@zibby/core 0.1.21 → 0.1.22

package/package.json

@@ -1,31 +1,47 @@
 {
   "name": "@zibby/core",
-  "version": "0.1.21",
+  "version": "0.1.22",
   "description": "Core test automation engine with multi-agent and multi-MCP support",
   "type": "module",
-  "main": "src/index.js",
+  "main": "dist/index.js",
   "exports": {
-    ".": "./src/index.js",
-    "./sync": "./src/sync/index.js",
-    "./framework/graph.js": "./src/framework/graph.js",
-    "./framework/state.js": "./src/framework/state.js",
-    "./framework/node.js": "./src/framework/node.js",
-    "./framework/graph-compiler.js": "./src/framework/graph-compiler.js",
-    "./framework/node-registry.js": "./src/framework/node-registry.js",
-    "./framework/skill-registry.js": "./src/framework/skill-registry.js",
-    "./framework/tool-resolver.js": "./src/framework/tool-resolver.js",
-    "./framework/function-bridge.js": "./src/framework/function-bridge.js",
-    "./framework/function-skill-registry.js": "./src/framework/function-skill-registry.js",
-    "./framework/code-generator.js": "./src/framework/code-generator.js",
-    "./utils/ast-utils.js": "./src/utils/ast-utils.js",
-    "./utils/mcp-config-writer.js": "./src/utils/mcp-config-writer.js",
-    "./utils/node-schema-parser.js": "./src/utils/node-schema-parser.js",
+    ".": "./dist/index.js",
+    "./backend-client.js": "./dist/backend-client.js",
+    "./sync": "./dist/sync/index.js",
+    "./framework/graph.js": "./dist/framework/graph.js",
+    "./framework/state.js": "./dist/framework/state.js",
+    "./framework/node.js": "./dist/framework/node.js",
+    "./framework/graph-compiler.js": "./dist/framework/graph-compiler.js",
+    "./framework/node-registry.js": "./dist/framework/node-registry.js",
+    "./framework/skill-registry.js": "./dist/framework/skill-registry.js",
+    "./framework/tool-resolver.js": "./dist/framework/tool-resolver.js",
+    "./framework/function-bridge.js": "./dist/framework/function-bridge.js",
+    "./framework/function-skill-registry.js": "./dist/framework/function-skill-registry.js",
+    "./framework/code-generator.js": "./dist/framework/code-generator.js",
+    "./utils/ast-utils.js": "./dist/utils/ast-utils.js",
+    "./utils/mcp-config-writer.js": "./dist/utils/mcp-config-writer.js",
+    "./utils/node-schema-parser.js": "./dist/utils/node-schema-parser.js",
+    "./utils/parallel-config.js": "./dist/utils/parallel-config.js",
+    "./utils/run-registry.js": "./dist/utils/run-registry.js",
+    "./utils/run-index-merge.js": "./dist/utils/run-index-merge.js",
+    "./utils/run-index-post-cli.js": "./dist/utils/run-index-post-cli.js",
+    "./utils/run-state-session.js": "./dist/utils/run-state-session.js",
+    "./utils/session-state-live-runs.js": "./dist/utils/session-state-live-runs.js",
