Canopus — AI skill for Claude Code

Canopus — keep long-running AI agents locked on their goal

Spacecraft on multi-year missions used a *Canopus star tracker* to lock onto one bright star and hold their orientation. **Canopus** does the same for AI coding agents such as Claude Code and Codex: a small, complete framework for running them with engineering discipline. It turns an ambiguous request into a frozen, acceptance-checkable goal; lets replaceable agents push that goal forward across sessions, engines, and machines; and **stops them deterministically when they drift, loop, or over-engineer** — while every project's data stays in its own repository.

It is plain standard-library Python and Markdown that you can adopt in about 15 minutes.

[`docs/EVIDENCE.md`](docs/EVIDENCE.md) maps every behavior to the test that proves it.

Why it exists

AI agents are productive inside one chat and unreliable across many. The failures are predictable:

Failure	What Canopus does instead
Chat becomes the database; a new session forgets what was accepted	Task state lives in the task's Issue and Task Footer; every session resumes from it
Commits, tests, and reviews look like progress	Only closed acceptance items and resolved blockers count as progress
Every "reasonable" finding grows the scope	The envelope is frozen; out-of-scope findings are deferred, widening needs re-admission
Review/fix loops never end	Finite allowances that never replenish, and a circuit breaker on stagnati