SpexCode

Spec-driven orchestration for your coding agents.

SpexCode gives a human, a coding agent, and a Git repository one shared account of intent. You state what should be true; the agent changes the code and the spec together; Git keeps the history; evals measure the result through the product surface. The agent can do the work, but the human still reviews the proposed merge.

The complete Prompts & guidance catalog keeps the system contract, CLI help, hooks, skills, and review guidance together in one verified, readable place.

One project, three useful layers

• The spec asset (L0) keeps intent next to the code in ordinary, reviewable Git files. It remains useful without a server or dashboard.
• The session substrate (L1) gives agents isolated worktrees, explicit lifecycle states, and a reviewable merge handoff.
• The dashboard (L2) is the shared view of that same data: the graph, sessions, issues, evals, and live terminals. It adds visibility; it does not become a second source of truth.

The layers are deliberately separable. Start with a plain agent in one repository, or add sessions and a dashboard when parallel work makes them valuable.

A normal agent remains the default

Run spex init --harness claude,codex in a Git repository, then launch the agent you already use. SpexCode materializes a small contract into the files each harness discovers automatically, so a normal claude or codex session knows to read the governing spec, update intent with code, measure the result, and propose rather than merge its own work.

You do not need to learn a second prompt language. Tell the agent what you want changed and review the diff it produces.

Intent, history, and evidence stay distinct

A spec node is a directory containing spec.md. Its prose says what the governed code must do now. Git supplies the durable record of why and when that intent changed. spex spec lint catches structural drift such as broken links, unclaimed source, and code that has moved beyond its spec. When behaviour matters, an eval records a scenario and its evidence against the real product.

That separation keeps documents from becoming stale progress reports and keeps automation from claiming work a human has not reviewed.

Start with a repository you already have

Getting started covers the install and one-command adoption path. Working with your agent explains direct use, managed sessions, and the dashboard. The Reference is projected from SpexCode's own spec tree, so it is the detailed contract behind this guide.