@zibby/core 0.1.21 → 0.1.23

package/src/enrichment/pipeline.js

@@ -1,192 +0,0 @@
-/**
- * Enrichment Pipeline Manager
- * Orchestrates multiple enrichers to add data to events
- */
-
-export class EnrichmentPipeline {
-  constructor(config = {}) {
-    this.enrichers = [];
-    this.config = config;
-    this.stats = {
-      totalEvents: 0,
-      enrichedEvents: 0,
-      skippedEvents: 0,
-      errors: {}
-    };
-  }
-
-  /**
-   * Register an enricher
-   * @param {EventEnricher} enricher
-   */
-  register(enricher) {
-    this.enrichers.push(enricher);
-    // Sort by priority (highest first)
-    this.enrichers.sort((a, b) => b.getPriority() - a.getPriority());
-    return this;
-  }
-
-  /**
-   * Unregister an enricher by name
-   * @param {string} name
-   */
-  unregister(name) {
-    this.enrichers = this.enrichers.filter(e => e.getName() !== name);
-    return this;
-  }
-
-  /**
-   * Get enricher by name
-   * @param {string} name
-   */
-  get(name) {
-    return this.enrichers.find(e => e.getName() === name);
-  }
-
-  /**
-   * Enable/disable specific enricher
-   * @param {string} name
-   * @param {boolean} enabled
-   */
-  setEnabled(name, enabled) {
-    const enricher = this.get(name);
-    if (enricher) {
-      enricher.enabled = enabled;
-    }
-    return this;
-  }
-
-  /**
-   * Enrich an event through the pipeline
-   * @param {Object} event - MCP event
-   * @param {Object} context - { page, element, ref, session }
-   * @returns {Promise<Object>} - Enriched event
-   */
-  async enrich(event, context) {
-    this.stats.totalEvents++;
-    
-    // Start with original event
-    const enriched = { ...event };
-    
-    // Track which enrichers ran
-    const enrichersRun = [];
-    const enrichersSkipped = [];
-    const enrichersFailed = [];
-    
-    // Run each enricher in priority order
-    for (const enricher of this.enrichers) {
-      try {
-        // Check if enricher can run
-        if (!enricher.canEnrich(context)) {
-          enrichersSkipped.push(enricher.getName());
-          continue;
-        }
-        
-        // Run enricher
-        const startTime = Date.now();
-        const data = await enricher.enrich(event, context);
-        const duration = Date.now() - startTime;
-        
-        if (data) {
-          // Merge enriched data
-          Object.assign(enriched, data);
-          enrichersRun.push({ name: enricher.getName(), duration });
-        } else {
-          enrichersSkipped.push(enricher.getName());
-        }
-      } catch (error) {
-        console.warn(`[EnrichmentPipeline] ${enricher.getName()} failed:`, error.message);
-        enrichersFailed.push(enricher.getName());
-        
-        // Track errors
-        this.stats.errors[enricher.getName()] = (this.stats.errors[enricher.getName()] || 0) + 1;
-      }
-    }
-    
-    // Add enrichment metadata
-    enriched._enrichment = {
-      version: '1.0',
-      timestamp: new Date().toISOString(),
-      enrichers: {
-        run: enrichersRun,
-        skipped: enrichersSkipped,
-        failed: enrichersFailed
-      }
-    };
-    
-    // Update stats
-    if (enrichersRun.length > 0) {
-      this.stats.enrichedEvents++;
-    } else {
-      this.stats.skippedEvents++;
-    }
-    
-    return enriched;
-  }
-
-  /**
-   * Enrich multiple events in batch
-   * @param {Array} events
-   * @param {Object} context
-   * @returns {Promise<Array>}
-   */
-  async enrichBatch(events, context) {
-    const results = [];
-    
-    for (const event of events) {
-      const enriched = await this.enrich(event, context);
-      results.push(enriched);
-    }
-    
-    return results;
-  }
-
-  /**
-   * Get pipeline statistics
-   */
-  getStats() {
-    return {
-      ...this.stats,
-      enrichers: this.enrichers.map(e => ({
-        name: e.getName(),
-        enabled: e.enabled,
-        priority: e.getPriority(),
-        errors: this.stats.errors[e.getName()] || 0
-      }))
-    };
-  }
-
-  /**
-   * Reset statistics
-   */
-  resetStats() {
-    this.stats = {
-      totalEvents: 0,
-      enrichedEvents: 0,
-      skippedEvents: 0,
-      errors: {}
-    };
-  }
-
-  /**
-   * Log pipeline status
-   */
-  logStatus() {
-    console.log('\n📊 Enrichment Pipeline Status:');
-    console.log(`   Total events: ${this.stats.totalEvents}`);
-    console.log(`   Enriched: ${this.stats.enrichedEvents}`);
-    console.log(`   Skipped: ${this.stats.skippedEvents}`);
-    console.log(`\n   Registered enrichers (${this.enrichers.length}):`);
-    
-    for (const enricher of this.enrichers) {
-      const status = enricher.enabled ? '✓' : '✗';
-      const errors = this.stats.errors[enricher.getName()] || 0;
-      const errorStr = errors > 0 ? ` (${errors} errors)` : '';
-      console.log(`     ${status} ${enricher.getName()} (priority: ${enricher.getPriority()})${errorStr}`);
-    }
-    
-    console.log();
-  }
-}
-
-export default EnrichmentPipeline;

