htmlgraph 0.23.4__py3-none-any.whl → 0.24.0__py3-none-any.whl

htmlgraph/orchestration/headless_spawner.py

@@ -1,8 +1,13 @@
 """Headless AI spawner for multi-AI orchestration."""
 
 import json
+import os
 import subprocess
 from dataclasses import dataclass
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from htmlgraph.sdk import SDK
 
 
 @dataclass
@@ -14,6 +19,7 @@ class AIResult:
     tokens_used: int | None
     error: str | None
     raw_output: dict | list | str | None
+    tracked_events: list[dict] | None = None  # Events tracked in HtmlGraph
 
 
 class HeadlessSpawner:
@@ -59,13 +65,316 @@ class HeadlessSpawner:
         """Initialize spawner."""
         pass
 
+    def _get_sdk(self) -> "SDK | None":
+        """
+        Get SDK instance for HtmlGraph tracking with parent session support.
+
+        Returns None if SDK unavailable.
+        """
+        try:
+            from htmlgraph.sdk import SDK
+
+            # Read parent session context from environment
+            parent_session = os.getenv("HTMLGRAPH_PARENT_SESSION")
+            parent_agent = os.getenv("HTMLGRAPH_PARENT_AGENT")
+
+            # Create SDK with parent session context
+            sdk = SDK(
+                agent=f"spawner-{parent_agent}" if parent_agent else "spawner",
+                parent_session=parent_session,  # Pass parent session
+            )
+
+            return sdk
+
+        except Exception:
+            # SDK unavailable or not properly initialized (optional dependency)
+            # This happens in test contexts without active sessions
+            # Don't log error to avoid noise in tests
+            return None
+
+    def _parse_and_track_gemini_events(
+        self, jsonl_output: str, sdk: "SDK"
+    ) -> list[dict]:
+        """
+        Parse Gemini stream-json events and track in HtmlGraph.
+
+        Args:
+            jsonl_output: JSONL output from Gemini CLI
+            sdk: HtmlGraph SDK instance for tracking
+
+        Returns:
+            Parsed events list
+        """
+        events = []
+
+        # Get parent context for metadata
+        parent_activity = os.getenv("HTMLGRAPH_PARENT_ACTIVITY")
+        nesting_depth_str = os.getenv("HTMLGRAPH_NESTING_DEPTH", "0")
+        nesting_depth = int(nesting_depth_str) if nesting_depth_str.isdigit() else 0
+
+        for line in jsonl_output.splitlines():
+            if not line.strip():
+                continue
+
+            try:
+                event = json.loads(line)
+                events.append(event)
+
+                # Track based on event type
+                event_type = event.get("type")
+
+                try:
+                    if event_type == "tool_use":
+                        tool_name = event.get("tool_name", "unknown_tool")
+                        parameters = event.get("parameters", {})
+                        payload = {
+                            "tool_name": tool_name,
+                            "parameters": parameters,
+                        }
+                        if parent_activity:
+                            payload["parent_activity"] = parent_activity
+                        if nesting_depth > 0:
+                            payload["nesting_depth"] = nesting_depth
+                        sdk.track_activity(
+                            tool="gemini_tool_call",
+                            summary=f"Gemini called {tool_name}",
+                            payload=payload,
+                        )
+
+                    elif event_type == "tool_result":
+                        status = event.get("status", "unknown")
+                        success = status == "success"
+                        tool_id = event.get("tool_id", "unknown")
+                        payload = {"tool_id": tool_id, "status": status}
+                        if parent_activity:
+                            payload["parent_activity"] = parent_activity
+                        if nesting_depth > 0:
+                            payload["nesting_depth"] = nesting_depth
+                        sdk.track_activity(
+                            tool="gemini_tool_result",
+                            summary=f"Gemini tool result: {status}",
+                            success=success,
+                            payload=payload,
+                        )
+
+                    elif event_type == "message":
+                        role = event.get("role")
+                        if role == "assistant":
+                            content = event.get("content", "")
+                            # Truncate for summary
+                            summary = (
+                                content[:100] + "..." if len(content) > 100 else content
+                            )
+                            payload = {"role": role, "content_length": len(content)}
+                            if parent_activity:
+                                payload["parent_activity"] = parent_activity
+                            if nesting_depth > 0:
+                                payload["nesting_depth"] = nesting_depth
+                            sdk.track_activity(
+                                tool="gemini_message",
+                                summary=f"Gemini: {summary}",
+                                payload=payload,
+                            )
+
+                    elif event_type == "result":
+                        stats = event.get("stats", {})
+                        payload = {"stats": stats}
+                        if parent_activity:
+                            payload["parent_activity"] = parent_activity
+                        if nesting_depth > 0:
+                            payload["nesting_depth"] = nesting_depth
+                        sdk.track_activity(
+                            tool="gemini_completion",
+                            summary="Gemini task completed",
+                            payload=payload,
+                        )
+                except Exception:
+                    # Tracking failure should not break parsing
+                    pass
+
+            except json.JSONDecodeError:
+                # Skip malformed lines
+                continue
+
+        return events
+
+    def _parse_and_track_codex_events(
+        self, jsonl_output: str, sdk: "SDK"
+    ) -> list[dict]:
+        """
+        Parse Codex JSONL events and track in HtmlGraph.
+
+        Args:
+            jsonl_output: JSONL output from Codex CLI
+            sdk: HtmlGraph SDK instance for tracking
+
+        Returns:
+            Parsed events list
+        """
+        events = []
+        parse_errors = []
+
+        # Get parent context for metadata
+        parent_activity = os.getenv("HTMLGRAPH_PARENT_ACTIVITY")
+        nesting_depth_str = os.getenv("HTMLGRAPH_NESTING_DEPTH", "0")
