Producer And Consumer Never Tested Together

What goes wrong

Two sides of a shared format — a field name in a payload, a message schema, a metric name — agree only because both contain the same literal, and each side is tested or faked against its own copy, never run against the other, so a rename on one side ships green while the other side silently stops matching in production.

Example

The frontend sends the booking id in camel case and the backend, after a rename, reads it in snake case; each test suite uses its own copy of the name and every booking arrives with a null id.

Not to be confused with

Both sides run, unlike SQL Without Executing Proof; the expected value is independent, unlike Self Derived Expected Value; unlike Helper Only Wiring Proof the gap is the cross-side check, not entry-point wiring.

Discovered by

@sks24git GitHub

Field notes · 8

@sks24git2026-11-02

A filter matched status strings that a cleanup job writes, but the test inserted those strings by hand. Reword the job's status and the filter stops matching while every test stays green.

Make the real writer produce the value under test. A hand-typed literal pins your copy, not the producer.

@sks24git2026-10-28beat 3 review rounds

A screen's tests used fixtures with a reference format and link key the real sensor never writes. In production every ticket row is labelled as a pull request and no evidence link renders.

Build fixtures from a captured producer payload, never from the type you wish it had. A hand-written fixture tests the author's assumption, not the contract.

@sks24git2026-10-28beat 4 review rounds

A client kept polling event kinds that the merged server deliberately filters out, and every test stubbed those kinds with fixture rows. The feed ships green and can never show what its design says it is for.

After a merge, re-run the contract against the real producer, not the fixture. A stub that invents the server's answer hides every narrowing the server made.

@sks24git2026-10-28beat 3 review rounds

A new caption was wired through the API and the screen, but the producer that feeds them never emits the flag that turns it on. The feature ships dark while tests that inject the flag by hand stay green.

Test the producer, not the flag. A test that injects a field proves nothing about the path that must set it.

@sks24git2026-10-17beat 2 review rounds

A relay posts a user's reply under a bot account, but the reader accepts only messages written by humans. Both sides pass their own tests, and every relayed reply is dropped until the waiting job times out.

Test the handoff, not the halves. Feed real writer output to the reader in one test.
List who may write each message. A human-only filter drops replies your own bots relay.

@sks24git2026-10-13

A routing test fed in labels the real alerts never produce, and it passed. The alerts reached people only through a catch-all.

Test with the labels your rules really emit. A green check on made-up input proves nothing.

@sks24git2026-09-06beat 2 review rounds

An alert rule read an unlabelled total counter while the worker comment and the migration named the labelled breakdown as its reader. The metric was right; three texts pointed operators at the wrong series.

Couple names across boundaries with a test. A rule file and a a language constant drift unless something fails when they disagree.

@sks24git2026-09-06beat 1 review round

A repository method grew to 25 parameters and an exactly-zero verification listed 16 matchers. The mock framework filled the rest with default-equality, so the guard passed against calls it should have rejected.

Omitted mock params are not wildcards. List every parameter in a zero-call guard, or the framework silently narrows it.