package/src/enrichment/trace-text-enricher.js

@@ -1,115 +0,0 @@
-import { EventEnricher } from './base.js';
-import { TraceParser } from '../utils/trace-parser.js';
-import { existsSync } from 'fs';
-import { join } from 'path';
-
-/**
- * TraceTextEnricher - Extracts ACTUAL element text from Playwright trace
- * 
- * Problem: AI might say "Login button" but actual DOM has "登录" (Chinese)
- * Solution: Parse trace.zip to find the actual element text at action time
- * 
- * This runs AFTER recording as a post-processing step.
- */
-export class TraceTextEnricher extends EventEnricher {
-  constructor(config = {}) {
-    super(config);
-    this.priority = 190; // Just below MCPRefEnricher
-    this.traceData = null;
-  }
-
-  getName() {
-    return 'TraceText';
-  }
-
-  getPriority() {
-    return this.priority;
-  }
-
-  /**
-   * Load trace data once for all events
-   */
-  async loadTrace(sessionPath) {
-    if (this.traceData) return;
-
-    const _tracePath = join(sessionPath, 'traces');
-    const traceZipPath = join(sessionPath, 'trace.zip');
-    
-    if (existsSync(traceZipPath)) {
-      try {
-        this.traceData = await TraceParser.parseTraceZip(traceZipPath);
-        console.log(`[TraceTextEnricher] ✅ Loaded trace with ${this.traceData.length} actions`);
-      } catch (e) {
-        console.log(`[TraceTextEnricher] ⚠️  Failed to parse trace: ${e.message}`);
-      }
-    }
-  }
-
-  async enrich(event, context) {
-    const ref = event.data?.params?.ref;
-    const eventId = event.id;
-    if (ref === undefined && eventId === undefined) return null;
-
-    // Load trace on first event
-    if (!this.traceData && context.sessionPath) {
-      await this.loadTrace(context.sessionPath);
-    }
-
-    if (!this.traceData) return null;
-
-    // Match by sequence order: event.id maps to traceData[id]
-    // This is more reliable than matching by ref (which trace doesn't have)
-    const traceAction = this.traceData[eventId];
-    if (!traceAction) {
-      console.log(`[TraceTextEnricher] ⚠️  No trace action for event ${eventId}`);
-      return null;
-    }
-
-    // NEW: Use the actualText extracted from accessibility tree
-    const actualText = traceAction.actualText || this._extractTextFromSelector(traceAction.selector);
-    const actualRole = traceAction.actualRole;
-    const actualAriaLabel = traceAction.actualAriaLabel;
-    
-    if (actualText || actualRole || actualAriaLabel) {
-      console.log(`[TraceTextEnricher] ✅ Event ${eventId}: text="${actualText}", role="${actualRole}", label="${actualAriaLabel}"`);
-      return {
-        traceActualText: actualText,
-        traceActualRole: actualRole,
-        traceActualAriaLabel: actualAriaLabel,
-        traceSelector: traceAction.selector,
-        traceStrategies: traceAction.strategies // Include all strategies
-      };
-    }
-
-    return null;
-  }
-
-  /**
-   * Extract actual text from Playwright's internal selector format
-   * Examples:
-   * - internal:label="登录" -> "登录"
-   * - getByText('Submit') -> "Submit"
-   * - getByRole('button', { name: '确定' }) -> "确定"
-   */
-  _extractTextFromSelector(selector) {
-    if (!selector) return null;
-
-    // Try internal:label format
-    const labelMatch = selector.match(/internal:label="([^"]+)"/);
-    if (labelMatch) return labelMatch[1];
-
-    // Try internal:text format
-    const textMatch = selector.match(/internal:text="([^"]+)"/);
-    if (textMatch) return textMatch[1];
-
-    // Try getByText format
-    const getByTextMatch = selector.match(/getByText\(['"]([^'"]+)['"]\)/);
-    if (getByTextMatch) return getByTextMatch[1];
-
-    // Try getByRole with name format
-    const getByRoleMatch = selector.match(/name:\s*['"]([^'"]+)['"]/);
-    if (getByRoleMatch) return getByRoleMatch[1];
-
-    return null;
-  }
-}

