testronaut 1.0.8 → 1.0.9

package/core/agent.js

@@ -2,33 +2,6 @@ import 'dotenv/config';
 import { ChromeBrowser } from '../tools/chromeBrowser.js';
 import { turnLoop } from './turnLoop.js';
 
-
-// export async function runAgent(goal, maxTurns = 10) {
-//   console.log(`\n🚀 Running Agent with goal: ${goal}`);
-//   const browser = new ChromeBrowser();
-//   await browser.start();
-
-//   const messages = [
-//     {
-//       role: 'system',
-//       content: `
-//         You are an autonomous web agent. Use function calls to complete the user's goal.
-//         If you are unsure of the selectors for inputs or buttons, call 'get_dom' to retrieve page HTML,
-//         analyze it, then make your best guess based on labels, names, types, and placeholder values.
-//         After completing the goal, respond with a final plain-text message starting with SUCCESS or FAILURE.
-//       `.trim(),
-//     },
-//     { role: 'user', content: goal },
-//   ];
-
-//   try {
-//     const result = await turnLoop(browser, messages, maxTurns);
-//     !result && console.log('🛑 Agent ran out of turns.')
-//   } finally {
-//     await browser.close();
-//   }
-// }
-
 export async function runAgent(goals, missionName, maxTurns = 20, ) {
   const browser = new ChromeBrowser();
   await browser.start();
@@ -86,13 +59,3 @@ export async function runAgent(goals, missionName, maxTurns = 20, ) {
     await browser.close();
   }
 }
-
-
-// (async () => {
-//   await runAgent(`
-//     Visit ${process.env.URL}.
-//     Log in using ${process.env.USERNAME} and password ${process.env.PASSWORD}.
-//     After clicking the login button, use 'check_text' to verify that "${process.env.AFTER_LOGIN_CHECK}" appears on the page.
-//     If found, report SUCCESS. Otherwise, report FAILURE.
-//   `);
-// })();

package/core/turnLoop.js

@@ -18,6 +18,7 @@ const openai = new OpenAI({ apiKey: process.env.OPENAI_API_KEY });
 
 let totalTokensUsed = 0;
 let turnTimestamps = [];
+let shouldBackoff;
 
 const pushDOMAssistant = async (browser, messages, agentMemory, { skipIfLastTool } = {}) => {
   if (skipIfLastTool && skipIfLastTool.includes(messages.at(-1)?.name)) {
@@ -61,48 +62,22 @@ export const turnLoop = async (browser, messages, maxTurns, currentTurn = 0, ret
   
   for (let turn = currentTurn; turn < maxTurns; turn++) {
     console.log(`\n🔄 Turn ${turn + 1}/${maxTurns}`);
-    // console.log("current tools", toolsSchema);
-
-    // const now = Date.now();
-    // turnTimestamps = turnTimestamps.filter(ts => now - ts < 60000);
+    let response;
+    try {
 
-    // Recalculate the rolling window before checking cooldown
-    const pruned = pruneOldTokenUsage(turnTimestamps);
-    turnTimestamps = pruned.turnTimestamps;
-    totalTokensUsed = pruned.totalTokensUsed;
+      // Recalculate the rolling window before checking cooldown
+      ({ turnTimestamps, totalTokensUsed } = pruneOldTokenUsage(turnTimestamps));
 
-    // Cool off check
-    const cooldownResult = await tokenUseCoolOff(totalTokensUsed, turnTimestamps);
-    totalTokensUsed = cooldownResult.totalTokensUsed;
-    turnTimestamps = cooldownResult.turnTimestamps;
-    if (cooldownResult.shouldBackoff) {
-      return await turnLoop(browser, messages, maxTurns, turn);
-    }
-    // if (await tokenUseCoolOff(totalTokensUsed, turnTimestamps)) {return await turnLoop(browser, messages, maxTurns, turn)}
+      // Cool off check
+      ({ totalTokensUsed, turnTimestamps, shouldBackoff } = await tokenUseCoolOff(totalTokensUsed, turnTimestamps));
+      if(shouldBackoff) return await turnLoop(browser, messages, maxTurns, turn);
 
-    const isValid = validateAndInsertMissingToolResponses(messages, {
-      insertPlaceholders: true, // safe fallback behavior
-    });
-    
-    if (!isValid) {
-      console.error('❌ Tool call structure invalid and no placeholders inserted.');
-      return false;
-    }
-    // console.log(JSON.stringify(messages, null, 2));
+      //Tool Validator
+      if (!validateAndInsertMissingToolResponses(messages, { insertPlaceholders: true })) {
+        console.error('❌ Tool call structure invalid and no placeholders inserted.');
+        return false;
+      }
 