+        nesting_depth = int(nesting_depth_str) if nesting_depth_str.isdigit() else 0
+
+        for line_num, line in enumerate(jsonl_output.splitlines(), start=1):
+            if not line.strip():
+                continue
+
+            try:
+                event = json.loads(line)
+                events.append(event)
+
+                event_type = event.get("type")
+
+                try:
+                    # Track item.started events
+                    if event_type == "item.started":
+                        item = event.get("item", {})
+                        item_type = item.get("type")
+
+                        if item_type == "command_execution":
+                            command = item.get("command", "")
+                            payload = {"command": command}
+                            if parent_activity:
+                                payload["parent_activity"] = parent_activity
+                            if nesting_depth > 0:
+                                payload["nesting_depth"] = nesting_depth
+                            sdk.track_activity(
+                                tool="codex_command",
+                                summary=f"Codex executing: {command[:80]}",
+                                payload=payload,
+                            )
+
+                    # Track item.completed events
+                    elif event_type == "item.completed":
+                        item = event.get("item", {})
+                        item_type = item.get("type")
+
+                        if item_type == "file_change":
+                            path = item.get("path", "unknown")
+                            payload = {"path": path}
+                            if parent_activity:
+                                payload["parent_activity"] = parent_activity
+                            if nesting_depth > 0:
+                                payload["nesting_depth"] = nesting_depth
+                            sdk.track_activity(
+                                tool="codex_file_change",
+                                summary=f"Codex modified: {path}",
+                                file_paths=[path],
+                                payload=payload,
+                            )
+
+                        elif item_type == "agent_message":
+                            text = item.get("text", "")
+                            summary = text[:100] + "..." if len(text) > 100 else text
+                            payload = {"text_length": len(text)}
+                            if parent_activity:
+                                payload["parent_activity"] = parent_activity
+                            if nesting_depth > 0:
+                                payload["nesting_depth"] = nesting_depth
+                            sdk.track_activity(
+                                tool="codex_message",
+                                summary=f"Codex: {summary}",
+                                payload=payload,
+                            )
+
+                    # Track turn.completed for token usage
+                    elif event_type == "turn.completed":
+                        usage = event.get("usage", {})
+                        total_tokens = sum(usage.values())
+                        payload = {"usage": usage}
+                        if parent_activity:
+                            payload["parent_activity"] = parent_activity
+                        if nesting_depth > 0:
+                            payload["nesting_depth"] = nesting_depth
+                        sdk.track_activity(
+                            tool="codex_completion",
+                            summary=f"Codex turn completed ({total_tokens} tokens)",
+                            payload=payload,
+                        )
+                except Exception:
+                    # Tracking failure should not break parsing
+                    pass
+
+            except json.JSONDecodeError as e:
+                parse_errors.append(
+                    {
+                        "line_number": line_num,
+                        "error": str(e),
+                        "content": line[:100],
+                    }
+                )
+                continue
+
+        return events
+
+    def _parse_and_track_copilot_events(
+        self, prompt: str, response: str, sdk: "SDK"
+    ) -> list[dict]:
+        """
+        Track Copilot execution (start and result only).
+
+        Args:
+            prompt: Original prompt
+            response: Response from Copilot
+            sdk: HtmlGraph SDK instance for tracking
+
+        Returns:
+            Synthetic events list for consistency
+        """
+        events = []
+
+        # Get parent context for metadata
+        parent_activity = os.getenv("HTMLGRAPH_PARENT_ACTIVITY")
+        nesting_depth_str = os.getenv("HTMLGRAPH_NESTING_DEPTH", "0")
+        nesting_depth = int(nesting_depth_str) if nesting_depth_str.isdigit() else 0
+
+        try:
+            # Track start
+            start_event = {"type": "copilot_start", "prompt": prompt[:100]}
+            events.append(start_event)
+            payload: dict[str, str | int] = {"prompt_length": len(prompt)}
+            if parent_activity:
+                payload["parent_activity"] = parent_activity
+            if nesting_depth > 0:
+                payload["nesting_depth"] = nesting_depth
+            sdk.track_activity(
+                tool="copilot_start",
+                summary=f"Copilot started with prompt: {prompt[:80]}",
+                payload=payload,
+            )
+        except Exception:
+            pass
+
+        try:
+            # Track result
+            result_event = {"type": "copilot_result", "response": response[:100]}
+            events.append(result_event)
+            payload_result: dict[str, str | int] = {"response_length": len(response)}
+            if parent_activity:
+                payload_result["parent_activity"] = parent_activity
+            if nesting_depth > 0:
+                payload_result["nesting_depth"] = nesting_depth
+            sdk.track_activity(
+                tool="copilot_result",
+                summary=f"Copilot completed: {response[:80]}",
+                payload=payload_result,
+            )
+        except Exception:
+            pass
+
+        return events
+
     def spawn_gemini(
         self,
         prompt: str,
-        output_format: str = "json",
+        output_format: str = "stream-json",
         model: str | None = None,
         include_directories: list[str] | None = None,
-        color: str = "auto",
+        track_in_htmlgraph: bool = True,
         timeout: int = 120,
     ) -> AIResult:
         """
@@ -73,15 +382,21 @@ class HeadlessSpawner:
 
