Building a production coding agent that lives in Linear. Wrapping Claude Code and Codex as child processes, surviving state loss from archived threads, chaining three fallback parsers for structured LLM output, debugging session IDs that break on working directory changes, and testing everything against in-memory replicas of Linear and GitHub.

daily.dev

daily.dev built Huginn, a coding agent integrated into Linear that automates the full workflow from ticket to PR. The post covers the hard engineering problems encountered: wrapping Claude Code and Codex as child processes with their undocumented streaming formats, building a three-tier fallback parser for structured LLM output, debugging session continuity failures caused by working directory changes, and using Linear labels as a crash-resilient state machine. The team also describes their 'Digital Twin Universe' (DTU) testing approach — in-memory replicas of Linear, GitHub, and KMS running in Docker containers — which made a 99% AI-generated codebase viable. Known limitations include ongoing output parsing fragility, rough BYOK credential handling, and poor fit for tight iterative or architecturally complex tasks.

How we built a Linear coding agent: the hard parts

What a 99% AI-generated codebase actually means

<p>Great read, thanks for sharing! I really like what Linear is doing in terms of embedding AI agents; I’m actually discussing the possibility of experimenting with this tool with my manager right now. If we decide to move in this direction, your article will be a huge help in setting up our workflow</p>
