CogniPrep adds assessment providers in rounds. A round is one agent per
provider, each in its own git worktree on its own branch with its own dev server
port, building the practice tests and all the public pages for that provider in
one branch. Then the branches are merged into an integration branch one at a
time from the main checkout.
The November round was four providers at once: the NYPD police officer exam, the
Ramsay trade tests, the Secret Service entrance exam, and the US Customs and
Border Protection exams. Branch sizes, measured from each merge base:
| branch | files changed |
|---|---|
| us-cbp | 58 |
| us-saee | 61 |
| us-ramsay | 71 |
| us-nypd | 73 |
Most of that is the provider's own territory: a folder of its own under the
games library, its own components, its own tests, its own route folders, its own
screenshots. Then I counted the files that appear in more than one branch's diff:
21 files were written by all four branches. One more, the question bank id
union, by two of them.
That number is the real cost of a parallel build, and it is worth listing,
because none of it is accidental. Every one of these exists so that something
about the site is derived rather than retyped:
- the provider id map, the game catalogue, the game config, the component registry
- the card copy on the browse page, which is an exhaustive record, so a missing provider is a type error rather than a blank card
- the registry behind the "can you cheat this?" cluster, including its family map
- the blog registry and the single lookup page that disambiguates providers
- the employer evidence file, the test-format pages, the dark-theme route list
- the generated screenshot manifest and the population statistics seed script
The tag convention, and how it actually gets used
The rule given to every worker is to mark each block it inserts into a shared
file with // [<round-tag>:<provider>], keep the block contiguous, and never
touch another provider's lines. The round tag here was us.
Counting what is in the repository now: 154 opening tags for these four
providers (ramsay 42, nypd 41, saee 36, cbp 35), spread over 28 files. Closing
tags: 16, four per provider.
So 138 of those 154 marks are not blocks at all. They are line annotations,
because the insertion really is one line:
'/games/cbp', // [us:cbp]
'/games/saee', // [us:saee]
'/games/nypd', // [us:nypd]
'/games/ramsay', // [us:ramsay]
cbp: true, // [us:cbp]
saee: true, // [us:saee]
nypd: true, // [us:nypd]
ramsay: true, // [us:ramsay]
That is the convention working rather than failing. Four workers appending one
annotated line each to the same array produce a conflict a text merge resolves
correctly, and the annotation is what lets the reviewer check afterwards that
nobody edited a neighbour's line. The two files with the most marks are the seed
script at 20 and the game catalogue at 12.
The three files where a text merge lies to you
A plain union of both sides is right for most of the 21. It is wrong, silently,
for the ones whose insertion point is inside an object literal rather than at the
end of a list. Written down after each one happened:
- The blog registry, the cheating registry and the employer evidence file conflict mid-object when two workers insert at the same point. Taking both sides produces an object that is not the union of two entries, it is one corrupt entry. The resolution is to take one side wholesale and then splice the other worker's whole tagged block in.
- The cheating registry's family map is a different problem: every worker rewrites the same line, because its families are arrays of provider ids and a new provider joins one of them. Both sides look like a complete line, so the merge picks one and quietly drops a member. The arrays have to be merged by hand.
-
The browse page's card record loses the
},that closes a provider's entry under a plain union. This one is at least loud:tscreports TS1005 and points at the next provider's key.
Two smaller ones: a conflict that falls inside a single import { statement
needs both halves plus the opening line re-added, and the screenshot manifest is
generated, so a conflict there is regenerated rather than resolved.
The audit nobody enjoys
The last merge step is reading each branch's diff for edits outside its own
tagged blocks, and dropping them. It is not theoretical: one worker in an
earlier round rewrote the word count field on twelve unrelated blog entries,
presumably while being helpful. Twelve small truthful-looking edits in a 70-file
diff are exactly what a review misses.
The other rule that came out of this is narrower and more useful: a
hand-maintained list keyed by provider needs a test, not care. The dark-theme
route list drifted twice before it got one, because a missing entry there does
not break a build, it just renders one page in the wrong colours.
See it
All four providers are live, and every item on that list of 21 has a visible
consequence you can check without an account:
- Count the new URLs. The sitemap gained exactly 26 entries for these four providers, in a uniform shape: one hub, one cheating page, one guide, and one page per named test.
curl -s https://cogniprep.app/sitemap.xml | grep -coE '<loc>[^<]*(nypd|ramsay|saee|cbp)'
That returns 25, not 26, and the missing one is the Border Patrol guide, whose
slug names the job rather than the provider.
- The exhaustive card record: find the four new cards on cogniprep.app/games.
- The cheating registry: /cheating/cbp and /cheating/nypd reach the same verdict in the same layout, and the paragraph that gets there is specific to the instrument. The CBP one argues from the exam being adaptive with no review at the end; the NYPD one cannot, so it argues from the testing centre, the sealed phone pouch and the fingerprints at the door.
- The lookup page, which is one file that every provider in every round has to append to: /blogs/which-assessment-provider-am-i-taking. Search it for "Ramsey", the misspelling, and for "7311", which is an exam series number a research pass corrected in the same round.
- The dark-theme route list is the one you can read directly. View source on
/games/ramsay and search for
forceDarkPaths: it is an array of 102 paths inlined into the document by a blocking script, and four of those entries are the annotated lines above.
Top comments (0)