         Args:
             prompt: Task description for Gemini
-            output_format: "json" or "stream-json"
+            output_format: "json" or "stream-json" (default: "stream-json" for real-time tracking)
             model: Model selection (e.g., "gemini-2.0-flash"). Default: None (uses default)
             include_directories: List of directories to include for context. Default: None
-            color: Color output control ("auto", "on", "off"). Default: "auto"
+            track_in_htmlgraph: Enable HtmlGraph activity tracking. Default: True
             timeout: Max seconds to wait
 
         Returns:
-            AIResult with response or error
+            AIResult with response or error and tracked events if tracking enabled
         """
+        # Initialize tracking if enabled
+        sdk: SDK | None = None
+        tracked_events: list[dict] = []
+        if track_in_htmlgraph:
+            sdk = self._get_sdk()
+
         try:
             # Build command based on tested pattern from spike spk-4029eef3
             cmd = ["gemini", "-p", prompt, "--output-format", output_format]
@@ -95,8 +410,20 @@ class HeadlessSpawner:
                 for directory in include_directories:
                     cmd.extend(["--include-directories", directory])
 
-            # Add color option
-            cmd.extend(["--color", color])
+            # CRITICAL: Add --yolo for headless mode (auto-approve all tools)
+            cmd.append("--yolo")
+
+            # Track spawner start if SDK available
+            if sdk:
+                try:
+                    sdk.track_activity(
+                        tool="gemini_spawn_start",
+                        summary=f"Spawning Gemini: {prompt[:80]}",
+                        payload={"prompt_length": len(prompt), "model": model},
+                    )
+                except Exception:
+                    # Tracking failure should not break execution
+                    pass
 
             # Execute with timeout and stderr redirection
             # Note: Cannot use capture_output with stderr parameter
@@ -116,9 +443,58 @@ class HeadlessSpawner:
                     tokens_used=None,
                     error=f"Gemini CLI failed with exit code {result.returncode}",
                     raw_output=None,
+                    tracked_events=tracked_events,
                 )
 