+    "./utils/mission-control-from-run-states.js": "./dist/utils/mission-control-from-run-states.js",
+    "./utils/live-frame-discovery.js": "./dist/utils/live-frame-discovery.js",
+    "./utils/run-capacity-coordinator.js": "./dist/utils/run-capacity-coordinator.js",
+    "./utils/run-capacity-queue.js": "./dist/utils/run-capacity-queue.js",
+    "./utils/cursor-mcp-isolated-home.js": "./dist/utils/cursor-mcp-isolated-home.js",
+    "./constants/zibby-scratch.js": "./dist/constants/zibby-scratch.js",
+    "./templates/browser-test-automation/pipeline-ids.js": "./templates/browser-test-automation/pipeline-ids.js",
+    "./templates/browser-test-automation/run-index.mjs": "./templates/browser-test-automation/run-index.mjs",
     "./templates/register-nodes.js": "./templates/register-nodes.js",
     "./templates": "./templates/index.js",
     "./templates/*": "./templates/*",
     "./package.json": "./package.json"
   },
   "scripts": {
+    "build": "node ../scripts/build.mjs",
     "test": "vitest run --exclude '**/memory/**'",
     "test:watch": "vitest",
     "test:state-schema": "vitest run src/framework/__tests__/state-schema.test.js",
@@ -52,15 +68,11 @@
     "url": "https://github.com/ZibbyHQ/zibby-agent/issues"
   },
   "files": [
-    "src/",
-    "!src/**/__tests__/",
-    "!src/**/*.test.js",
-    "!src/**/*.spec.js",
+    "dist/",
     "templates/",
     "!templates/**/__tests__/",
     "!templates/**/*.test.js",
     "!templates/**/*.spec.js",
-    "scripts/",
     "README.md",
     "LICENSE"
   ],
@@ -68,26 +80,28 @@
     "node": ">=18.0.0"
   },
   "dependencies": {
-    "@anthropic-ai/claude-agent-sdk": "^0.2.50",
-    "@anthropic-ai/sdk": "^0.71.2",
-    "@modelcontextprotocol/sdk": "^1.27.1",
-    "@playwright/mcp": "^0.0.54",
+    "@anthropic-ai/claude-agent-sdk": "^0.2.104",
+    "@anthropic-ai/sdk": "^0.88.0",
+    "@modelcontextprotocol/sdk": "^1.29.0",
+    "@openai/codex-sdk": "^0.120.0",
+    "@playwright/mcp": "^0.0.70",
+    "@zibby/mcp-browser": "^0.1.6",
     "acorn": "^8.15.0",
-    "acorn-walk": "^8.3.4",
-    "axios": "^1.13.3",
+    "acorn-walk": "^8.3.5",
+    "axios": "^1.15.0",
     "chalk": "^5.3.0",
-    "dotenv": "^16.4.0",
-    "handlebars": "^4.7.8",
-    "zod": "^3.23.0",
-    "@zibby/mcp-browser": "^0.1.0",
-    "zod-to-json-schema": "^3.25.1"
+    "dotenv": "^17.4.1",
+    "handlebars": "^4.7.9",
+    "zod": "^4.3.6",
+    "zod-to-json-schema": "^3.25.2"
   },
   "peerDependencies": {
     "@playwright/test": ">=1.49.0",
     "playwright": ">=1.49.0"
   },
   "devDependencies": {
-    "@playwright/test": "^1.49.0",
-    "vitest": "^4.0.18"
+    "@playwright/test": "^1.59.1",
+    "esbuild": "^0.28.0",
+    "vitest": "^4.1.4"
   }
 }

