Diagnostics

Most build tools hand you a wall of text and leave you to reason out whether the failure is your change, your environment, or a bug in the tool. magus treats that reasoning as the tool's job, not yours. Every diagnostic magus emits is a pointable coded error with a name, a cause, and a resolution, designed for a human to act on, not for a machine to parse.

Anatomy of a diagnostic

[MGS4005] generated output drifted but its declared inputs are unchanged; the committed
form is produced by the pinned release and you are running a dev build (v0.1.0-5-gabc123)
- not your change, do not commit (see: https://.../codes/race/MGS4005.md)

Three things come with every code, for free:

• A stable code (MGSxxxx) you can search, grep your CI logs for, or paste to a teammate. Codes never move once assigned.
• A handwritten resolution page at the see: URL: a real Why and Resolution written by a person, so neither you nor an agent has to reverse-engineer the fix.
• A queryable graph node. Every code is a node in the knowledge graph: magus explain MGS4005 prints the card, and the code page is a first-class doc-site page. Nothing is ad-hoc free text.

Codes are grouped into families

The prefix tells you the domain at a glance:

A worked example: drift classification

The clearest expression of the philosophy is what magus does when a generated file drifts. A generate gate re-runs the generators and checks whether the tree went dirty. When it did, vcs\diagnoseDrift names why, instead of just failing:

• MGS4006: stale generated output. A declared input changed. Real drift: regenerate and commit.
• MGS4005: environmental drift. The declared inputs are byte-identical to what is committed, but a dev build (or a locally installed tool at a different version than the pinned release) rendered them differently, the classic markdown-emphasis case, *x* versus _x_. Not your change; do not commit.
• MGS4003: non-deterministic output. Same inputs, same generator version, output still differs: a reproducibility bug in the generator.

magus already holds every fact to make that call: a content hash of the declared inputs, a version fingerprint of the generator and tools, and the produces/consumes edges in the graph, so it makes the call for you. An agent no longer reads a 25-file diff and reasons about markdown emphasis to conclude "toolchain noise, ignore." The code says it.

Why this matters

The same diagnostic serves three audiences with no extra work:

• A human gets a named cause and a linked resolution instead of a hunt.
• An agent gets a stable code it can branch on, so it stops burning context re-deriving "is this my problem?" every session.
• CI gets an identical, coded signal: the drift gate on a clean checkout with the pinned release stays a true content-drift gate, now with a code attached.

This is deliberate design: the cost of diagnosing a failure should be paid once, in the tool, not re-paid by every person and every agent that hits it.