A coding agent stripped down to six readable moving parts

What it does

mini_coding_agent.py is a self-contained coding agent loop that edits code inside a workspace by calling a local Ollama model. It collects repo context, manages prompts, validates tool calls, and persists memory and transcripts—all without importing anything beyond Python’s standard library. The goal is readability and pedagogy, not production robustness.

The interesting bit

Instead of hiding mechanics inside a framework, the code is organized around six explicit building blocks: live repo context, prompt caching, structured tools with approval gates, context reduction, transcript resumption, and bounded subagent delegation. It is essentially a dissected frog for agent architecture.

Key highlights

• Runs against local Ollama models with zero third-party Python dependencies.
• Risky actions like shell commands and file writes are gated by configurable approval modes (ask, auto, never).
• Sessions are saved as durable JSON transcripts that can be resumed later, paired with a distilled working memory.
• Long outputs are clipped and repeated reads deduplicated to keep prompts within bounds.
• The model is expected to emit <tool> or <final> tags; reliability varies by model strength.

Caveats

• Explicitly optimized for readability over robustness, so edge-case handling is thin.
• Relies on the local model’s ability to follow strict output formats, which weaker models may ignore.

Verdict

Grab this if you are trying to understand agent internals or teaching the concept. Skip it if you need a battle-tested coding assistant for daily production work.