-    let response;
-    try {
-      // console.log(
-      //   '🔍 Verifying message structure before OpenAI call:\n',
-      //   messages.map(m => ({
-      //     role: m.role,
-      //     name: m.name,
-      //     tool_call_id: m.tool_call_id,
-      //     tool_calls: m.tool_calls?.map(t => t.id),
-      //     content: m.content?.slice?.(0, 100), // truncate for readability
-      //   }))
-      // );
-      
       const unrespondedCalls = messages.filter(
         (msg, i) =>
           msg.role === 'assistant' &&
@@ -140,8 +115,8 @@ export const turnLoop = async (browser, messages, maxTurns, currentTurn = 0, ret
         return await turnLoop(browser, messages, maxTurns, turn, retryCount + 1);
       } else if (err.status === 400) {
         console.error('❌ Bad request:', err.message);
-        console.log("current tools", toolsSchema);
-        console.log("messages: ", messages)
+        // console.log("current tools", toolsSchema);
+        // console.log("messages: ", messages)
         return false;
       } else {
         throw err;
@@ -149,40 +124,23 @@ export const turnLoop = async (browser, messages, maxTurns, currentTurn = 0, ret
     }
     console.log('Response received from OpenAI');
     const usage = response.usage;
-    // if (usage) {
-    //   console.log(`📊 Token Usage This Turn → Prompt: ${usage.prompt_tokens}, Completion: ${usage.completion_tokens}, Total: ${usage.total_tokens}`);
-    //   totalTokensUsed += usage?.total_tokens || 0;
-    //   turnTimestamps.push(now);
-    //   console.log(`📈 Running Total Tokens Used: ${totalTokensUsed}`);
-    // }
     if (usage) {
       const tokensUsed = usage.total_tokens || 0;
       console.log(`📊 Token Usage This Turn → Total: ${tokensUsed}`);
       recordTokenUsage(turnTimestamps, tokensUsed);
-      const pruned = pruneOldTokenUsage(turnTimestamps);
-      turnTimestamps = pruned.turnTimestamps;
-      totalTokensUsed = pruned.totalTokensUsed;
+      ({ turnTimestamps, totalTokensUsed } = pruneOldTokenUsage(turnTimestamps));
+
       console.log(`📈 Running Total Tokens Used (Rolling 60s): ${totalTokensUsed}`);
     }
     
     const msg = response.choices[0].message;
 
-    // if (!msg.tool_calls?.length) {
-    //   console.error('❌ Assistant message had empty or missing tool_calls array');
-    //   return false;
-    // }   
-
     if (msg.tool_calls?.length) {
       console.log('Processing tool calls...');
-      // messages.push(msg);
       const toolResponses = [];
       let lastToolName = null;
 
       for (const call of msg.tool_calls) {
-        const toolCallId = call.id;
-        // const hasResponse = messages.some(m =>
-        //   m.role === 'tool' && m.tool_call_id === toolCallId
-        // );
         const fnName = call.function.name;
         lastToolName = fnName;
         const args = JSON.parse(call.function.arguments || '{}');
@@ -194,15 +152,6 @@ export const turnLoop = async (browser, messages, maxTurns, currentTurn = 0, ret
           // console.log("agentMemory: ", agentMemory);
           result = await CHROME_TOOL_MAP[fnName](browser, args, agentMemory);
 
-          // Smart DOM-ready wait after major actions like click
-          // if (['click', 'click_text', 'fill'].includes(fnName)) {
-          //   try {
-          //     await browser.page.waitForLoadState?.('domcontentloaded', { timeout: 5000 });
-          //     await browser.page.waitForTimeout?.(2000); // allow animations/navigation
-          //   } catch (err) {
-          //     console.warn(`⚠️ DOM readiness wait failed after ${fnName}: ${err.message}`);
-          //   }
-          // }
           if (typeof result !== 'string') {
             result = JSON.stringify(result ?? '');
           }
@@ -211,26 +160,9 @@ export const turnLoop = async (browser, messages, maxTurns, currentTurn = 0, ret
           result = errorMessage;
         }
 
-        // 🔍 Diagnostic: log whether this tool call succeeded or failed
         console.log(`[tool ] ← ${fnName} result:`, errorMessage ? '❌ Failed' : '✅ Success');
-
         const truncated = result.length > 100 ? result.slice(0, 100) + '…' : result;
-
-        // if (result.length > 100) {
-        //   // const logFile = `missions/mission_reports/tool-result-${Date.now()}.log`;
-        //   // fs.writeFileSync(logFile, result);
-        //   console.log(`[tool ] ← ${truncated} (full output written to ${logFile})`);
-        // } else {
-          console.log(`[tool ] ← ${truncated}`);
-        // }
-
-        // messages.push({
-        //   role: 'tool',
-        //   tool_call_id: call.id,
-        //   name: fnName,
-        //   type: 'function',
-        //   content: result,
-        // });
+        console.log(`[tool ] ← ${truncated}`);
 
         toolResponses.push({
           role: 'tool',
@@ -249,7 +181,7 @@ export const turnLoop = async (browser, messages, maxTurns, currentTurn = 0, ret
           const domHtml = await CHROME_TOOL_MAP.get_dom(browser, {
             limit: 100000,
             exclude: true,
-            focus: [], // or try focusing on ['main'] to keep it light
+            focus: [],
           }, agentMemory);
           await tokenEstimate('gpt-4o', domHtml);
           // fs.writeFileSync(`missions/mission_reports/debug-expanded-${Date.now()}.html`, domHtml);
@@ -258,18 +190,9 @@ export const turnLoop = async (browser, messages, maxTurns, currentTurn = 0, ret
           });
           console.log(`[auto] → DOM size after ${fnName}: ${domHtml.length} chars`);
         }
-
-
-        
-        // if (!hasResponse) {
-          //   console.error(`❌ Tool call ${toolCallId} for ${call.function.name} was not followed by a response`);
-          //   return false;
-          // }
-        }
         
