AI Pipeline Ticket Builder

VSCode-compatible extension that spins up a planning pipeline (product manager + UX agents) grounded in the open workspace, producing a Ticket.md you can hand to your implementation agent. Works in Cursor, Windsurf, and vanilla VSCode.

Features

• Index the current workspace with OpenAI embeddings (configurable chunking & ignores).
• Respects .gitignore, optional .contextignore, and VSCode settings to ensure noisy files stay out of the context.
• Streams and chunks very large text files (>10 MB) without exhausting memory while still capturing their contents.
• Prioritizes critical entry points, config, and documentation files first for faster high-signal indexing.
• Builds a dependency graph and symbol index (JS/TS-first, graceful fallbacks for other languages) to surface relationships, comments, and complexity insights.
• Performs semantic chunking at class/function boundaries with dynamic sizing, deduplication, and metadata about the symbols covered.
• Flags likely generated assets, high-complexity areas, and exports without downstream consumers so you know where to focus planning time.
• Background watcher reacts to saves/renames/deletes, incrementally refreshing only changed files while reusing cached embeddings.
• Git-aware metadata surfaces branch, status summaries, and merge-conflict hints directly in the stored index.
• Monorepo-aware topology metadata highlights workspace packages, language modules, and git submodules for cross-repo planning.
• Parallel chunking workers respect CPU and memory budgets so even very large monorepos stay responsive.
• Product manager and UX designer agents reason over repo context to output a structured spec.
• AI Pipeline: Generate Ticket command creates a Markdown ticket in-editor.
• AI Pipeline: Rebuild Repo Index lets you refresh embeddings after refactors.

Setup

1. npm install
2. Expose an OpenAI API key either via VSCode settings (aiPipeline.openaiApiKey) or environment variable OPENAI_API_KEY.
3. npm run compile
4. Launch the extension (F5) or package with npm run package.

Usage

1. Open any project in Cursor or Windsurf.
2. Run command palette → AI Pipeline: Generate Ticket.
3. Enter the feature goal (e.g., “Build AI insights page”).
4. Review the generated Ticket.md and pass it to your implementation agent.

Use AI Pipeline: Rebuild Repo Index whenever repository content changes dramatically.

Configuration

• aiPipeline.model: chat model for agents (default gpt-4.1).
• aiPipeline.embeddingModel: embedding model for indexing (default text-embedding-3-large).
• aiPipeline.ignoredGlobs: extend the ignore list beyond node_modules, dist, etc.
• aiPipeline.embeddingBatchSize: number of chunks sent per embedding request (default 64). Increase if you rarely hit rate limits; decrease if you run into provider throttling.
• aiPipeline.maxConcurrentWorkers: cap parallel file analyzers (default 4). Raise on beefy machines to speed large repos.
• aiPipeline.memoryBudgetMb: optional soft cap for file processing (0 = auto based on system RAM). Streamed chunking kicks in after the budget is crossed.

Context Acquisition Defaults

• Walks the entire workspace (including dot-directories) while respecting .gitignore, .contextignore, and aiPipeline.ignoredGlobs.
• Skips binary content and safely ignores symlinks to avoid infinite recursion.
• Streams files larger than 10 MB so even huge sources can be chunked without loading them fully into memory.
• Applies a priority heuristic so core entry points, configuration, and documentation surface first.
• Extracts imports/exports, symbols, comments, and basic complexity metrics per file (deep TypeScript/JavaScript analysis, heuristic fallback for other languages).
• Enriches the repo index with dependency adjacency data and unused-export hints to inform future planning steps.
• Semantic chunker keeps related logic together (class + methods, function + helpers) and deduplicates redundant slices before embedding.
• Incremental rebuilds reuse cached metadata and embeddings when file size/mtime is unchanged (or untouched), making large-workspace refreshes fast.
• Background saves, creates, deletes, and renames automatically schedule a silent index refresh with status-bar feedback.
• Stored metadata includes per-file hashes, git status codes, and conflict markers so downstream agents can prioritize risky diffs.
• Repo topology metadata captures workspace manifests (npm/pnpm/yarn), Go modules, Cargo crates, Python projects, and git submodule paths for multi-repo context.

Roadmap Ideas

• Add implementation & review loops directly in-extension.
• Support Anthropic models with an alternate provider setting.
• Persist richer metadata (git blame, tests) for deeper planning context.