Redtropig/harness-anchor

14 stars · Last commit 2026-07-31

A warn-only runtime layer for Claude Code agents. Keeps feature state, scope and handoffs on disk so they survive compaction, and requires an evidence path before a done claim. Never blocks a tool call. Companion to superpowers; bash + git, no MCP.

PluginWorkflow & Automation TDD & Testing Documentation Design & UI DevOps & Infra Productivity AI & Prompting

README preview

# harness-anchor

[![validate](https://github.com/Redtropig/harness-anchor/actions/workflows/validate.yml/badge.svg)](https://github.com/Redtropig/harness-anchor/actions/workflows/validate.yml)
[![license: MIT](https://img.shields.io/badge/license-MIT-green.svg)](LICENSE)

**A warn-only runtime layer for Claude Code agents.**
[`superpowers`](https://github.com/obra/superpowers) gives an agent *process*
(brainstorm → plan → TDD → review). `harness-anchor` gives it *environment and
state discipline*: where am I, what is the active feature, what counts as
"done", and what to do when stuck.

It never blocks a tool call. It keeps state on disk, injects it back at session
start, and makes the agent show its evidence before it claims anything.

## What the agent sees

Every session opens with this — real output, from an anchored project:

```
<harness-anchor-state>

View full repository on GitHub →