package/src/framework/AGENTS.md

@@ -1,98 +0,0 @@
-# Workflow Execution Architecture
-
-## Single Source of Truth: Template Node Files
-
-Workflows are **compiled from source code**, not loaded from JSON files.
-
-The template node definitions in `packages/core/templates/` are the single source of truth.
-There is no static JSON workflow file — any such file (e.g. `default-analysis-workflow.json`) is stale and unused.
-
-## How It Works
-
-### 1. Graph Builders Compile Templates into Workflows
-
-Each workflow type has a graph builder that imports node definitions and wires them together:
-
-```
-packages/core/templates/code-analysis/graph.js     → buildAnalysisGraph()
-packages/core/templates/code-implementation/graph.js → buildImplementationGraph()
-```
-
-These builders import the actual node files:
-
-```
-packages/core/templates/code-analysis/nodes/
-  ├── setup-node.js
-  ├── analyze-ticket-node.js
-  ├── generate-code-node.js
-  ├── generate-test-cases-node.js
-  └── finalize-node.js
-```
-
-Each node file exports `{ name, outputSchema, execute }`. The `execute` function is the actual runtime code.
-
-### 2. Backend Resolves Workflow at Execution Time
-
-When a user triggers an analysis (`POST /projects/:projectId/tickets/:ticketKey/analyze`), the backend resolves the workflow in this order:
-
-1. **WORKFLOWS_TABLE** (DynamoDB) — user-customized workflow saved via the UI workflow editor
-2. **project.analysisGraphConfig** (PROJECTS_TABLE) — legacy per-project config
-3. **null** — no saved workflow; the CLI will use the compiled default
-
-The resolved config (or null) is uploaded to S3 as `context.json` and passed to the ECS container via `CONTEXT_PRESIGNED_URL`.
-
-### 3. ECS Container Compiles and Runs the Graph
-
-The CLI (`packages/cli/src/commands/analyze-graph.js`) picks the workflow source:
-
-```
-if (options.workflow)        → local file (--workflow flag, dev only)
-else if (execCtx.graphConfig) → custom saved workflow from DynamoDB (serialized JSON)
-else                          → getDefaultGraph('analysis') → buildAnalysisGraph() from templates
-```
-
-When using the default path, `buildAnalysisGraph()` imports the node files directly — the `execute` functions run as compiled JavaScript, not as eval'd strings.
-
-When using a saved custom workflow, `compileGraph()` deserializes the JSON and wraps `executeCode` strings via `new Function()`.
-
-### 4. Key Implication: Saved Workflows Can Go Stale
-
-If a user saves a workflow via the UI editor, the `executeCode` is serialized as a string snapshot into DynamoDB. Future code fixes to the template node files **do not** propagate to saved workflows.
-
-**To pick up template fixes**: reset the workflow in the UI settings (Project → Settings → Workflow → Reset to Default). This deletes the DynamoDB record and forces the CLI to recompile from templates.
-
-## State Access Pattern
-
-The graph stores each node's output under `state[nodeName]`:
-
-```js
-// In graph.js after node execution:
-state.update({ [currentNode]: result.output });
-
-// So to access previous node results:
-state.analyze_ticket?.analysis        // NOT state.analysis
-state.generate_code?.codeImplementation  // NOT state.codeImplementation
-```
-
-## invokeAgent Return Shape
-
-When `invokeAgent` is called with a `schema` option, it returns `{ raw, structured }` (not the structured data directly):
-
-```js
-const result = await invokeAgent(prompt, { state, model, schema: SomeSchema });
-const output = result?.structured || result;  // unwrap
-```
-
-Without a schema, it returns a plain string.
-
-## File-Based Structured Output (Cursor Strategy)
-
-For Cursor agent, structured output uses a file-based approach:
-
-1. Framework generates a unique file path: `.zibby/tmp/zibby-result-<timestamp>.json`
-2. Instructions appended to prompt telling agent to write JSON to that file
-3. After agent completes, framework reads the file and validates with Zod
-4. If validation fails and `strictMode` is enabled, falls back to OpenAI proxy for schema enforcement
-5. Result files are kept (not deleted) for debugging
-
-The Claude strategy uses its SDK's native structured output — completely separate path.

