Merge pull request #63 from SentienceAPI/cloud_sync

cloud tracing support

Author: rcholic

.gitignore

@@ -40,6 +40,9 @@ ENV/
 htmlcov/
 .tox/
 
+# Traces (runtime and test-generated)
+traces/
+
 # Jupyter
 .ipynb_checkpoints
 

examples/click_rect_demo.py

@@ -37,7 +37,12 @@ def main():
             print("   Clicking at center of element's bbox...")
             result = click_rect(
                 browser,
-                {"x": link.bbox.x, "y": link.bbox.y, "w": link.bbox.width, "h": link.bbox.height},
+                {
+                    "x": link.bbox.x,
+                    "y": link.bbox.y,
+                    "w": link.bbox.width,
+                    "h": link.bbox.height,
+                },
             )
             print(f"   Result: success={result.success}, outcome={result.outcome}")
             print(f"   URL changed: {result.url_changed}\n")

examples/cloud_tracing_agent.py

@@ -0,0 +1,111 @@
+"""
+Example: Agent with Cloud Tracing
+
+Demonstrates how to use cloud tracing with SentienceAgent to upload traces
+and screenshots to cloud storage for remote viewing and analysis.
+
+Requirements:
+- Pro or Enterprise tier API key (SENTIENCE_API_KEY)
+- OpenAI API key (OPENAI_API_KEY) for LLM
+
+Usage:
+    python examples/cloud_tracing_agent.py
+"""
+
+import os
+
+from sentience import SentienceAgent, SentienceBrowser
+from sentience.agent_config import AgentConfig
+from sentience.llm_provider import OpenAIProvider
+from sentience.tracer_factory import create_tracer
+
+
+def main():
+    # Get API keys from environment
+    sentience_key = os.environ.get("SENTIENCE_API_KEY")
+    openai_key = os.environ.get("OPENAI_API_KEY")
+
+    if not sentience_key:
+        print("❌ Error: SENTIENCE_API_KEY not set")
+        print("   Cloud tracing requires Pro or Enterprise tier")
+        print("   Get your API key at: https://sentience.studio")
+        return
+
+    if not openai_key:
+        print("❌ Error: OPENAI_API_KEY not set")
+        return
+
+    print("🚀 Starting Agent with Cloud Tracing Demo\n")
+
+    # 1. Create tracer with automatic tier detection
+    # If api_key is Pro/Enterprise, uses CloudTraceSink
+    # If api_key is missing/invalid, falls back to local JsonlTraceSink
+    run_id = "cloud-tracing-demo"
+    tracer = create_tracer(api_key=sentience_key, run_id=run_id)
+
+    print(f"🆔 Run ID: {run_id}\n")
+
+    # 2. Configure agent with screenshot capture
+    config = AgentConfig(
+        snapshot_limit=50,
+        capture_screenshots=True,  # Enable screenshot capture
+        screenshot_format="jpeg",  # JPEG for smaller file size
+        screenshot_quality=80,  # 80% quality (good balance)
+    )
+
+    # 3. Create browser and LLM
+    browser = SentienceBrowser(api_key=sentience_key, headless=False)
+    llm = OpenAIProvider(api_key=openai_key, model="gpt-4o-mini")
+
+    # 4. Create agent with tracer
+    agent = SentienceAgent(browser, llm, tracer=tracer, config=config)
+
+    try:
+        # 5. Navigate and execute agent actions
+        print("🌐 Navigating to Google...\n")
+        browser.start()
+        browser.page.goto("https://www.google.com")
+        browser.page.wait_for_load_state("networkidle")
+
+        # All actions are automatically traced!
+        print("📝 Executing agent actions (all automatically traced)...\n")
+        agent.act("Click the search box")
+        agent.act("Type 'Sentience AI agent SDK' into the search field")
+        agent.act("Press Enter key")
+
+        # Wait for results
+        import time
+
+        time.sleep(2)
+
+        agent.act("Click the first non-ad search result")
+
+        print("\n✅ Agent execution complete!")
+
+        # 6. Get token usage stats
+        stats = agent.get_token_stats()
+        print("\n📊 Token Usage:")
+        print(f"   Total tokens: {stats.total_tokens}")
+        print(f"   Prompt tokens: {stats.total_prompt_tokens}")
+        print(f"   Completion tokens: {stats.total_completion_tokens}")
+
+    except Exception as e:
+        print(f"\n❌ Error during execution: {e}")
+        raise
+
+    finally:
+        # 7. Close tracer (uploads to cloud)
+        print("\n📤 Uploading trace to cloud...")
+        try:
+            tracer.close(blocking=True)  # Wait for upload to complete
+            print("✅ Trace uploaded successfully!")
+            print(f"   View at: https://studio.sentienceapi.com (run_id: {run_id})")
+        except Exception as e:
+            print(f"⚠️  Upload failed: {e}")
+            print(f"   Trace preserved locally at: ~/.sentience/traces/pending/{run_id}.jsonl")
+
+        browser.close()
+
+
+if __name__ == "__main__":
+    main()

examples/test_local_llm_agent.py

@@ -65,7 +65,10 @@ def test_local_llm_basic():
     user_prompt = "What is the next step to achieve the goal?"
 