package/templates/browser-test-automation/README.md

@@ -2,7 +2,7 @@
 
 This is YOUR workflow graph. You can customize it however you want!
 
-Works with **Claude** or **Cursor** agents (configured in `.zibby.config.js`).
+Works with **Claude** or **Cursor** agents (configured in `.zibby.config.mjs`).
 
 ## Default Flow
 
@@ -86,18 +86,40 @@ graph.addParallelEdges('verify_script', [
 
 ## Configuration
 
-Edit `.zibby/config.js` to set default agents per node:
+Edit `.zibby.config.mjs` to set your default agent and optional per-node model overrides:
 
 ```javascript
 export default {
-  agents: {
-    execute_live: { type: 'cursor' },
-    verify_script: { type: 'deepseek', model: 'deepseek-coder' },
-    update_jira: { type: 'ollama', model: 'llama3' }
-  }
+  agent: {
+    cursor: { model: 'auto' }, // or claude: { model: 'auto' }
+    strictMode: false,
+  },
+  models: {
+    default: 'auto',
+    execute_live: 'auto',
+    generate_script: 'auto',
+  },
 };
 ```
 
+## Studio / Scripts tab (code discovery)
+
+Runs write `generate_script/result.json` with a `scriptPath` (often under your repo `tests/`). After the graph finishes, **`BrowserTestResultHandler.ensureStudioCodegenMirror`** copies that file into the session folder under stable names so tools don’t need Studio running at generation time:
+
+| File (under `.zibby/output/sessions/<sessionId>/generate_script/`) | Role |
+|---------------------------------------------------------------------|------|
+| `generated-test.spec.js` | Playwright (`.js`) |
+| `playwright.spec.ts` | Playwright (`.ts` / `.tsx` source) |
+| `test.selenium.py` | Selenium |
+
+**Electron Studio** resolves these via `discoverCodegenArtifactsElectron` (after `session/codegen/`).
+
+**Web Studio** (`VITE_STUDIO_API_ORIGIN`, e.g. `:3847`) should implement `GET /api/sessions/:id/codegen/playwright` (and `/selenium`) by reading, in order:
+
+1. `sessions/<id>/codegen/` legacy JIT names (`test.spec.ts`, `generated-test.spec.js`, …)
+2. **`sessions/<id>/generate_script/`** canonical names above
+3. `scriptPath` from `generate_script/result.json` (resolve relative to session / `cwd` from session meta)
+
 ## Documentation
 
 - [Full Graph Framework Design](../../docs/GRAPH_FRAMEWORK_DESIGN.md)

package/templates/browser-test-automation/chat.mjs

@@ -0,0 +1,36 @@
+/**
+ * Zibby Chat Agent
+ *
+ * Interactive conversational node that acts as the default entry point
+ * when users type `zibby` with no subcommand.
+ *
+ * This is a plain chat bot — no MCP servers, no middleware, no structured output.
+ * Just streamed text conversation with the AI agent.
+ *
+ * The skill-installer skill injects its promptFragment so the LLM knows which
+ * skills are available and can install/uninstall them via natural conversation.
+ * Users can customize this file after `zibby init` copies it to .zibby/chat.mjs
+ */
+
+import { SKILLS } from '@zibby/core';
+
+export const CHAT_CONFIG = {
+  name: 'zibby_chat',
+  skills: [SKILLS.CORE_TOOLS, SKILLS.SKILL_INSTALLER, SKILLS.CHAT_MEMORY],
+  timeout: 0,
+
+  systemPrompt: `You are Zibby, a helpful AI assistant. Capabilities come from installed skills.
+
+## How you work
+1. When you need data, call tools. You can chain up to 5 calls per turn.
+2. After each tool result, decide: "Would I be embarrassed to give this answer to a coworker?" If yes, call another tool.
+3. Only respond once you have something genuinely useful.
+4. Never claim you did something without actually calling the tool.
+5. After EVERY response, self-evaluate: is the user's goal fully achieved? Is anything still pending or running? If yes, DO NOT ASK — autonomously poll: call wait (you decide how long), then check status, then respond with an update. Repeat until done or the user interrupts.
+
+## How you talk
+- Talk like a teammate in Slack, not a report generator.
+- Summarize and paraphrase. Never copy-paste field values or list raw steps verbatim.
+- Short paragraphs, not numbered lists (unless the user specifically asks for steps).
+- Match the user's tone and energy. Be concise.`,
+};

package/templates/browser-test-automation/graph.mjs

@@ -13,17 +13,9 @@ import {
 } from './nodes/index.mjs';
 import { BrowserTestResultHandler } from './result-handler.mjs';
 
-let memoryMiddleware = null;
-try {
-  const mem = await import('@zibby/memory');
-  memoryMiddleware = mem.memoryMiddleware;
-} catch { /* @zibby/memory not installed */ }
-
 export class BrowserTestAutomationAgent extends WorkflowAgent {
   buildGraph() {
-    const graph = new WorkflowGraph({
-      middleware: [memoryMiddleware?.()].filter(Boolean),
-    });
+    const graph = new WorkflowGraph();
 
     graph.addNode('preflight', preflightNode);
     graph.addNode('execute_live', executeLiveNode);
@@ -46,6 +38,10 @@ export class BrowserTestAutomationAgent extends WorkflowAgent {
     const cwd = result.state.cwd || process.cwd();
     BrowserTestResultHandler.saveTitle(result, cwd);
     await BrowserTestResultHandler.saveExecutionData(result);
+    BrowserTestResultHandler.ensureStudioCodegenMirror(
+      result.state?.sessionPath,
+      result.state?.cwd || cwd,
+    );
 
