SimpleCoder

Overview

SimpleCoder is a deliberately simple terminal coding assistant intended to demonstrate how equipping a model with a decent prompt & tool set can enable "agentic" AI behavior. The purpose isn't to showcase a production-ready agent, but rather to supplement a future blog post on the topic for non-technical readers. Around 95% of the code was put together over the course of a few hours.

For example,

• Message history management could be improved since in the current state, the entire message history is sent to the model for each request. A naive but "good enough" improvement for this simple project would be to track token usage & either truncate the message history or summarize it at certain thresholds to avoid context pollution & token waste on subsequent messages.
• More sophisticated stop conditions could be implemented as opposed to stopWhen: stepCountIs(Infinity)
• More deterministic flow control could be implemented as opposed to relying on the system prompt to dictate the flow
• Shell command execution could and should be further scrutinized

Again, outside of going a bit overboard with shell command execution & diff rendering, the intent is to just be a simple example.

Highlights

• Terminal-based chat with file system access via structured tools & powered by GPT-5
• Streaming token display with live tool execution feedback
• Basic file operations: read, edit, list, and command execution with safety guardrails
• Diff rendering for file edits

Stack

• Node v24.2.0
• TypeScript v5.9.2
• biome for linting & formatting
• vitest for unit testing pure utility functions
• ai sdk v5 (beta) for LLM interactions
• zod for schema validation

Footnotes

Claude Code (Claude Opus 4.1) was used to expedite the development of: command execution, diff rendering, and unit tests.