     response = llm.generate(
-        system_prompt=system_prompt, user_prompt=user_prompt, max_new_tokens=20, temperature=0.0
+        system_prompt=system_prompt,
+        user_prompt=user_prompt,
+        max_new_tokens=20,
+        temperature=0.0,
     )
 
     print(f"Agent Response: {response.content}")

sentience/__init__.py

@@ -9,6 +9,9 @@
 # Agent Layer (Phase 1 & 2)
 from .base_agent import BaseAgent
 from .browser import SentienceBrowser
+
+# Tracing (v0.12.0+)
+from .cloud_tracing import CloudTraceSink
 from .conversational_agent import ConversationalAgent
 from .expect import expect
 
@@ -43,8 +46,7 @@
 from .recorder import Recorder, Trace, TraceStep, record
 from .screenshot import screenshot
 from .snapshot import snapshot
-
-# Tracing (v0.12.0+)
+from .tracer_factory import SENTIENCE_API_URL, create_tracer
 from .tracing import JsonlTraceSink, TraceEvent, Tracer, TraceSink
 
 # Utilities (v0.12.0+)
@@ -107,7 +109,10 @@
     "Tracer",
     "TraceSink",
     "JsonlTraceSink",
+    "CloudTraceSink",
     "TraceEvent",
+    "create_tracer",
+    "SENTIENCE_API_URL",
     # Utilities (v0.12.0+)
     "canonical_snapshot_strict",
     "canonical_snapshot_loose",

sentience/actions.py

@@ -3,14 +3,13 @@
 """
 
 import time
-from typing import Any, Dict, Optional
 
 from .browser import SentienceBrowser
 from .models import ActionResult, BBox, Snapshot
 from .snapshot import snapshot
 
 
-def click(
+def click(  # noqa: C901
     browser: SentienceBrowser,
     element_id: int,
     use_mouse: bool = True,
@@ -141,7 +140,10 @@ def click(
         error=(
             None
             if success
-            else {"code": "click_failed", "reason": "Element not found or not clickable"}
+            else {
+                "code": "click_failed",
+                "reason": "Element not found or not clickable",
+            }
         ),
     )
 
@@ -371,7 +373,10 @@ def click_rect(
             success=False,
             duration_ms=0,
             outcome="error",
-            error={"code": "invalid_rect", "reason": "Rectangle width and height must be positive"},
+            error={
+                "code": "invalid_rect",
+                "reason": "Rectangle width and height must be positive",
+            },
         )
 
     start_time = time.time()
@@ -426,6 +431,9 @@ def click_rect(
         error=(
             None
             if success
-            else {"code": "click_failed", "reason": error_msg if not success else "Click failed"}
+            else {
+                "code": "click_failed",
+                "reason": error_msg if not success else "Click failed",
+            }
         ),
     )

sentience/agent.py

@@ -5,7 +5,7 @@
 
 import re
 import time
-from typing import TYPE_CHECKING, Any, Dict, List, Optional, Union
+from typing import TYPE_CHECKING, Any, Optional
 
 from .actions import click, press, type_text
 from .base_agent import BaseAgent
@@ -93,8 +93,11 @@ def __init__(
         # Step counter for tracing
         self._step_count = 0
 
-    def act(
-        self, goal: str, max_retries: int = 2, snapshot_options: SnapshotOptions | None = None
+    def act(  # noqa: C901
+        self,
+        goal: str,
+        max_retries: int = 2,
+        snapshot_options: SnapshotOptions | None = None,
     ) -> AgentActionResult:
         """
         Execute a high-level goal using observe → think → act loop
@@ -116,9 +119,9 @@ def act(
             42
         """
         if self.verbose:
-            print(f"\n{'='*70}")
+            print(f"\n{'=' * 70}")
             print(f"🤖 Agent Goal: {goal}")
-            print(f"{'='*70}")
+            print(f"{'=' * 70}")
 
         # Generate step ID for tracing
         self._step_count += 1
@@ -234,7 +237,7 @@ def act(
                 self._track_tokens(goal, llm_response)
 
