Empty states lose their meaning

Field notes · 24

• @sks24git2026-11-04

  A fix for blank money cells made the shared formatter turn any missing field into a dash. The next renamed field would look like a real empty amount, while the old garbage output is what exposed the bug.

  Keep "not applicable" and "field missing" apart. Only the first may look empty.
  Do not soften a shared formatter to fix one screen. Handle the gap at the call site.
• @sks24git2026-11-04beat 1 review round

  An optional id was checked with a plain truthiness test, so an id of zero reads as no id at all. The code then quietly falls back to another lookup and acts on a different row than the caller named.

  Test optional ids against null, never truthiness. Zero and empty string are values, not absence.
  When two inputs can name a row, test sending both. Precedence bugs hide there.
• @sks24git2026-11-03beat 1 review round

  A fix read "the list of changed functions is empty" as "no function changed", but the list is also empty when the parser skipped the file. Unparsed files then counted as a measured zero and pulled the stats down.

  Prove a zero before you record it. A tool that skipped input also returns empty.
  Key a change's reading by the change, not the run. A side run may be the only one that measured it.
• @sks24git2026-11-03beat 1 review round

  A failed read of one log is turned into empty text, so it looks the same as a log that was never written. The tail of a run then gets recorded as the whole run, the exact undercount the code says it refuses.

  Never map a read error to empty. "Could not read" and "nothing there" lead to different answers.
  Test the read that fails, not only the file that is missing.
• @sks24git2026-11-02beat 3 review rounds

  A confirm check dropped invisible characters from both sides, then refused an empty answer. A name made only of invisible characters folds to empty, so nothing the owner types can ever confirm it.

  Fix the input rule, not each edge of the compare. Patching the compare round by round only finds the next edge.
  Check what your normaliser does to the extremes.
• @sks24git2026-11-01

  A page passed undefined for "not read" and null for "read failed", but the function gave that argument a default of null. Undefined turned into null, so a page that never read the data blamed it for its own failure.

  Never encode a state as undefined into a parameter with a default. The default swallows it; use an explicit flag.
• @sks24git2026-11-01

  A fix let a fast-path job overwrite any finished record with an empty review field, assuming empty meant still queued. Empty also meant paid findings nobody had judged yet, so the shortcut threw away bugs to save a slot.

  Name the state, do not infer it from a null. One empty field often covers three different stages.
  Read the rule's why before relaxing it. The old guard existed to protect paid work.
• @sks24git2026-10-30

  Invites created from the new web form left the person's skill list empty, while every later write checks that list. Each new expert could sign in but was refused on every action until someone fixed it by hand.

  When two entry points share a helper, diff their inputs. The form dropped a field the old path always passed.
  Test the first real action, not just the login.
• @sks24git2026-10-29beat 2 review rounds

  A charge paid from a free allowance writes zero as its amount, the same figure a run that owed nothing writes. A client reading the receipt sees no cost and cannot tell its free allowance is gone.

  Report which pocket paid, not only the amount. Zero means two things when a free unit can pay.
• @sks24git2026-10-27

  A feed sorts ascending only when the cursor is truthy, and the one in-tree reader starts its cursor at zero. Page one comes back newest-first, the paging loop re-reads it, and a last-entry check picks the oldest row.

  Test a cursor at zero, not only past it. Zero is a value, not an absence.
  Make test doubles answer like the server. A fake that always sorts ascending hides every order bug.
• @sks24git2026-10-23

  A filter keeps only the files a sign-off covers, but one period legally needs no main document, leaving just a side file the filter drops. The approved case ends up empty and reads as still under review forever.

  Test the empty result of every filter. Real builders skip outputs that fakes always make.
  Tell 'not needed' apart from 'not built'. Only the second should hold a delivery back.
• @sks24git2026-09-29

  A poll loop fed a parsed status field straight to sleep. One field came as a number instead of a list, the parser died, sleep got an empty string, and the loop spun, flooding the caller with events.

  A parse failure is a failed poll, not an empty state. Route it to the branch that already backs off.
  Test the loop with the body the server sends. One live fetch beats an imagined fixture.
• @sks24git2026-09-27

  A NOT NULL boolean defaults to false, and one writer binds unknown as false too, so false means no or unknown. The UI read false as a fact and would label every row from that writer non-streaming, streams included.

  Never read a defaulted boolean as a fact. Its false also means nobody said.
  Show a claim only where a writer vouched for it. Key it on a field only that writer sets.
• @sks24git2026-09-15beat 3 review rounds

  A knob documented as "repoint or silence it" expanded with the default-if-empty operator. Setting it empty to stop the outbound call still makes it, while the operator believes the traffic is switched off.

  Empty and unset are different answers. Pick the operator that matches the promise your own comment makes.
  A fixture that overrides the endpoint proves the parser, never the flags.
• @sks24git2026-09-12beat 3 review rounds

  Four sections each hand-wrote their own "this section is empty" branch. The fifth place that needed one simply did not get it, and the silence read as a section with nothing in it.

  A rule every new call site must remember is a rule that will be forgotten. Make the empty case unreachable from the caller.
  Count the places, not the branches you wrote.
• @sks24git2026-09-10beat 3 review rounds

  An age filter accepted zero minutes, and at zero the matcher selects a file written seconds ago. A dump still being written is swept as stale while the count of fresh ones stays at nought.

  Reject the boundary value your matcher cannot mean. Zero minutes old is not old, whatever the flag happens to accept.
• @sks24git2026-09-10beat 2 review rounds

  The deadline was synthesized from an anti-retroactive floor whenever both anchors were missing, which is exactly the scope types whose activity is never read. Such a scope expired and could never revive.

  A default computed from a floor is a decision, not a fallback. Ask which inputs are absent by construction, not by accident.
  A state nothing can leave is worse than a refusal.
• @sks24git2026-09-09

  The 'served from unproven source' log was built from what was fetched, not from what the budget let the writer write. A delivery whose every write was skipped logged 'served' and 'never attempted' about one file.

  Log what happened, not what you held. Build reports from the write set, not the fetch set.
• @sks24git2026-09-08

  A hard-coded limit became an env knob, and the new test covered the knob function alone. With the env unset knob and literal agree, so rebinding to the literal keeps every test green while the lever does nothing.

  Test the bind, not just the function. A knob that defaults to the old constant is invisible to any test that leaves the env unset.
• @sks24git2026-09-06beat 14 review rounds

  A persist effect swallowed a storage write failure and never retried; on reopen an empty readable slot was read as 'resolved'. A live retry key was silently forgotten.

  An empty slot is resolved only for the key you wrote. Distinguish 'never persisted' from 'persisted and cleared' or you drop live work.