Update app.py

app.py

@@ -26,44 +26,26 @@ if not GROQ_API_KEY:
     raise RuntimeError("GROQ_API_KEY not set.")
 client = AsyncGroq(api_key=GROQ_API_KEY)
 
-# ---------- System prompt ----------
+# ---------- System prompt (simple) ----------
 SYSTEM_PROMPT = """
-You are an expert web automation agent. Your goal is to complete the user's task efficiently and robustly.
-
-**BEFORE ACTING, THINK:**
-1. What is the current page state? (URL, visible text, any popups/forms)
-2. What have I tried so far? (review history)
-3. What is the most logical next step? If I have a plan, follow it; if not, create one.
-4. If an action previously failed, try an alternative selector (e.g., use data-test, id, class, or text content).
-
-**Available actions:**
-- navigate, click, fill, select, scroll, press, hover, back, wait, screenshot, done
-- wait_for_selector: {"action":"wait_for_selector","selector":"...","timeout":5000}
-- extract_table: {"action":"extract_table","selector":"table","variable":"tableData"} (returns array of rows)
-- store_variable: {"action":"store_variable","variable":"name","value":"..."} (manual store)
-- log: {"action":"log","message":"..."} (debugging)
-- execute_script (advanced)
-
-**You MUST output a JSON object with two fields:**
-{
-  "thought": "Your reasoning about the current situation and what you plan to do.",
-  "action": { ... one of the above actions ... }
-}
-
-**Examples:**
-- Thought: "The login form is visible. I need to fill username and password and click submit."
-  Action: {"action":"fill","selector":"#username","value":"testuser"}
-
-- Thought: "The table of results is on the page. I'll extract it and store for later."
-  Action: {"action":"extract_table","selector":".results-table","variable":"results"}
-
-- Thought: "I've tried clicking the 'Next' button twice, but it's not working. Maybe the pagination uses a different selector. I'll try the 'next' link with text."
-  Action: {"action":"click","selector":"a:has-text('Next')"}
-
-**Always prefer `data-test` attributes, then `id`, then class, then text.**
-**If you get a CAPTCHA or a modal, try to close it or wait; if impossible, return 'CAPTCHA' in result.**
-**When extracting data, use `extract_table` for structured data, or `get_text` for single values.**
-**You can store values with `store_variable` and use them later in `fill` or other actions.**
+You are an autonomous web agent. You have access to a browser via Playwright.
+Your goal is to complete the user's task by issuing a sequence of actions.
+
+Available actions (return a JSON object with exactly one of these):
+- {"action": "navigate", "url": "https://..."}
+- {"action": "click", "selector": "css_selector"}
+- {"action": "fill", "selector": "css_selector", "value": "text"}
+- {"action": "select", "selector": "css_selector", "value": "option_text"}
+- {"action": "scroll", "amount": 300}   (positive = down, negative = up)
+- {"action": "press", "key": "Enter"}
+- {"action": "hover", "selector": "..."}
+- {"action": "back"}
+- {"action": "wait", "seconds": 2}
+- {"action": "screenshot"}
+- {"action": "done", "result": "final answer"}
+
+You will receive the current page URL and visible text.
+Think step by step. Only output the JSON action, nothing else.
 """
 
 # ---------- Session state ----------
@@ -75,9 +57,7 @@ class SessionState:
         self.steps = 0
         self.last_result = ""
         self.done = False
-        self.action_history = []  # list of {"thought": ..., "action": ..., "result": ...}
-        self.variables = {}       # store extracted text
-        self.plan = []           # optional step plan
+        self.action_history = []  # list of {"action": ..., "result": ...}
 
 sessions = {}
 
@@ -90,20 +70,16 @@ async def get_screenshot_b64(page):
     return base64.b64encode(screenshot_bytes).decode("utf-8")
 
 def build_history_text(session):
-    """Return a string of last 5 actions with thoughts and results."""
     history = session.action_history[-5:]
     if not history:
         return "No actions taken yet."
     lines = []
     for i, entry in enumerate(history):
-        thought = entry.get('thought', '')
-        act = entry.get('action', {})
-        result = entry.get('result', '')
-        lines.append(f"{i+1}. Thought: {thought}\n   Action: {json.dumps(act)}\n   Result: {result}")
+        lines.append(f"{i+1}. Action: {json.dumps(entry['action'])} → Result: {entry['result']}")
     return "\n".join(lines)
 
 async def ask_llm_for_action(session, user_instruction=None):
-    """Ask Groq for next action with thought. If user_instruction is given, treat as new task."""
+    """Ask Groq for the next action."""
     page = session.page
     url = page.url
     page_text = await get_page_text(page)
@@ -123,12 +99,12 @@ Action history (most recent first):
 Current visible text:
 {page_text}
 