     // Memory end-run hook (if @zibby/memory is installed)
     try {

package/templates/browser-test-automation/nodes/execute-live.mjs

@@ -6,7 +6,7 @@
  * Configuration:
  * - capabilities: Declares ['browser'] — framework injects the appropriate MCP server
  * - outputSchema: Structured JSON with execution results, actions, assertions
- * - Model: Configured in .zibby.config.js → agent.claude.model or agent.cursor.model
+ * - Model: Configured in .zibby.config.mjs → agent.claude.model or agent.cursor.model
  */
 
 import { z, SKILLS } from '@zibby/core';
@@ -31,14 +31,6 @@ ${ctx.env ? `Environment Config:\n${JSON.stringify(ctx.env, null, 2)}\n` : ''}
     
     return `⚠️  CRITICAL: At the END, output ONLY the JSON object. NO explanations after the JSON.
 
-Execute this test using ONLY mcp_playwright-official_browser_* tools.
-
-🚨 HONESTY REQUIREMENT (STRICT):
-- If you DO NOT have access to browser tools → return {"success": false, "steps": [], "browserClosed": false, "notes": "No browser tools available"}
-- DO NOT hallucinate or pretend you executed the test
-- DO NOT return success: true unless you ACTUALLY called browser tools
-- BE HONEST - it's better to admit you can't do it than to lie
-
 🎯 YOUR GOAL: Execute the test steps and collect evidence for script generation.
    You don't need perfect verification - just capture the key actions and results.
    The next node will generate the actual test script from your execution.
@@ -151,46 +143,6 @@ IMPORTANT for 'evidenceScreenshots' (array) - OPTIONAL:
 - If you take screenshots, use descriptive filenames
 - Filename pattern: "{step-number}-{action-or-state}.png"
 - Keep it minimal - test execution is more important than documentation
-
-════════════════════════════════════════════════════════════
-🚨 CRITICAL JSON OUTPUT RULES 🚨
-
-YOU MUST OUTPUT JSON USING ONE OF THESE TWO FORMATS:
-
-✅ FORMAT 1 (BEST - Use This!):
-Think/plan/explain first, THEN output ONLY JSON:
-
-I'll navigate to the login page and fill the form...
-[... use browser tools ...]
-[... complete test execution ...]
-
-{"success": true, "steps": [...], "browserClosed": true}
-
-NO TEXT AFTER THE JSON! Stop immediately after }.
-
-✅ FORMAT 2 (If you need to explain after):
-Use delimiters to separate JSON from explanations:
-
-I'm executing the test now...
-[... use browser tools ...]
-
-===JSON_START===
-{"success": true, "steps": [...], "browserClosed": true}
-===JSON_END===
-
-Now let me explain what happened...
-
-❌ WRONG - DO NOT DO THIS:
-{"success": true, "steps": [...]} followed by more explanations
-
-❌ WRONG - DO NOT STREAM JSON LETTER BY LETTER:
-The test completed successfully, here's the result: { "success": t
-
-✅ CORRECT - Output complete JSON in one block:
-{"success": true, "steps": ["step 1", "step 2"], "browserClosed": true}
-
-REMEMBER: After the final }, you MUST STOP or use ===JSON_END===
-════════════════════════════════════════════════════════════
 `;
   },
   
@@ -202,11 +154,27 @@ REMEMBER: After the final }, you MUST STOP or use ===JSON_END===
       .describe('Array of test steps executed'),
     
     finalUrl: z.string()
-      .optional()
+      .nullish()
       .describe('Final URL after test execution'),
     