-            # Parse JSON response
+            # Handle stream-json format with real-time tracking
+            if output_format == "stream-json" and sdk:
+                try:
+                    tracked_events = self._parse_and_track_gemini_events(
+                        result.stdout, sdk
+                    )
+                    # Only use stream-json parsing if we got valid events
+                    if tracked_events:
+                        # For stream-json, we need to extract response differently
+                        # Look for the last message or result event
+                        response_text = ""
+                        for event in tracked_events:
+                            if event.get("type") == "result":
+                                response_text = event.get("response", "")
+                                break
+                            elif event.get("type") == "message":
+                                content = event.get("content", "")
+                                if content:
+                                    response_text = content
+
+                        # Token usage from stats in result event
+                        tokens = None
+                        for event in tracked_events:
+                            if event.get("type") == "result":
+                                stats = event.get("stats", {})
+                                if stats and "models" in stats:
+                                    total_tokens = 0
+                                    for model_stats in stats["models"].values():
+                                        model_tokens = model_stats.get(
+                                            "tokens", {}
+                                        ).get("total", 0)
+                                        total_tokens += model_tokens
+                                    tokens = total_tokens if total_tokens > 0 else None
+                                break
+
+                        return AIResult(
+                            success=True,
+                            response=response_text,
+                            tokens_used=tokens,
+                            error=None,
+                            raw_output={"events": tracked_events},
+                            tracked_events=tracked_events,
+                        )
+
+                except Exception:
+                    # Fall back to regular JSON parsing if tracking fails
+                    pass
+
+            # Parse JSON response (for json format or fallback)
             try:
                 output = json.loads(result.stdout)
             except json.JSONDecodeError as e:
@@ -128,6 +504,7 @@ class HeadlessSpawner:
                     tokens_used=None,
                     error=f"Failed to parse JSON output: {e}",
                     raw_output={"stdout": result.stdout},
+                    tracked_events=tracked_events,
                 )
 
             # Extract response and token usage from parsed output
@@ -150,15 +527,22 @@ class HeadlessSpawner:
                 tokens_used=tokens,
                 error=None,
                 raw_output=output,
+                tracked_events=tracked_events,
             )
 
-        except subprocess.TimeoutExpired:
+        except subprocess.TimeoutExpired as e:
             return AIResult(
                 success=False,
                 response="",
                 tokens_used=None,
                 error=f"Gemini CLI timed out after {timeout} seconds",
-                raw_output=None,
+                raw_output={
+                    "partial_stdout": e.stdout.decode() if e.stdout else None,
+                    "partial_stderr": e.stderr.decode() if e.stderr else None,
+                }
+                if e.stdout or e.stderr
+                else None,
+                tracked_events=tracked_events,
             )
         except FileNotFoundError:
             return AIResult(
@@ -167,6 +551,7 @@ class HeadlessSpawner:
                 tokens_used=None,
                 error="Gemini CLI not found. Ensure 'gemini' is installed and in PATH.",
                 raw_output=None,
+                tracked_events=tracked_events,
             )
         except Exception as e:
             return AIResult(
@@ -175,24 +560,24 @@ class HeadlessSpawner:
                 tokens_used=None,
                 error=f"Unexpected error: {type(e).__name__}: {e}",
                 raw_output=None,
+                tracked_events=tracked_events,
             )
 
     def spawn_codex(
         self,
         prompt: str,
-        approval: str = "never",
         output_json: bool = True,
         model: str | None = None,
         sandbox: str | None = None,
-        full_auto: bool = False,
+        full_auto: bool = True,
         images: list[str] | None = None,
-        color: str = "auto",
         output_last_message: str | None = None,
         output_schema: str | None = None,
         skip_git_check: bool = False,
         working_directory: str | None = None,
         use_oss: bool = False,
         bypass_approvals: bool = False,
+        track_in_htmlgraph: bool = True,
         timeout: int = 120,
     ) -> AIResult:
         """
@@ -200,24 +585,29 @@ class HeadlessSpawner:
 
         Args:
             prompt: Task description for Codex
-            approval: Approval mode ("never", "always")
-            output_json: Use --json flag for JSONL output
+            output_json: Use --json flag for JSONL output (enables real-time tracking)
             model: Model selection (e.g., "gpt-4-turbo"). Default: None
             sandbox: Sandbox mode ("read-only", "workspace-write", "danger-full-access"). Default: None
-            full_auto: Enable full auto mode (--full-auto). Default: False
+            full_auto: Enable full auto mode (--full-auto). Default: True (required for headless)
             images: List of image paths (--image). Default: None
-            color: Color output control ("auto", "on", "off"). Default: "auto"
             output_last_message: Write last message to file (--output-last-message). Default: None
             output_schema: JSON schema for validation (--output-schema). Default: None
             skip_git_check: Skip git repo check (--skip-git-repo-check). Default: False
             working_directory: Workspace directory (--cd). Default: None
             use_oss: Use local Ollama provider (--oss). Default: False
             bypass_approvals: Dangerously bypass approvals (--dangerously-bypass-approvals-and-sandbox). Default: False
+            track_in_htmlgraph: Enable HtmlGraph activity tracking. Default: True
             timeout: Max seconds to wait
 
