Zed is a fast, GPU-accelerated code editor written in Rust.

HelixML

Helix forked the Zed code editor to run fleets of headless coding agents inside Docker containers, controlled via a WebSocket protocol connecting a central Go API to each Zed instance and an LLM via the Agent Control Protocol (ACP). The post details three bugs encountered: multi-message accumulation (overwrite vs. append semantics), a completion hang caused by session-level event rejection logic, and garbled text from byte vs. UTF-16 offset mismatches in patch-based streaming. It also covers a streaming performance optimization that reduced wire traffic from O(N) full-payload sends to O(delta) patches, achieving roughly a 5000x reduction per update for long responses, plus throttling on both the Go and Rust sides to reduce redundant events.

How We Forked Zed to Run a Fleet of Headless Coding Agents

Bug 1: The Multi-Message Accumulation Problem

Shared Protocol Code: Eliminating Test Drift

Streaming Performance: O(N²) to O(delta)

<p>This is the real challenge of AI systems, plumbing, not prompts. State sync, encoding, and event flow break things long before model quality does. The 5000x delta optimization is where real leverage lives.</p>