-    actions: z.array(z.any())
-      .optional()
+    actions: z.array(z.object({
+      type: z.string()
+        .describe('Action type: navigate, click, fill, type, select, keypress, hover, drag'),
+      description: z.string()
+        .describe('Human-readable description of the action'),
+      reasoning: z.string().nullish()
+        .describe('Why this action was performed'),
+      selectors: z.object({
+        role: z.object({
+          role: z.string().describe('ARIA role (e.g. button, link, textbox, generic)'),
+          name: z.string().nullish().describe('Accessible name of the element')
+        }).nullish().describe('Role-based selector for fallback matching')
+      }).nullish()
+        .describe('Element selectors captured during the action'),
+      value: z.string().nullish()
+        .describe('Value entered for fill/type actions')
+    }))
+      .nullish()
       .describe('Detailed array of actions performed with descriptions and reasoning'),
     
     assertions: z.array(z.object({
@@ -217,14 +185,18 @@ REMEMBER: After the final }, you MUST STOP or use ===JSON_END===
       verifiedAfterAction: z.number()
         .describe('Index of the action after which this was verified (0-based, matches actions array index) - REQUIRED'),
       evidence: z.string()
-        .optional()
+        .nullish()
         .describe('Brief evidence of what was observed')
     }))
-      .optional()
+      .nullish()
       .describe('Array of assertions made during test'),
     
-    waits: z.array(z.any())
-      .optional()
+    waits: z.array(z.object({
+      description: z.string().describe('What the wait is for'),
+      duration: z.number().nullish().describe('Wait duration in milliseconds'),
+      condition: z.string().nullish().describe('Wait condition expression')
+    }))
+      .nullish()
       .describe('Array of waits needed for proper test execution'),
     
     evidenceScreenshots: z.array(z.object({
@@ -237,14 +209,14 @@ REMEMBER: After the final }, you MUST STOP or use ===JSON_END===
       verdict: z.enum(['pass', 'fail', 'info'])
         .describe('Test verdict: pass/fail for validation points, info for checkpoints')
     }))
-      .optional()
+      .nullish()
       .describe('Array of screenshots taken at key validation points throughout the test'),
     
     browserClosed: z.boolean()
       .describe('Whether the browser was properly closed (should always be true)'),
     
     notes: z.string()
-      .optional()
+      .nullish()
       .describe('Additional notes or observations. REQUIRED when success=false to explain why test failed or could not execute')
   })
 };

package/templates/browser-test-automation/nodes/generate-script.mjs

@@ -1,5 +1,5 @@
-import { z } from '@zibby/core';
-import { formatRecordedActions, formatAssertionsWithResults } from './utils.mjs';
+import { z, SKILLS } from '@zibby/core';
+import { formatRecordedActions, formatAssertionsWithResults, loadRecordedActions, detectLoginPattern, formatSetupHint } from './utils.mjs';
 
 const GenerateScriptOutputSchema = z.object({
   success: z.boolean(),
@@ -9,8 +9,9 @@ const GenerateScriptOutputSchema = z.object({
 
 export const generateScriptNode = {
   name: 'generate_script',
+  skills: [SKILLS.MEMORY],
   outputSchema: GenerateScriptOutputSchema,
-  timeout: 360000,
+  timeout: 1200000,
 
   prompt: (state) => {
     const exec = state.execute_live || {};
@@ -24,6 +25,9 @@ export const generateScriptNode = {
       exec.finalUrl
     );
 
+    const recorded = loadRecordedActions(state.sessionPath);
+    const setupHint = formatSetupHint(detectLoginPattern(recorded));
+
     return `Generate and verify Playwright test at ${state.outputPath}
 
 Test Spec:
@@ -35,18 +39,23 @@ Live Execution Summary:
 - Final URL: ${exec.finalUrl || 'unknown'}
 ${actionsBlock}
 ${assertionsBlock}
-
+${setupHint}
 IMPORTS AND PATTERN:
 \`\`\`javascript
 import { test, expect } from '@playwright/test';
 import { StableIdRuntime } from '@zibby/core';
 
