
OpenSpec (Part 1): Spec-Driven Development, One Change at a Time
Why I picked OpenSpec after reading the Thoughtworks Technology Radar, and how it works in theory: specs, changes, deltas, archive, and the loop that ties them together.
Tag archive

Why I picked OpenSpec after reading the Thoughtworks Technology Radar, and how it works in theory: specs, changes, deltas, archive, and the loop that ties them together.

OpenSpec Workbench 0.108.0 puts the committed diff, the PR target, and recorded verification in one review.

OpenSpec Workbench assigns a sandbox to each Codex stage and reads the result of the turn through the native CLI or ACP.

A run inside one of the repository's worktrees is given that worktree's own agents, whose sandbox is that worktree. Any other run gets the workspace's, whose sandbox still refuses a directory that is neither.
AI coding assistants like Claude Code, Cursor, and GitHub Copilot are remarkably fast—but without...

Agent workflow rules now say when to commit and push. A new change is committed and pushed as it is made. A card of a change's own worktree offers Commit Change, and the git stage commits what the stages left.

What the product says to a person is now written in one register, every message carries an OSW-XXX code, and a generated page lets you look up the identifier you saw.

Two new shared core tools, search_web and fetch_webpage, let local-model agents consult a SearXNG endpoint and retrieve HTML pages as Markdown. No remote agents are changed.
AI coding agents report 'done' far more often than they actually changed system state. Exit codes and green test runs are not evidence. The cheap gate: ask for the diff plus the output that proves the state moved.

OpenSpec Workbench now ships a versioned catalogue of rules, checks that run only when the host is idle, and findings that explain what is wrong without moving a single file.

A chain has four stages you can hand to four different agents, at four different prices. The harness can prove the savings on two of the three agents involved, and is honestly blind to the third.

A 54 MB leftover that survived a whole day of cleanup passes, and a context window nobody was reading until DeepSeek's own agent told us how full it was.