package/src/framework/agents/base.js

@@ -1,72 +0,0 @@
-/**
- * Base class for AI agent strategies.
- * All provider implementations (Cursor, Claude, etc.) must extend this class.
- *
- * @abstract
- */
-export class AgentStrategy {
-  /**
-   * @param {string} name - Provider identifier (e.g. 'cursor', 'claude')
-   * @param {string} description - Human-readable description
-   * @param {number} [priority=0] - Selection priority (higher = preferred)
-   */
-  constructor(name, description, priority = 0) {
-    this.name = name;
-    this.description = description;
-    this.priority = priority;
-  }
-
-  /**
-   * Execute a prompt against this agent.
-   *
-   * @abstract
-   * @param {string} prompt - The prompt text
-   * @param {AgentInvokeOptions} options
-   * @returns {Promise<string | AgentStructuredResult>}
-   *   - Without schema: returns raw output string
-   *   - With schema: returns { raw: string, structured: object }
-   *   - On failure: throws Error
-   *
-   * @typedef {Object} AgentInvokeOptions
-   * @property {string}  [model]       - Model name or 'auto'
-   * @property {string}  [workspace]   - Working directory
-   * @property {object}  [schema]      - Zod schema for structured output
-   * @property {Array}   [tools]       - MCP tools available to the agent
-   * @property {Array}   [images]      - Image attachments (provider-specific)
-   * @property {string}  [sessionPath] - Session artifact directory
-   * @property {number}  [timeout]     - Execution timeout in ms
-   * @property {object}  [config]      - Full workflow config
-   *
-   * @typedef {Object} AgentStructuredResult
-   * @property {string} raw        - Raw agent output
-   * @property {object} structured - Parsed and validated output
-   */
-  async invoke(prompt, _options = {}) {
-    throw new Error('AgentStrategy.invoke() must be implemented by subclass');
-  }
-
-  /**
-   * Check if this agent strategy is available in the current environment.
-   *
-   * @abstract
-   * @param {object} [context] - Environment context (unused by most providers)
-   * @returns {boolean}
-   */
-  canHandle(_context) {
-    throw new Error('AgentStrategy.canHandle() must be implemented by subclass');
-  }
-
-  getName() {
-    return this.name;
-  }
-
-  getDescription() {
-    return this.description;
-  }
-
-  getPriority() {
-    return this.priority;
-  }
-}
-
-export default AgentStrategy;