Ordewell uses read-only mode to stop coding agents rewriting code while exploring it

·1 views

Developer and Ordewell creator explains how AI coding agents often make unintended edits to a codebase while simply exploring it to answer a question. The problem arises because agents tend to test hypotheses by modifying files, blurring the line between findings and inventions in their output. Ordewell, an open-source coding orchestrator, addresses this by enforcing strict read-only permissions at the process level during planning sessions, not merely through model instructions. The platform applies the strongest available read-only flags across tools like Claude Code, OpenCode, and Codex, so write requests are refused outright rather than prompted for approval. This separation ensures that exploration and mutation remain distinct phases, with only dedicated runner tasks permitted to alter the codebase.

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)

Dev Drops AI Race Agent Idea, Opts for Classic Rule-Based Tactics System

A developer building EvoWild Run, a creature-breeding racing game, initially planned to let players deploy AI agents to make real-time decisions during races. The concept aimed to replace direct player control so that a creature's bred traits, not the player's reflexes, would determine race outcomes. To avoid splitting the game into two systems for AI and non-AI users, the developer unified the approach by converting UI-selected tactics into a common 'Agent File' format. Upon closer inspection, the runtime logic — checking conditions and executing predefined commands — turned out to be a classic rule-based system rather than true AI. The developer concluded that the 'Race Agent' label was misleading, and that games have long used similar tactics-programming mechanics without calling them AI.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Releases node-raft-rsm SDK to Add Raft Consensus to Existing Node.js Services

A developer has published node-raft-rsm, an open-source Node.js SDK designed to embed Raft consensus logic into existing services without requiring a full system rewrite. The library targets applications such as schedulers, metadata services, and control planes that need leader election and log replication but do not fit neatly around external systems like etcd. It exposes a small typed API surface — including RaftNode.create, propose, and read — while leaving storage, transport, state machine, and domain logic entirely under the application's control. The project is currently under active development and is explicitly not production-ready, intended for integration testing and API evaluation. According to the author, the goal is to let teams route existing mutation paths through Raft semantics without redesigning their service architecture around a general-purpose platform.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer finds 14 bugs in real-world testing after 1,800 automated tests passed

A developer building Filament Studio, a Laravel plugin with automation and MCP server features, discovered 14 defects by manually testing the app in a realistic environment after an 1,800-test automated suite and mutation testing had indicated the code was in good shape. The real-world setup used HTTPS webhooks, a live queue worker, MySQL, and a browser-driven admin panel — conditions the automated tests had never replicated. One critical bug caused flows triggered without the visual designer to silently complete with zero steps executed, because the graph-walking engine only handled edge formats produced by its own UI. A second serious flaw left webhook HMAC authentication effectively open to anyone, since a missing config field meant webhook secrets were never generated, and the verifier fell back to validating against an empty string. The developer concluded that nearly every bug shared the same pattern: the system reported success while either doing nothing or doing the wrong thing.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How Developers Are Quietly Building Personal AI Hierarchies for Different Tasks

Many developers have unconsciously assigned different AI tools to specific tasks based on trust, cost, and capability rather than a deliberate plan. One developer describes using Claude Opus for complex planning and hard bugs, while Gemini 3.1 Pro handles routine coding tasks despite hallucinating more, simply because it offers greater usage headroom. Outside the IDE, ChatGPT serves quick throwaway questions, Claude handles urgent drafts, and Gemini is reserved exclusively for calculation-heavy problems where it has proven consistently reliable. The system only became visible when an agentic AI over-engineered a simple formatting check, burning multiple actions on a task ChatGPT resolved in a single reply. The experience highlighted that effective AI routing is less about raw intelligence and more about matching the shape of the tool to the shape of the task.

0 comments Read more at DEV Community