block in content_blocks:
+                    if block.get("type") == "tool_result":
+                        tool_content = block.get("content", "")
+                        is_error = block.get("is_error", False)
+                        if tool_content and len(str(tool_content)) < 500:
+                            prefix = "[Error]" if is_error else "[Output]"
+                            messages.append(
+                                SessionMessage(
+                                    role="tool",
+                                    content=f"{prefix}: {str(tool_content)[:500]}",
+                                )
+                            )
+
+            elif event_type == "result":
+                # Final result with usage info
+                result_usage = event.get("usage", {})
+
+                # Map Claude's usage fields to our standard format
+                usage["input_tokens"] = result_usage.get("input_tokens", 0)
+                usage["output_tokens"] = result_usage.get("output_tokens", 0)
+                usage["cache_read_input_tokens"] = result_usage.get("cache_read_input_tokens", 0)
+                usage["cache_creation_input_tokens"] = result_usage.get("cache_creation_input_tokens", 0)
+                usage["total_tokens"] = usage["input_tokens"] + usage["output_tokens"]
+
+                # Check for errors
+                if event.get("is_error"):
+                    error = event.get("result", "Unknown error")
+                elif event.get("subtype") == "error":
+                    error = event.get("result", "Unknown error")
+
+        return messages, usage, error
+
+    def get_trajectory(
+        self, session_id: str, full: bool = False, max_output_len: int = 200
+    ) -> list[dict]:
+        """
+        Get the full trajectory of a session - all steps in order.
+
+        Args:
+            session_id: Session to get trajectory for
+            full: If True, include full untruncated content
+            max_output_len: Max length for outputs when full=False
+
+        Returns a list of step dicts with type, summary, and details.
+        """
+        if full:
+            max_output_len = 999999
+
+        session = self.get_session(session_id)
+        if not session:
+            return []
+
+        trajectory = []
+
+        for turn in range(1, session.turn + 1):
+            output_path = self._output_path(session.id, turn)
+            if not output_path.exists():
+                continue
+
+            content = output_path.read_text()
+            step_num = 0
+
+            for line in content.strip().split("\n"):
+                if not line.strip():
+                    continue
+
+                try:
+                    event = json.loads(line)
+                except json.JSONDecodeError:
+                    continue
+
+                event_type = event.get("type", "")
+
+                if event_type == "assistant":
+                    msg = event.get("message", {})
+                    content_blocks = msg.get("content", [])
+
+                    for block in content_blocks:
+                        block_type = block.get("type", "")
+                        step_num += 1
+
+                        if block_type == "text":
+                            text = block.get("text", "")
+                            summary_len = max_output_len if full else 200
+                            trajectory.append({
+                                "turn": turn,
+                                "step": step_num,
+                                "type": "message",
+                                "summary": text[:summary_len] + ("..." if len(text) > summary_len else ""),
+                                "full_text": text if full else None,
+                                "full_length": len(text),
+                            })
+
+                        elif block_type == "tool_use":
+                            tool_name = block.get("name", "unknown")
+                            args = block.get("input", {})
+                            args_str = str(args)
+                            args_len = max_output_len if full else 100
+                            trajectory.append({
+                                "turn": turn,
+                                "step": step_num,
+                                "type": "tool_call",
+                                "tool": tool_name,
+                                "args_preview": args_str[:args_len] + ("..." if len(args_str) > args_len else ""),
+                                "full_args": args if full else None,
+                            })
+
+                elif event_type == "user":
+                    # Tool results
+                    msg = event.get("message", {})
+                    content_blocks = msg.get("content", [])
+
+                    for block in content_blocks:
+                        if block.get("type") == "tool_result":
+                            step_num += 1
+                            output = str(block.get("content", ""))
+                            output_preview = output[:max_output_len]
+                            if len(output) > max_output_len:
+                                output_preview += "..."
+                            trajectory.append({
+                                "turn": turn,
+                                "step": step_num,
+                                "type": "tool_output",
+                                "output": output_preview,
+                                "is_error": block.get("is_error", False),
+                            })
+
+        return trajectory