Can a prompt replace Windsurf?

Build a VS Code extension that runs a multi-step agentic coding loop — distinct from a single-shot 'generate this component' assistant. Use TypeScript, and read a model key from ANTHROPIC_API_KEY or OPENAI_API_KEY. Given a task description, have the model produce a numbered plan (e.g. edit file A, run npm test, edit file B based on test output), and execute the plan one step at a time — pausing before every file edit and every shell command for explicit user approval, never running commands unattended. Index the whole repository's file tree and let the model request specific files' contents as needed rather than dumping the whole repo into context every time. Represent file edits as diffs (accept/reject/undo) and shell command output as a readable log entry, both stored in a local session file so a full run can be replayed or audited later. Add a hard-coded command allowlist pattern (e.g. only npm/pnpm/git/test-runner commands by default) that the user can extend, and refuse anything outside it without an explicit one-time override. Do not build unattended or background execution, a custom IDE fork, or team policy controls — those are out of scope; this is a single-user, always-supervised extension. Requires an Anthropic or OpenAI API key — without one, the extension runs in a read-only 'plan preview' mode with no execution.