Why AI Coding Agents Should Leave Replay Scripts, Not Just Chat Logs

·10 views

A backend engineer merged a fix generated by an AI coding agent, but reviewers found the build failed on a clean machine with no trace of the original chat session. The core problem is that chat transcripts record intent rather than reproducible steps, meaning work validated inside one AI session may not survive outside it. To address this, a structured 'spike' ritual is proposed where the agent must produce a replay script — a shell script that restores a clean environment, applies only committed changes, and exits successfully on any machine without agent involvement. The spike runs within a fixed time budget, starts with a written hypothesis and a pre-existing failing test, and is killed outright if the replay script is missing or unrunnable elsewhere. The framework treats reproducibility, not conversational confidence, as the true measure of whether engineering work is done.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Developer Seeks Help Fixing Execution Flaws in Polymarket Latency Arbitrage Bot

A developer building a trading bot on Polymarket's prediction market platform reports that live performance is consistently negative despite a strategy that tests well on paper. The bot targets Bitcoin short-term price direction markets using a latency arbitrage approach via the Polymarket CLOB API on the Polygon blockchain. Key execution problems identified include large gaps between expected and actual fill prices, slow cancel-and-replace cycles, and API or rate-limit errors during live trading. The developer has already attempted fixes such as using limit orders, reducing order size, and improving logging, but the issues persist. They are now seeking community guidance on order manager design, realistic fill modelling, execution quality measurement, and open-source tooling for CLOB-based bots.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Go Developers Explore Zig: Key Differences in Memory, Errors, and Control

A technical comparison published on DEV Community examines what Go developers encounter when learning Zig, a systems programming language gaining traction among backend engineers. While both languages favour simplicity and a single official toolchain, their core priorities differ: Go targets developer velocity and scale, whereas Zig emphasises predictable performance and direct hardware control. The most significant shift for Go developers is memory management — Zig has no garbage collector, requiring programmers to explicitly pass and manage allocators for every allocation. Error handling follows a similar philosophy to Go's explicit error values, but Zig enforces it at the compiler level through typed error unions, making it impossible to silently ignore errors. The article argues that developers who have experienced GC pauses or struggled with cgo in Go are the most likely audience to find Zig's trade-offs worthwhile.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why staying at one job longer than two years builds judgment you can't get elsewhere

A software engineer argues that the common advice to change jobs every couple of years causes developers to miss a critical feedback loop that only emerges over 18 months or more. Watching your own architectural and design decisions age in production is described as the only reliable way to distinguish good instincts from bad ones. Beyond technical lessons, longer tenure also earns engineers informal influence — being consulted before decisions rather than after — and a working knowledge of how their specific organisation actually functions. However, the author cautions that staying can stop being productive, recommending a twice-yearly self-check: if you cannot name what you learned or what you would do differently, familiarity may be quietly costing you years of growth.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Rewrites Abandoned 2015 Rust CRC-32 Crate With Modern Features

A software developer discovered that the original Rust crc32 crate, last published in 2015 and lacking CI or a changelog, was still widely used as a dependency despite being unmaintained. The developer rewrote the crate from scratch using modern Rust, implementing multiple performance-optimized table variants including slicing-by-4, slicing-by-8, and slicing-by-16. The revamped library, named crc32-v2, supports no_std environments, includes a streaming Digest API, a build-time proc-macro codegen subcrate, and bindings for both Python and Node.js. The entire codebase was written in 100% safe Rust with unsafe code explicitly forbidden at the crate root. Following its release, the project gained unexpected viral attention within the Rust developer community.

0 comments Read more at DEV Community