MGS4003: non-deterministic output

A project's declared outputs produced different content across two consecutive runs of the same dispatch. The same inputs should yield the same outputs; magus treats determinism as a build-system invariant.

[MGS4003] non-deterministic output (see .../MGS4003.md)
  project=api target=build differing_paths=[api/dist/bundle.js api/dist/manifest.json]

Why

This is an error, not a warning: it fails the run. Determinism is not a preference here, because three things magus does are unsound without it: drift gating compares regenerated output against what is committed, a cache hit replays stored bytes instead of rerunning, and a conflict in generated output is resolved by regenerating. A generator that emits different bytes for the same inputs breaks all three. The fix is always in the generator.

It is only emitted by --race=replay, which:

1. Runs the affected set normally (concurrent, cache-aware).
2. Content-hashes every declared output of every project.
3. Re-executes the same set sequentially, bypassing the cache.
4. Content-hashes the outputs again.
5. Compares the two hash sets per project.

Any project whose outputs differ between the two runs is non-deterministic. Common causes:

• Embedded timestamps. A build that injects time.Now() or __DATE__ into the artifact (linker build IDs, pkg/build.go ldflags, source maps with generation timestamps).
• Map iteration order. Go and Python iterate maps in random order; a generator that emits code from a map without sorting keys produces different bytes each run.
• Parallel work-stealing leaking into output. Worker IDs, goroutine IDs, or the order of fan-in results landing in a manifest.
• PID / hostname / cwd in the output. Some toolchains embed these by default (e.g. pip embeds the build tempdir path into compiled .pyc).
• Random salt / nonce. Crypto helpers used at build time without a fixed seed.

This check is expensive: it doubles the wall-clock time of the affected build, and the second pass cannot benefit from the cache. That is why it is opt-in and belongs on a scheduled audit rather than on every push. Asking for it and getting a warning would be the worst of both, so when you ask, you get a verdict.

Resolution

1. Identify the source

The differing_paths list narrows the search. Open one of the differing files in both runs and diff them. The divergence is often visible at first glance (a timestamp, a permuted list, a tempdir path).

For binary outputs, use diffoscope (the reproducible-builds.org tool) which unpacks archives, decompresses sections, and diffs the contents recursively:

magus run build
cp -r api/dist /tmp/run1
magus run build --race=replay   # second run rebuilds
diffoscope /tmp/run1/bundle.js api/dist/bundle.js

2. Remove the source of non-determinism

Common fixes:

• For embedded timestamps, ldflag the build time to a fixed value (-X main.BuildTime=$(git log -1 --format=%cI) ties it to the commit timestamp, which is deterministic).
• For map iteration, sort keys before emitting (slices.Sort(maps.Keys(m))).
• For tempdir paths, use a fixed relative path or strip the absolute prefix.
• For build IDs and PIDs, pass -buildid= to go build and -Wl,--build-id=none to the linker.

3. If non-determinism is unavoidable

Some artifacts (e.g. compiler outputs with debug info on Windows) are non-deterministic by design. For these, exclude the path from the project's declared outputs so the cache doesn't track them.

What this is not

• Not a runtime crash detector. It only checks that outputs are bit-for-bit identical, not that they behave the same.
• Not run by default. You must pass --race=replay explicitly.
• Not a substitute for reproducible-builds.org tooling. That community has stronger guarantees (cross-machine, cross-time, cross-locale). MGS4003 is a same-machine same-session check.

See also

• MGS4001.md: runtime filesystem race detector.
• reproducible-builds.org: the gold standard for build determinism, plus diffoscope for inspecting differences.
• --race=replay mode on magus run.