Assumed property the producer never guarantees

What goes wrong

A consumer depends on a property of data it receives — an ordering, a key that resolves the same way on both sides, counts that line up — which the producing component never promised to hold. The property holds by accident on ordinary input, so the dependence stays invisible until the producer meets an unusual case or changes shape.

Example

A screen treats the first row of a returned list as the most relevant one, although the query that builds the list never sorts it.

Not to be confused with

Not a documented guarantee the code fails to keep: here nothing was ever promised, the producer is behaving correctly, and the consumer is the side asserting more than it was given.

Discovered by

@sks24git GitHub

Field notes · 15

@sks24git2026-11-04

A lookup query skips a filter because today only one writer creates the matching rows. The day a second writer appears, the query returns rows it was never meant to see.

Filter in the query, not in your knowledge of the writers. Writers change; queries are rarely re-read.

@sks24git2026-11-04

A cache-hit share added the cached tokens to the input count, but one provider's input count already holds them. Its runs showed about half the real hit rate, so a healthy cache read as a weak one.

Check each provider's token semantics before you sum counts. Disjoint for one, inclusive for another, and the same formula lies.

@sks24git2026-10-30beat 1 review round

When a log hit its size cap we moved new entries to a second log, but close marks were still applied in log order. A close mark written to the old log missed entries that sat in the new one, so done work came back.

Order events by time, not by where they are stored. Splitting storage breaks positional order.
Retest every reader when writes start to spill over.

@sks24git2026-10-28beat 1 review round

A spec claimed the existing event rows already carried the new fields, so no new writer was needed. The only writer emits other field names, so a reader built to the spec finds an empty parameter history.

Check the writer before claiming back-compat. A mapped kind name is not a mapped payload.
Read the cited line before calling it a divergence. A value re-read each round is one source.

@sks24git2026-10-27

A mirror switched from full re-reads to a delta keyed on each record's last-updated time. Fields that change without touching that time, and upstream deletions, now stay stale forever while the flag says live.

List what the delta key does not cover before you switch. Then keep a periodic full pass for exactly those.
A freshness flag reports the last full pass. Not the last delta.

@sks24git2026-10-27

An intake model pinned an identifier to letters only, while the producer's default vocabulary included a name with a digit. One such record and the strict parser rejects the whole batch, so the live view goes dark.

Test the contract with the producer's real values. A regex probed with made-up strings proves nothing.
Reject the record, not the batch. One bad field must not blank the whole screen.

@sks24git2026-10-25beat 1 review round

A spec hung a long audit on 'after the deploy workflow', but that workflow is a per-app button with a free ref and any outcome. As written, one release spawns several audits and a failed deploy audits the wrong commit.

Name the trigger's real shape before you hang work on it. A button with free inputs is not 'after release'.
One audit per shipped commit, never per click.

@sks24git2026-10-23

A cleanup cancels unsent notices whose key is not the current one, but edits get a suffixed key and cancelled keys stay taken. Each run kills edited copies, and one blip in status cancels the notice for good.

Match by identity, not exact key. Derived keys look foreign to an equality filter.
Never cancel on a transient state. A unique key turns every cancel into a permanent one.

@sks24git2026-10-18

A list of past findings was sorted by when each run happened, while its docs promised oldest round first. Re-run an old version after a newer one and the rounds come out shuffled, and the size cap cuts the wrong rows.

Sort by the key your docs name. A stand-in key reads right until the two disagree.
Cap after the final sort. A cut made in another order drops rows you meant to keep.

@sks24git2026-10-15

A view took labels from a list of options that changes from one step of a process to the next. In a later step the list held one action, so one row was named after the action and the rest showed raw ids.

Take labels from the data itself, not from nearby UI state. Option lists change by step; labels must not.
Test every step that shows the view. One happy-path step hides the rest.

@sks24git2026-10-13beat 2 review rounds

A cost rule relies on a newer behavior of a pinned tool, but the code never checks which version actually ran. Roll back the pin or ship an image with an older copy, and part of each run's spend vanishes.

Gate version-specific parsing on the version that ran. A pin states intent, not fact.
Keep a test for the older output format. It catches a rollback before the numbers drift.

@sks24git2026-10-08beat 3 review rounds

A shell helper lists changed files, then feeds those names back to a tool as paths. From a subfolder the root-relative names match nothing, and an accented name comes back quoted, so the review gets an empty diff.

Never round-trip file names through text. Pass the filter to the tool as its own path rules.
Test from a subfolder too. A root-only test hides every cwd bug.

@sks24git2026-10-04

Two sources of one item change price a second apart, and the screen merges their intervals. The old price now ends after the new one starts, so a "compare with what it replaced" rule sees no change.

Never read overlap as parallel when sources keep separate clocks. One edit lands at two instants.
Test the staggered case. Same change, one second apart.

@sks24git2026-09-28

A safety check compared a generated report's section count with the total in its header, but the generator skips sections for trivial items. On every normal week the check refused, so the tool marked nothing at all.

Build a consistency check from the code that makes both numbers. Two counts drawn from one list always agree; an assumed pair does not.

@sks24git2026-09-27

A job asks for the newest version of a prompt, but each version number is a hash of the prompt text. Edit the prompt and the old one wins half the time, so paid runs keep using text the code no longer has.

Pin the exact version you mean. "Newest" only works when numbers grow.
Never order by a hash. A hash is an id, not a clock.