-      // Only after ALL tool responses are ready
+      }
       messages.push(msg, ...toolResponses);
-
       continue;
     }
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "testronaut",
-  "version": "1.0.8",
+  "version": "1.0.9",
   "main": "./runner/testronaut.js",
   "bin": {
     "testronaut": "./bin/cli.js"
@@ -25,5 +25,12 @@
     "openai": "^5.10.2",
     "playwright": "^1.54.1",
     "prompts": "^2.4.2"
-  }
+  },
+  "keywords": [
+    "cli",
+    "testing",
+    "automation",
+    "node",
+    "qa"
+  ]
 }

package/runner/missionLaunch.js

@@ -6,4 +6,4 @@ mission('Collegium Dashboard Login Test', () => {
   objective('should land on dashboard after login', dashboardWorkflow);
 });
 
-await launch(loginWorkflow); // Setup: login
+await launch(loginWorkflow);

package/runner/testronaut.js

@@ -16,38 +16,6 @@ export async function launch(preMissionSetup) {
   await runSuite(objectives, preMissionSetup);
 }
 
-
-// export async function runMissions({ preMission, mission, postMission }) {
-//   const goals = [];
-
-//   if (preMission) {
-//     console.log('\n🧭 Running pre-mission setup:', preMission);
-//     goals.push({ goal: preMission, label: 'pre-mission' });
-//   }
-
-//   if (mission) {
-//     console.log('\n🧭 Running main mission:', mission);
-//     goals.push({ goal: mission, label: 'mission' });
-//   }
-
-//   if (postMission) {
-//     console.log('\n🧭 Running post-mission cleanup:', postMission);
-//     goals.push({ goal: postMission, label: 'post-mission' });
-//   }
-
-// console.log('\n🧭 Running mission flow:\n', goals);
-
-//   // for (const { goal, label } of goals) {
-//     // console.log(`\n🧭 Running ${label}:\n`);
-//   const success = await runAgent(goals);
-//   if (!success) {
-//     console.log(`❌ Aborting after failed goal.`);
-//     return;
-//   }
-//   // }
-//   console.log('\n✅ Mission flow complete.');
-// }
-
 export async function runMissions({ preMission, mission, postMission }, missionName) {
   const normalizeToArray = (x) => (Array.isArray(x) ? x : x ? [x] : []);
   
@@ -58,17 +26,14 @@ export async function runMissions({ preMission, mission, postMission }, missionN
   const goals = [];
 
   if (pre.length) {
-    // console.log('\n🧭 Running pre-mission setup:', pre);
     goals.push(...pre.map(fn => ({ goal: fn, label: 'pre-mission' })));
   }
 
   if (main.length) {
-    // console.log('\n🧭 Running main mission(s):', main);
     goals.push(...main.map(fn => ({ goal: fn, label: 'mission' })));
   }
 
   if (post.length) {
-    // console.log('\n🧭 Running post-mission cleanup:', post);
     goals.push(...post.map(fn => ({ goal: fn, label: 'post-mission' })));
   }
 
@@ -84,5 +49,4 @@ export async function runMissions({ preMission, mission, postMission }, missionN
   }
   console.log('\n✅ Mission flow complete.');
   return success;
-
 }

package/tools/chromeBrowser.js

@@ -92,7 +92,7 @@ export class ChromeBrowser {
         await el.click();
         await this.page.screenshot({ path: `missions/mission_reports/screenshot_after_click_${Date.now()}.png` });
         await this.page.waitForTimeout(5000);
-        // await this.page.screenshot({ path: `screenshot_after_waiting_click_${Date.now()}.png` });
+        await this.page.screenshot({ path: `screenshot_after_waiting_click_${Date.now()}.png` });
         return `clicked text "${text}"`;
         
       }
@@ -147,11 +147,6 @@ export class ChromeBrowser {
     await this.page.waitForSelector(selector, { timeout: 30_000 });
     await this.page.click(selector);
     await this.page.waitForTimeout(delayMs);
-
-    const expandedDom = await this.page.content();
-    // const filename = `missions/mission_reports/expand_menu_dump_${Date.now()}.html`;
-    // fs.writeFileSync(filename, expandedDom);
-    // console.log(`[expand_menu] → Dumped DOM after expand click to ${filename}`);
     await this.page.screenshot({ path: `missions/mission_reports/screenshot_expand_menu_${Date.now()}.png` });
 
     return `expanded menu using ${selector}`;