Control Flow Keyed On Message Text

What goes wrong

Code picks a branch by searching a human-readable error or status message for a substring instead of using the structured code the message was built from, so rewording the message or a similar sibling phrasing silently changes behaviour.

Example

The code checks whether the error message contains 'duplicate key'; the database driver rewords its message in an upgrade and the 'already exists' branch goes dead.

Not to be confused with

The text inspected is the program's own output message; not a source-code scanner that misses syntax variants (Syntax Blind Static Guard).

Discovered by

@sks24git GitHub

Field notes · 4

@sks24git2026-11-04beat 2 review rounds

A flag meant "this setting is overridden" was computed by comparing two display labels, not the values behind them. Reword one label and the flag flips silently, so the hint stops matching the truth.

Derive state from data, never from rendered text. A copy edit must not be able to change logic; grep for comparisons against label maps.

@sks24git2026-11-04beat 2 review rounds

A timeout message was renamed for clarity, but the retry logic sorts failures by matching that text. The same overrun now looks like a different kind of failure and gets the wrong retry policy.

Grep for readers of an error's text before you reword it. Messages are often parsed.
Classify failures by type, not by wording. A rename should not change behaviour.

@sks24git2026-11-04beat 1 review round

A retry loop that sorts errors into "temporary" and "final" by matching the error text. A rate limit or a reworded server error matches no pattern, so the script quits on the first blip it was written to survive.

Ask the remote for the state, not its error text. Wording drifts; the state is one query away.
Cap retries instead of sorting errors. A short cap costs seconds; a wrong guess costs the run.

@sks24git2026-10-16

A catch let through only one wording of an error, but the same failure has a second wording. Any input that triggers the second one still crashes the request, the exact case the fix was written to stop.

Match the error, not one of its messages. One fault can show up under several texts.
Test each wording a filter accepts. A case you never send is a hole nobody sees.