-{user_instruction if user_instruction else "What is your next action? Include your thought and the action."}
+{user_instruction if user_instruction else "What is your next action?"}
 
-Output a JSON with "thought" and "action".
+Output a JSON action.
 """
     completion = await client.chat.completions.create(
-        model="mixtral-8x7b-32768",   # smarter model
+        model="llama-3.3-70b-versatile",
         messages=[
             {"role": "system", "content": SYSTEM_PROMPT},
             {"role": "user", "content": user_prompt}
@@ -138,76 +114,52 @@ Output a JSON with "thought" and "action".
     )
     raw = completion.choices[0].message.content
     try:
-        data = json.loads(raw)
-        # Ensure it has "thought" and "action"
-        if "thought" not in data:
-            data["thought"] = "No thought provided."
-        if "action" not in data:
-            data["action"] = {"action": "done", "result": "No action provided."}
-        return data
+        action = json.loads(raw)
     except json.JSONDecodeError:
         match = re.search(r'\{.*\}', raw, re.DOTALL)
         if match:
-            data = json.loads(match.group())
-            if "thought" not in data:
-                data["thought"] = "Extracted from raw."
-            if "action" not in data:
-                data["action"] = {"action": "done", "result": f"Invalid LLM output: {raw}"}
-            return data
+            action = json.loads(match.group())
         else:
-            return {
-                "thought": "Could not parse LLM output.",
-                "action": {"action": "done", "result": f"Invalid LLM output: {raw}"}
-            }
+            action = {"action": "done", "result": f"Invalid LLM output: {raw}"}
+    return action
 
 async def execute_one_action(session, override=None):
     """
     Execute one action.
     override can be:
       - None or "auto" → use LLM.
-      - a JSON string → treat as action directly (with optional thought).
-      - free text → treat as new instruction, then ask LLM.
+      - a JSON string → execute that action directly.
+      - free text → treat as new task, then ask LLM for action.
     """
     page = session.page
 
     if override and override.lower() != "auto":
         # Try to parse as JSON
         try:
-            data = json.loads(override)
-            # If it has "action" field, use it; if not, treat as a single action.
-            if "action" in data:
-                thought = data.get("thought", "User override.")
-                action = data["action"]
-            else:
-                # Assume the whole thing is the action
-                thought = "User override with JSON action."
-                action = data
+            action = json.loads(override)
+            # If it's valid JSON, treat as action
         except json.JSONDecodeError:
             # Not JSON – treat as new task description
             session.task = override
-            data = await ask_llm_for_action(session)
-            thought = data.get("thought", "Replanned after user instruction.")
-            action = data.get("action", {"action": "done", "result": "No action from LLM"})
+            action = await ask_llm_for_action(session)
     else:
         # Auto – ask LLM
-        data = await ask_llm_for_action(session)
-        thought = data.get("thought", "Auto decision.")
-        action = data.get("action", {"action": "done", "result": "No action from LLM"})
+        action = await ask_llm_for_action(session)
 
     act = action.get("action")
     result_text = ""
     screenshot_b64 = ""
 
-    # Execute the action
+    # Execute
     try:
         if act == "done":
             session.done = True
             session.last_result = action.get("result", "Task completed.")
             result_text = session.last_result
             screenshot_b64 = await get_screenshot_b64(page)
-            session.action_history.append({"thought": thought, "action": action, "result": result_text})
+            session.action_history.append({"action": action, "result": result_text})
             return {"done": True, "result": result_text, "screenshot": screenshot_b64,
-                    "action_taken": action, "thought": thought, "steps": session.steps}
+                    "action_taken": action, "steps": session.steps}
 
         elif act == "navigate":
             url = action.get("url")
@@ -224,12 +176,7 @@ async def execute_one_action(session, override=None):
                     await page.click(sel, timeout=5000)
                     result_text = f"Clicked {sel}"
                 except Exception as e:
-                    # Try alternative: find by text
-                    try:
-                        await page.click(f"text='{sel}'", timeout=5000)
-                        result_text = f"Clicked by text: {sel}"
-                    except:
-                        result_text = f"Click failed: {str(e)}"
+                    result_text = f"Click failed: {str(e)}"
             else:
                 result_text = "Error: missing 'selector'"
 