                 # Parse action from LLM response
-                action_str = llm_response.content.strip()
+                action_str = self._extract_action_from_response(llm_response.content)
 
                 # 4. EXECUTE: Parse and run action
                 result_dict = self._execute_action(action_str, filtered_snap)
@@ -392,6 +395,34 @@ def _build_context(self, snap: Snapshot, goal: str) -> str:
 
         return "\n".join(lines)
 
+    def _extract_action_from_response(self, response: str) -> str:
+        """
+        Extract action command from LLM response, handling cases where
+        the LLM adds extra explanation despite instructions.
+
+        Args:
+            response: Raw LLM response text
+
+        Returns:
+            Cleaned action command string
+        """
+        import re
+
+        # Remove markdown code blocks if present
+        response = re.sub(r"```[\w]*\n?", "", response)
+        response = response.strip()
+
+        # Try to find action patterns in the response
+        # Pattern matches: CLICK(123), TYPE(123, "text"), PRESS("key"), FINISH()
+        action_pattern = r'(CLICK\s*\(\s*\d+\s*\)|TYPE\s*\(\s*\d+\s*,\s*["\'].*?["\']\s*\)|PRESS\s*\(\s*["\'].*?["\']\s*\)|FINISH\s*\(\s*\))'
+
+        match = re.search(action_pattern, response, re.IGNORECASE)
+        if match:
+            return match.group(1)
+
+        # If no pattern match, return the original response (will likely fail parsing)
+        return response
+
     def _query_llm(self, dom_context: str, goal: str) -> LLMResponse:
         """
         Query LLM with standardized prompt template
@@ -415,23 +446,30 @@ def _query_llm(self, dom_context: str, goal: str) -> LLMResponse:
 - {{CLICKABLE}}: Element is clickable
 - {{color:X}}: Background color name
 
-RESPONSE FORMAT:
-Return ONLY the function call, no explanation or markdown.
-
-Available actions:
+CRITICAL RESPONSE FORMAT:
+You MUST respond with ONLY ONE of these exact action formats:
 - CLICK(id) - Click element by ID
 - TYPE(id, "text") - Type text into element
 - PRESS("key") - Press keyboard key (Enter, Escape, Tab, ArrowDown, etc)
 - FINISH() - Task complete
 
-Examples:
-- CLICK(42)
-- TYPE(15, "magic mouse")
-- PRESS("Enter")
-- FINISH()
+DO NOT include any explanation, reasoning, or natural language.
+DO NOT use markdown formatting or code blocks.
+DO NOT say "The next step is..." or anything similar.
+
+CORRECT Examples:
+CLICK(42)
+TYPE(15, "magic mouse")
+PRESS("Enter")
+FINISH()
+
+INCORRECT Examples (DO NOT DO THIS):
+"The next step is to click..."
+"I will type..."
+```CLICK(42)```
 """
 
-        user_prompt = "What is the next step to achieve the goal?"
+        user_prompt = "Return the single action command:"
 
         return self.llm.generate(system_prompt, user_prompt, temperature=0.0)
 
@@ -460,7 +498,9 @@ def _execute_action(self, action_str: str, snap: Snapshot) -> dict[str, Any]:
 
         # Parse TYPE(42, "hello world")
         elif match := re.match(
-            r'TYPE\s*\(\s*(\d+)\s*,\s*["\']([^"\']*)["\']\s*\)', action_str, re.IGNORECASE
+            r'TYPE\s*\(\s*(\d+)\s*,\s*["\']([^"\']*)["\']\s*\)',
+            action_str,
+            re.IGNORECASE,
         ):
             element_id = int(match.group(1))
             text = match.group(2)
@@ -486,7 +526,11 @@ def _execute_action(self, action_str: str, snap: Snapshot) -> dict[str, Any]:
 
         # Parse FINISH()
         elif re.match(r"FINISH\s*\(\s*\)", action_str, re.IGNORECASE):
-            return {"success": True, "action": "finish", "message": "Task marked as complete"}
+            return {
+                "success": True,
+                "action": "finish",
+                "message": "Task marked as complete",
+            }
 
         else:
             raise ValueError(

sentience/cli.py

@@ -104,7 +104,9 @@ def main():
         "--snapshots", action="store_true", help="Capture snapshots at each step"
     )
     record_parser.add_argument(
-        "--mask", action="append", help="Pattern to mask in recorded text (e.g., password)"
+        "--mask",
+        action="append",
+        help="Pattern to mask in recorded text (e.g., password)",
     )
     record_parser.set_defaults(func=cmd_record)