         Returns:
-            AIResult with response or error
+            AIResult with response, error, and tracked events if tracking enabled
         """
+        # Initialize tracking if enabled
+        sdk: SDK | None = None
+        tracked_events: list[dict] = []
+        if track_in_htmlgraph and output_json:
+            sdk = self._get_sdk()
+
         cmd = ["codex", "exec"]
 
         if output_json:
@@ -240,9 +630,6 @@ class HeadlessSpawner:
             for image in images:
                 cmd.extend(["--image", image])
 
-        # Add color option
-        cmd.extend(["--color", color])
-
         # Add output last message file if specified
         if output_last_message:
             cmd.extend(["--output-last-message", output_last_message])
@@ -267,7 +654,24 @@ class HeadlessSpawner:
         if bypass_approvals:
             cmd.append("--dangerously-bypass-approvals-and-sandbox")
 
-        cmd.extend(["--approval", approval, prompt])
+        # Add prompt as final argument
+        cmd.append(prompt)
+
+        # Track spawner start if SDK available
+        if sdk:
+            try:
+                sdk.track_activity(
+                    tool="codex_spawn_start",
+                    summary=f"Spawning Codex: {prompt[:80]}",
+                    payload={
+                        "prompt_length": len(prompt),
+                        "model": model,
+                        "sandbox": sandbox,
+                    },
+                )
+            except Exception:
+                # Tracking failure should not break execution
+                pass
 
         try:
             result = subprocess.run(
@@ -286,16 +690,34 @@ class HeadlessSpawner:
                     tokens_used=None,
                     error=None if result.returncode == 0 else "Command failed",
                     raw_output=result.stdout,
+                    tracked_events=tracked_events,
                 )
 
             # Parse JSONL output
             events = []
-            for line in result.stdout.splitlines():
-                if line.strip():
-                    try:
-                        events.append(json.loads(line))
-                    except json.JSONDecodeError:
-                        continue
+            parse_errors = []
+
+            # Use tracking parser if SDK is available
+            if sdk:
+                tracked_events = self._parse_and_track_codex_events(result.stdout, sdk)
+                events = tracked_events
+            else:
+                # Fallback to regular parsing without tracking
+                for line_num, line in enumerate(result.stdout.splitlines(), start=1):
+                    if line.strip():
+                        try:
+                            events.append(json.loads(line))
+                        except json.JSONDecodeError as e:
+                            parse_errors.append(
+                                {
+                                    "line_number": line_num,
+                                    "error": str(e),
+                                    "content": line[
+                                        :100
+                                    ],  # First 100 chars for debugging
+                                }
+                            )
+                            continue
 
             # Extract agent message
             response = None
@@ -318,7 +740,11 @@ class HeadlessSpawner:
                 response=response or "",
                 tokens_used=tokens,
                 error=None if result.returncode == 0 else "Command failed",
-                raw_output=events,
+                raw_output={
+                    "events": events,
+                    "parse_errors": parse_errors if parse_errors else None,
+                },
+                tracked_events=tracked_events,
             )
 
         except FileNotFoundError:
@@ -328,14 +754,30 @@ class HeadlessSpawner:
                 tokens_used=None,
                 error="Codex CLI not found. Install from: https://github.com/openai/codex",
                 raw_output=None,
+                tracked_events=tracked_events,
             )
-        except subprocess.TimeoutExpired:
+        except subprocess.TimeoutExpired as e:
             return AIResult(
                 success=False,
                 response="",
                 tokens_used=None,
                 error=f"Timed out after {timeout} seconds",
+                raw_output={
+                    "partial_stdout": e.stdout.decode() if e.stdout else None,
+                    "partial_stderr": e.stderr.decode() if e.stderr else None,
+                }
+                if e.stdout or e.stderr
+                else None,
+                tracked_events=tracked_events,
+            )
+        except Exception as e:
+            return AIResult(
+                success=False,
+                response="",
+                tokens_used=None,
+                error=f"Unexpected error: {type(e).__name__}: {e}",
                 raw_output=None,
+                tracked_events=tracked_events,
             )
 