@@ -241,12 +188,7 @@ async def execute_one_action(session, override=None):
                     await page.fill(sel, val, timeout=5000)
                     result_text = f"Filled {sel} with '{val}'"
                 except Exception as e:
-                    # try by placeholder or label
-                    try:
-                        await page.fill(f"input[placeholder='{sel}']", val, timeout=5000)
-                        result_text = f"Filled by placeholder: {sel}"
-                    except:
-                        result_text = f"Fill failed: {str(e)}"
+                    result_text = f"Fill failed: {str(e)}"
             else:
                 result_text = "Error: missing 'selector' or 'value'"
 
@@ -295,80 +237,7 @@ async def execute_one_action(session, override=None):
             await asyncio.sleep(seconds)
             result_text = f"Waited {seconds}s"
 
-        elif act == "wait_for_selector":
-            sel = action.get("selector")
-            timeout = action.get("timeout", 5000)
-            if sel:
-                try:
-                    await page.wait_for_selector(sel, timeout=timeout)
-                    result_text = f"Selector '{sel}' appeared."
-                except Exception as e:
-                    result_text = f"Wait for selector failed: {str(e)}"
-            else:
-                result_text = "Error: missing 'selector'"
-
-        elif act == "extract_table":
-            sel = action.get("selector")
-            var = action.get("variable", "table_data")
-            if sel:
-                try:
-                    # Extract table as array of objects
-                    rows = await page.evaluate(f"""
-                        (sel) => {{
-                            const table = document.querySelector(sel);
-                            if (!table) return [];
-                            const headers = Array.from(table.querySelectorAll('thead th, tr:first-child th')).map(th => th.innerText.trim());
-                            const dataRows = table.querySelectorAll('tbody tr');
-                            const result = [];
-                            dataRows.forEach(row => {{
-                                const cells = row.querySelectorAll('td');
-                                if (cells.length) {{
-                                    const obj = {{}};
-                                    cells.forEach((cell, i) => {{
-                                        obj[headers[i] || `col_${{i}}`] = cell.innerText.trim();
-                                    }});
-                                    result.push(obj);
-                                }}
-                            }});
-                            return result;
-                        }}
-                    """, sel)
-                    session.variables[var] = rows
-                    result_text = f"Extracted table with {len(rows)} rows into '{var}'"
-                except Exception as e:
-                    result_text = f"Extract table failed: {str(e)}"
-            else:
-                result_text = "Error: missing 'selector'"
-
-        elif act == "store_variable":
-            var = action.get("variable")
-            val = action.get("value")
-            if var and val is not None:
-                session.variables[var] = val
-                result_text = f"Stored '{val}' in '{var}'"
-            else:
-                result_text = "Error: missing 'variable' or 'value'"
-
-        elif act == "log":
-            msg = action.get("message", "")
-            result_text = f"LOG: {msg}"
-            # Also print to console (will appear in Space logs)
-            print(f"AGENT LOG: {msg}")
-
-        elif act == "execute_script":
-            code = action.get("code")
-            if code:
-                try:
-                    result = await page.evaluate(code)
-                    session.variables["_last_script"] = result
-                    result_text = f"Script executed, result stored in _last_script"
-                except Exception as e:
-                    result_text = f"Script failed: {str(e)}"
-            else:
-                result_text = "Error: missing 'code'"
-
         elif act == "screenshot":
-            # just capture later
             result_text = "Screenshot captured"
 
         else:
@@ -378,7 +247,7 @@ async def execute_one_action(session, override=None):
         result_text = f"Action execution error: {str(e)}"
 
     # Record history
-    session.action_history.append({"thought": thought, "action": action, "result": result_text})
+    session.action_history.append({"action": action, "result": result_text})
     session.steps += 1
     session.last_result = result_text
 
@@ -386,22 +255,17 @@ async def execute_one_action(session, override=None):
     screenshot_b64 = await get_screenshot_b64(page)
 
     # Ask LLM for next action (to show the user)
-    next_data = await ask_llm_for_action(session)
-    next_thought = next_data.get("thought", "")
-    next_action = next_data.get("action", {"action": "wait", "seconds": 1})
+    next_action = await ask_llm_for_action(session)
 
     return {
         "done": session.done,
         "result": result_text,
         "screenshot": screenshot_b64,
         "steps": session.steps,
-        "thought": thought,
-        "action_taken": action,
-        "proposed_thought": next_thought,
         "proposed_next_action": next_action,
         "current_url": page.url,
-        "history": session.action_history[-5:],
-        "variables": session.variables
+        "action_taken": action,
+        "history": session.action_history[-5:]
     }
 