+async function clickSafe(page, stableId, fallback) {
+  try { await StableIdRuntime.clickWithRetry(page, stableId); }
+  catch { await fallback.click(); }
+}
+
 test('Test Name', async ({ page }) => {
   await page.goto('https://...');
   await StableIdRuntime.injectStableIds(page);
-  // Elements WITH stable IDs — use StableIdRuntime:
+  // Elements WITH stable IDs + fallback — use clickSafe:
+  await clickSafe(page, 'zibby-xxxxx', page.getByRole('button', { name: '...' }));
   await StableIdRuntime.fillWithRetry(page, 'zibby-xxxxx', 'value');
-  await StableIdRuntime.clickWithRetry(page, 'zibby-xxxxx');
   // Elements WITHOUT stable IDs (NO_STABLE_ID) — use native Playwright selectors:
   await page.getByText('visible text').click();
   await page.getByRole('button', { name: 'Submit' }).click();
@@ -58,18 +67,29 @@ test('Test Name', async ({ page }) => {
 RULES:
 1. First navigate → page.goto(), skip subsequent navigates
 2. After goto, call StableIdRuntime.injectStableIds(page)
-3. Use EXACT stable IDs from recorded actions when available
-4. For [NO_STABLE_ID] actions, use the fallback selector (getByText, getByRole, getByPlaceholder). These are typically non-semantic elements like spans acting as buttons — use the visible text to target them.
+3. Selector priority:
+   a. If memory/insights flag a stableId as unreliable → use the fallback selector instead
+   b. If action is marked [DUPLICATE_STABLE_ID] → always use the provided fallback
+   c. Cross-reference stableIds against memory "Reliable Selectors" and "Flaky Selectors" — prefer proven selectors, avoid flaky ones
+   d. Otherwise use EXACT stable IDs from recorded actions
+4. For [NO_STABLE_ID] actions, use the fallback selector (getByText, getByRole, getByPlaceholder)
 5. Skip duplicate consecutive clicks on same stableId
 6. No comments in generated code
 7. Implement ALL assertions from the list above
 8. If an assertion fails after retries, comment it out with a TODO (don't delete it)
+9. Selector failure handling:
+   - When a stableId fails and you switch to a fallback, IMMEDIATELY call memory_save_insight (category: selector_tip) with which stableId failed, which fallback worked, and the page URL
+   - Note the specific Playwright locator strategy that succeeded
+10. Navigation order: always complete setup/login FIRST from the base URL, then navigate to the target page. Never go to a deep URL before setup is done.
+11. The generated test runs in a FRESH browser with no prior state. Even if the live execution skipped setup steps, the test must include them. Check memory insights for any required setup.
 
 WORKFLOW:
-1. Write test to ${state.outputPath}
-2. Run: PLAYWRIGHT_HEADLESS=1 npx playwright test ${state.outputPath} --reporter=line --timeout=30000
-3. If fails: make ONE targeted fix (longer timeout, different selector, short wait)
-4. MAX 2 ATTEMPTS then STOP
+1. Study the codebase FIRST — search tests/ for existing helpers, fixtures, and shared setup files. Read them. Reuse what exists. Do NOT create files that duplicate existing ones.
+2. Write test to ${state.outputPath} (after the run, a copy is mirrored under ${state.sessionPath}/generate_script/ for Studio — you may also write directly there if you prefer)
+3. Verify syntax: run node --check on the file. If it fails, fix and re-check before proceeding.
+4. Run: PLAYWRIGHT_HEADLESS=1 npx playwright test ${state.outputPath} --reporter=line --timeout=60000
+5. If fails: try selectors in order — (a) getByRole (b) getByText (c) getByTestId (d) add waitForSelector. Never retry the same selector twice.
+6. MAX 2 ATTEMPTS then STOP
 
 The test runs in: ${state.cwd || 'project root'}
 `;