Accepted Setting Silently Inert

What goes wrong

A setting, flag or config key is accepted at the entrance — validated, echoed, logged — but nothing at runtime reads it, or the default is silently used instead, and the run reports itself under the requested label.

Example

The retries flag is set and echoed in the startup log, but the code reads a slightly different variable name, so the run labelled 'with retries' ran without them.

Not to be confused with

The operation ran, only not with what was asked; not a setting refused loudly, not a value outside what storage can hold (Storage range escapes input validation), not a policy that fails open when unreadable (Unreadable policy opens the gate), and not a false report about the operation itself (Unearned Completion Report).

Discovered by

@sks24git GitHub

Field notes · 11

@sks24git2026-11-04

A fallback path reads the hard-coded default instead of the environment override. Callers that take the fallback ignore the operator's setting while the main path honours it.

Resolve a setting once and pass the value down. A second read in a fallback picks the default.

@sks24git2026-11-04

A settings endpoint dropped fields a profile does not allow instead of refusing them. The caller gets a success, believes the setting took, and only learns otherwise when the run ignores it.

Refuse unsupported input; never drop it silently. A quiet drop reads as success to the caller.
Show what was saved, not what was sent. The reply is the caller's only check.

@sks24git2026-10-28beat 1 review round

A new concurrency knob shipped with a code default but was never added to the settings registry, so the API lookup fell back to the hard-coded value. The owner's dial does nothing and only a redeploy changes it.

A knob is not a knob until the registry knows it. Register the key in the same change that reads it.
Test settings through the real lookup. A stub that returns the default hides a dead dial.

@sks24git2026-10-27beat 2 review rounds

A quarantine registry gained a second test runner nobody reads yet, and its auto-entries still count toward the cap. The flaky tests keep failing while the cap fills with entries that do nothing.

Land the consumer before the producer. An entry nobody reads is noise that takes a slot.
Count only what works. A cap over no-ops starves the real ones.

@sks24git2026-10-25

A test module swapped explicit imports for a glob collector; a class nobody had ever imported woke up without the skip guard its siblings carry. Runs without the optional database now fail two tests instead of skipping.

A new collector runs what was never run. List every class it picks up and check each guard.
Count tests with and without the optional backend. A gain seen only with it hides a red run.

@sks24git2026-10-25beat 1 review round

A new 'feature without a story' sensor was specified to flag every merged change older than a window with no story naming it. With no start date, its first pass would raise one alert per change in the whole history.

Every backlog sensor needs a start point. Define 'since when' or the first run audits all of history.
Aggregate alerts that scale with history.

@sks24git2026-10-21

A balance-chain rule said 'compare with the neighbouring statement' but never said the neighbour must be adjacent. One missing month and a correct file is rejected as a balance error instead of a coverage gap.

Define 'neighbour' by contiguity, not order. Gaps belong to the coverage check.
Copy every gate clause into the binding spec. Prose in a design doc is not a test target.

@sks24git2026-09-12beat 1 review round

A script refused a timestamp younger than ten minutes with a message blaming the operator for passing the wrong instant, while the runbook told them to apply at once. Its count gate also read two different snapshots.

An error must name the condition it tested, not the mistake you expect.
Take both sides of a consistency check in one snapshot, or races read as operator error.

@sks24git2026-09-11

The insert intent was derived from a metadata field that only escalations used to write, and a new path started writing it too. That path then took the escalation intent and skipped the under-lock duplicate read.

Derive intent from what the request is, not from a field another path may start writing. Carry the fact on the job explicitly.

@sks24git2026-09-10

A warmup step asked the build tool to run with no background daemon, but passed its memory options through a variable the client never reads. Unable to honour them itself, it forks a daemon anyway and leaves it behind.

Check which process reads each option variable. A flag on the wrong one is ignored in silence.
Verify that 'no daemon' actually left none.

@sks24git2026-09-07

An adapter capped the top reasoning level one step down, trusting a stale comment about what the CLI accepts. Runs then logged the level asked for, not the one run, so the table measured a level that never executed.

Check option ladders on the pinned artifact, not a comment. A version bump changes what a flag accepts.
Record what launched, not what was asked. Intent measures the plan, not the run.