 # ---------- FastAPI Endpoints ----------
@@ -412,7 +276,6 @@ class StartRequest(BaseModel):
 class StartResponse(BaseModel):
     session_id: str
     screenshot: str
-    proposed_thought: str
     proposed_next_action: dict
     current_url: str
     steps: int
@@ -434,13 +297,12 @@ async def start_session(req: StartRequest):
         sessions[session_id] = session
 
         screenshot_b64 = await get_screenshot_b64(page)
-        data = await ask_llm_for_action(session)
+        next_action = await ask_llm_for_action(session)
 
         return StartResponse(
             session_id=session_id,
             screenshot=screenshot_b64,
-            proposed_thought=data.get("thought", ""),
-            proposed_next_action=data.get("action", {"action": "wait", "seconds": 1}),
+            proposed_next_action=next_action,
             current_url=page.url,
             steps=0
         )
@@ -456,13 +318,10 @@ class StepResponse(BaseModel):
     result: str
     screenshot: str
     steps: int
-    thought: str
-    action_taken: dict
-    proposed_thought: str
     proposed_next_action: dict
     current_url: str
+    action_taken: dict
     history: list
-    variables: dict
 
 @app.post("/step", response_model=StepResponse)
 async def step(req: StepRequest):
@@ -479,13 +338,10 @@ async def step(req: StepRequest):
             result=result["result"],
             screenshot=result["screenshot"],
             steps=result["steps"],
-            thought=result.get("thought", ""),
-            action_taken=result["action_taken"],
-            proposed_thought=result.get("proposed_thought", ""),
             proposed_next_action=result["proposed_next_action"],
             current_url=result["current_url"],
-            history=result["history"],
-            variables=result.get("variables", {})
+            action_taken=result["action_taken"],
+            history=result["history"]
         )
     except Exception as e:
         raise HTTPException(status_code=500, detail=str(e))
@@ -498,13 +354,13 @@ async def stop_session(session_id: str):
         return {"status": "stopped"}
     return {"status": "not_found"}
 
-# ---------- Serve HTML (with thought and variables) ----------
+# ---------- Serve HTML (with working override) ----------
 HTML_PAGE = """
 <!DOCTYPE html>
 <html>
 <head>
     <meta charset="UTF-8">
-    <title>Ultra‑Intelligent Agent</title>
+    <title>Interactive Agent</title>
     <style>
         * { box-sizing: border-box; margin: 0; padding: 0; }
         body { font-family: system-ui; background: #0b0e14; color: #e8edf5; padding: 20px; }
@@ -519,17 +375,14 @@ HTML_PAGE = """
         .status { margin: 10px 0; }
         .action-card { background: #11171f; padding: 15px; border-radius: 12px; margin: 10px 0; }
         pre { background: #0f141d; padding: 10px; border-radius: 8px; overflow: auto; max-height: 200px; }
-        .history-item { background: #0f141d; padding: 8px 12px; border-radius: 6px; margin: 6px 0; font-size: 13px; border-left: 3px solid #4f7cff; }
-        .history-item .thought { color: #ffd54f; }
-        .history-item .action { color: #4fc3f7; }
-        .history-item .result { color: #8bc34a; }
-        .variables-box { background: #0f141d; padding: 12px; border-radius: 12px; margin-top: 15px; }
-        .variables-box pre { margin: 5px 0; }
+        .history-item { background: #0f141d; padding: 5px 10px; border-radius: 6px; margin: 4px 0; font-size: 13px; }
+        .history-item .act { color: #4f7cff; }
+        .history-item .res { color: #8bc34a; }
     </style>
 </head>
 <body>
 <div class="container">
-    <h1>🧠 Ultra‑Intelligent Agent</h1>
+    <h1>🧠 Interactive Agent</h1>
     <p>Start a session, see the screenshot, and guide the agent step by step.</p>
     <hr style="border-color:#2f384a; margin:20px 0;">
 
@@ -547,11 +400,6 @@ HTML_PAGE = """
                 <h4>📜 Action History</h4>
                 <div id="historyList"></div>
             </div>
-
-            <div id="variablesBox" style="margin-top:15px; display:none;">
-                <h4>📦 Variables</h4>
-                <div class="variables-box"><pre id="variablesContent">{}</pre></div>
-            </div>
         </div>
 
