Plandex

Plandex is a terminal-based coding agent that actually remembers what it's doing across hundreds of files. The autonomy here is cumulative: you describe a feature or refactor, it generates changes across your codebase, then you review them in a sandbox diff before applying. That review step is the whole point. Unlike Cursor or Aider, which apply changes immediately, Plandex builds a changeset you can inspect, reject parts of, and iterate on without polluting your working directory. I used it to migrate a legacy Express API to Fastify across 40 route files. Gave it the migration plan in plain English, pointed it at the codebase, and it proposed changes to every route handler, middleware registration, and error boundary. The 2 million token context window meant it cross-referenced how I'd handled similar patterns elsewhere. The sandbox let me reject its overly cautious error handling in three files and approve the rest. That workflow - describe, review, apply - beats the usual AI chat loop where you're manually copying diffs into your editor. Failure modes: it occasionally misses implicit dependencies. In the Fastify migration, it updated route signatures but forgot to change a shared validation helper that expected Express's `req` shape. The diff caught it, but only because I knew to look. It also struggles with ambiguous instructions. 'Make this faster' produces vague changes; 'replace this O(n²) loop with a hash map' works better. Compared to Aider, Plandex is slower but safer. Aider applies changes instantly, which is faster for small edits but riskier for multi-file refactors. Compared to GitHub Copilot Workspace, Plandex runs locally and costs nothing if you self-host. The cloud tier adds collaboration features and faster models, but the open-source version is fully functional. The terminal UI takes getting used to. No syntax highlighting in diffs, just plain text. But for large refactors where you need to audit every change before committing, that cumulative review sandbox is worth the friction.