Traycer-mini

Problem

AI code generators tend to write straight into your codebase — no plan to inspect, no diff to approve, and no check that the result even compiles.

Solution

A staged workflow: plan → generate → review → approve → verify. Plans are persisted as JSON, generated code lands in a staging area as unified diffs, you apply files selectively, and verification runs the TypeScript compiler and ESLint.

Architecture

Key engineering decisions

• Planning and generation are separate commands with a persisted plan, so the plan is reviewable and re-runnable.
• A staging directory with diffs instead of in-place writes: nothing touches the codebase without approval.
• Provider abstraction — OpenAI, Anthropic or Groq selected with one environment variable.

Features

• Natural-language task → structured plan with steps and files to touch
• Code generation per planned file, saved as proposals with unified diffs
• Interactive and batch modes built from modular Commander.js commands
• Selective approval — apply only the files you accept
• Automated verification with tsc and ESLint

Tech stack

• TypeScript
• Node.js
• Commander.js
• LLM APIs