         <div class="right">
@@ -560,18 +408,16 @@ HTML_PAGE = """
                 <div class="screenshot-box"><img id="screenshotImg" src="" alt="Screenshot" /></div>
             </div>
             <div id="actionContainer" style="display:none;">
-                <h3>🤖 Thought</h3>
-                <div style="background:#0f141d; padding:12px; border-radius:12px; margin-bottom:10px;"><span id="thoughtDisplay">—</span></div>
-                <h3>🎯 Proposed Next Action</h3>
+                <h3>🤖 Proposed Next Action</h3>
                 <pre id="proposedAction">—</pre>
                 <div>
                     <button class="btn" id="continueBtn">▶ Continue (auto)</button>
                     <button class="btn" id="overrideBtn">✏️ Override</button>
                 </div>
                 <div id="overrideArea" style="display:none; margin-top:10px;">
-                    <input type="text" id="overrideInput" placeholder='JSON with thought+action OR free text' style="width:70%;" />
+                    <input type="text" id="overrideInput" placeholder='JSON action OR new task description' style="width:70%;" />
                     <button class="btn" id="sendOverrideBtn">Send</button>
-                    <small style="display:block; color:#8892a8;">Example: {"thought":"I want to navigate","action":{"action":"navigate","url":"https://example.com"}}<br>Or free text: "now search for Python"</small>
+                    <small style="display:block; color:#8892a8;">Example JSON: {"action":"click","selector":"#login"}<br>Or free text: "now search for Python"</small>
                 </div>
                 <div id="resultBox" style="margin-top:15px; background:#11171f; padding:15px; border-radius:12px; display:none;">
                     <strong>✅ Last result:</strong> <span id="resultText">—</span>
@@ -592,15 +438,12 @@ HTML_PAGE = """
     const screenshotImg = document.getElementById('screenshotImg');
     const actionContainer = document.getElementById('actionContainer');
     const proposedAction = document.getElementById('proposedAction');
-    const thoughtDisplay = document.getElementById('thoughtDisplay');
     const resultBox = document.getElementById('resultBox');
     const resultText = document.getElementById('resultText');
     const actionTakenBox = document.getElementById('actionTakenBox');
     const actionTaken = document.getElementById('actionTaken');
     const historyBox = document.getElementById('historyBox');
     const historyList = document.getElementById('historyList');
-    const variablesBox = document.getElementById('variablesBox');
-    const variablesContent = document.getElementById('variablesContent');
 
     const startBtn = document.getElementById('startBtn');
     const stopBtn = document.getElementById('stopBtn');
@@ -620,10 +463,6 @@ HTML_PAGE = """
         screenshotImg.src = 'data:image/png;base64,' + b64;
     }
 
-    function showThought(thought) {
-        thoughtDisplay.textContent = thought || 'No thought provided.';
-    }
-
     function showProposedAction(action) {
         proposedAction.textContent = JSON.stringify(action, null, 2);
     }
@@ -635,29 +474,14 @@ HTML_PAGE = """
         }
         historyBox.style.display = 'block';
         historyList.innerHTML = history.map(h => 
-            `<div class="history-item">
-                <div class="thought">💭 ${h.thought || ''}</div>
-                <div class="action">⚡ ${JSON.stringify(h.action)}</div>
-                <div class="result">✅ ${h.result}</div>
-            </div>`
+            `<div class="history-item"><span class="act">${JSON.stringify(h.action)}</span> → <span class="res">${h.result}</span></div>`
         ).join('');
     }
 
-    function updateVariables(vars) {
-        if (!vars || Object.keys(vars).length === 0) {
-            variablesBox.style.display = 'none';
-            return;
-        }
-        variablesBox.style.display = 'block';
-        variablesContent.textContent = JSON.stringify(vars, null, 2);
-    }
-
     function updateUI(data) {
         showScreenshot(data.screenshot);
-        showThought(data.proposed_thought || '');
         showProposedAction(data.proposed_next_action);
         if (data.history) updateHistory(data.history);
-        if (data.variables) updateVariables(data.variables);
         if (data.action_taken) {
             actionTakenBox.style.display = 'block';
             actionTaken.textContent = JSON.stringify(data.action_taken, null, 2);
@@ -755,7 +579,6 @@ HTML_PAGE = """
             actionContainer.style.display = 'none';
             screenshotContainer.style.display = 'none';
             historyBox.style.display = 'none';
-            variablesBox.style.display = 'none';
         } catch (e) {
             setStatus('Stop error: ' + e.message, true);
         }