Can I vibecode Tabnine?

Build a closest honest personal substitute for Tabnine in an empty repository, focusing on privacy-first local AI autocomplete and chat.
Use TypeScript, Node.js 22, a VS Code extension, SQLite, and support for a local inference engine (like Ollama or Llama.cpp) or one user-supplied model API; do not offer alternative stacks.
The core loop is: provide a bounded coding assistant with local repository completions (FIM - Fill-in-the-Middle) and chat through a local or user-selected model endpoint, index the current repository for context, propose patches, and preserve an auditable session log locally.
Make the first run work locally with one documented command to build and package the VS Code extension.
Store all user data, vector embeddings, and chat history locally in SQLite and make export straightforward.
Put secrets in .env, ship .env.example, and never commit credentials.
Create a VS Code sidebar with chat, selected-code actions (explain, refactor, generate tests), repository search, and a patch preview.
Implement real-time ghost-text autocomplete that triggers on typing, fetching suggestions from the configured model with low latency.
Index only the open repository and respect .gitignore plus a separate `.tabnineignore` file. Require explicit approval before running any command.
Represent edits as unified diffs with accept/reject. Capture tool calls, model requests, and patch decisions in a local session log.
Add token estimates, local model downloading instructions, tests, and an offline data-flow diagram proving data never leaves the machine unless configured.
Include clear empty, loading, success, and recoverable error states.
Write focused tests for the text-document synchronization and core transformation logic.
Create a README with setup, architecture, permissions, and backup steps. Do not add accounts, billing, telemetry, or analytics.
Finish by running the tests and listing the exact commands used.