     def spawn_copilot(
@@ -344,6 +786,7 @@ class HeadlessSpawner:
         allow_tools: list[str] | None = None,
         allow_all_tools: bool = False,
         deny_tools: list[str] | None = None,
+        track_in_htmlgraph: bool = True,
         timeout: int = 120,
     ) -> AIResult:
         """
@@ -354,11 +797,18 @@ class HeadlessSpawner:
             allow_tools: List of tools to auto-approve (e.g., ["shell(git)", "write(*.py)"])
             allow_all_tools: Auto-approve all tools (--allow-all-tools). Default: False
             deny_tools: List of tools to deny (--deny-tool). Default: None
+            track_in_htmlgraph: Enable HtmlGraph activity tracking. Default: True
             timeout: Max seconds to wait
 
         Returns:
-            AIResult with response or error
+            AIResult with response, error, and tracked events if tracking enabled
         """
+        # Initialize tracking if enabled
+        sdk = None
+        tracked_events = []
+        if track_in_htmlgraph:
+            sdk = self._get_sdk()
+
         cmd = ["copilot", "-p", prompt]
 
         # Add allow all tools flag
@@ -375,6 +825,18 @@ class HeadlessSpawner:
             for tool in deny_tools:
                 cmd.extend(["--deny-tool", tool])
 
+        # Track spawner start if SDK available
+        if sdk:
+            try:
+                sdk.track_activity(
+                    tool="copilot_spawn_start",
+                    summary=f"Spawning Copilot: {prompt[:80]}",
+                    payload={"prompt_length": len(prompt)},
+                )
+            except Exception:
+                # Tracking failure should not break execution
+                pass
+
         try:
             result = subprocess.run(
                 cmd,
@@ -406,12 +868,19 @@ class HeadlessSpawner:
                     tokens = 0  # Placeholder
                     break
 
+            # Track Copilot execution if SDK available
+            if sdk:
+                tracked_events = self._parse_and_track_copilot_events(
+                    prompt, response, sdk
+                )
+
             return AIResult(
                 success=result.returncode == 0,
                 response=response,
                 tokens_used=tokens,
                 error=None if result.returncode == 0 else result.stderr,
                 raw_output=result.stdout,
+                tracked_events=tracked_events,
             )
 
         except FileNotFoundError:
@@ -421,14 +890,30 @@ class HeadlessSpawner:
                 tokens_used=None,
                 error="Copilot CLI not found. Install from: https://docs.github.com/en/copilot/using-github-copilot/using-github-copilot-in-the-command-line",
                 raw_output=None,
+                tracked_events=tracked_events,
             )
-        except subprocess.TimeoutExpired:
+        except subprocess.TimeoutExpired as e:
             return AIResult(
                 success=False,
                 response="",
                 tokens_used=None,
                 error=f"Timed out after {timeout} seconds",
+                raw_output={
+                    "partial_stdout": e.stdout.decode() if e.stdout else None,
+                    "partial_stderr": e.stderr.decode() if e.stderr else None,
+                }
+                if e.stdout or e.stderr
+                else None,
+                tracked_events=tracked_events,
+            )
+        except Exception as e:
+            return AIResult(
+                success=False,
+                response="",
+                tokens_used=None,
+                error=f"Unexpected error: {type(e).__name__}: {e}",
                 raw_output=None,
+                tracked_events=tracked_events,
             )
 
     def spawn_claude(
@@ -548,19 +1033,24 @@ class HeadlessSpawner:
                 error="Claude CLI not found. Install Claude Code from: https://claude.com/claude-code",
                 raw_output=None,
             )
-        except subprocess.TimeoutExpired:
+        except subprocess.TimeoutExpired as e:
             return AIResult(
                 success=False,
                 response="",
                 tokens_used=None,
                 error=f"Timed out after {timeout} seconds",
-                raw_output=None,
+                raw_output={
+                    "partial_stdout": e.stdout.decode() if e.stdout else None,
+                    "partial_stderr": e.stderr.decode() if e.stderr else None,
+                }
+                if e.stdout or e.stderr
+                else None,
             )
         except Exception as e:
             return AIResult(
                 success=False,
                 response="",
                 tokens_used=None,
-                error=f"Unexpected error: {str(e)}",
+                error=f"Unexpected error: {type(e).__name__}: {e}",
                 raw_output=None,
             )