Generated from cat-harness/skills/ui/theming/theme-art-intake.md — do not edit here.
✎ Edit this page’s source

Theme art intake

Art arrives — three crops for a theme or an avatar — and the answer is either accepted, with the declaration written, or refused, with what to fix.

make skills for avatar theme ingestion (need 3 meeting certain formatting constraints), return sucches or explnation of falire
…make as part of docuemtn ingestion skill.
— owner, 2026-09-20

This is a step in document ingestion, not a process beside it

Document ingestion already answers the question art arrival poses — something unstructured has arrived; is it acceptable, and if not, why not — and already has the shape of the answer: a pipeline with a completeness gate. Theme art is the same event with a different payload.

So what exists is a check keyed on the payload, not a second pipeline. A parallel “avatar ingestion” would be a second answer to one question, free to disagree with the first.

	where
the check	`schemas/theme-art-intake.ts` — pure, no I/O, no clock
the CLI face	`scripts/check-theme-art.ts`, `bun run cat check:theme-art`
the arrival	`uploads/`, per `content-acquisition` and `document-intake`

Every constraint was paid for

Each one is a failure this repository actually took. Do not add a constraint here that has not cost something — a rule with no incident behind it is a rule whose next maintainer cannot tell whether it still matters.

constraint	verdict	the incident
all three layouts	refuse	`landing-architecture` shipped laptop + card and no mobile from 2026-09-20 to 2026-09-24. `resolveThemeBackdrop` refuses an incomplete backdrop wholesale, so the theme renders with no art — not two thirds of it.
distinct content	refuse	commit `1b62b57` delivered three files of which two were byte-identical (same digest, both 1,606,269 bytes): two distinct images for three named avatars.
orientation per layout	refuse	a portrait crop is not a landscape crop scaled down. The quiet area the sticky’s text sits in moves.
a declared destination	refuse	commit `0301fbd2` put three PNGs at the repository root, undeclared, with spaces and commas in the names. No gate saw them: `check-declared-assets` walks declared→disk, and the root is not an instance.
dimensions measured	structural	`landing-engineer` is declared 1672×941 where its siblings are 1671×941 — one pixel, copied from a neighbour. Intake takes bytes and emits the declaration, so there is nowhere for a retyped number to enter.
weight and format	report	the arriving PNGs are 1.5–2.3 MB against ~100 KB webp. Heavy art still renders, so refusing it would block work over a bandwidth judgement.

Refuse, report, or neither — and the line matters

Refuse when the art cannot do its job, or when intake cannot tell which of two readings is intended. A missing crop and two identical files are both “somebody has to decide something”.
Report when the art works but costs. Weight and format are the cases: a 1.6 MB PNG renders correctly.
Neither — say nothing — for anything intake did not actually measure.

A refusal that says invalid teaches nobody what to fix, so remedy is a required field on every failure, not an optional one. All failures are collected rather than thrown one at a time: art arrives as a batch, and reporting the first problem of three means three round trips to learn three things.

Do not repair somebody’s art

deletion-requires-confirmation applies, and so does its spirit. When intake refuses, the outputs are the report and a work-plan item — not a re-crop, not a converted file, and not a quietly removed declaration. What looks like a missing crop may be a duplicate that was uploaded by mistake, and that is a question for the person who has the original.

check:theme-art:check was held back from CI for the same reason: it refused landing-architecture, and a gate added in the same change as the check would have made CI red over art that was missing rather than over a regression somebody just introduced. The portrait crop arrived on 2026-09-24 and the gate went in with it — a refusal now IS a regression.

Reading the report

✓ landing-engineer: 3 layouts accepted
  · laptop 1672×941 docs/assets/img/harness/landing-engineer-laptop.png
  ! [heavy] laptop: … is 1560 KB, over the 391 KB budget
      → convert to webp — the measured gap on this art was about 16×

✗ landing-architecture: refused, 1 problem(s)
  · [missing-layout] mobile: no file offered for the mobile layout
      → supply all three layouts …

Three rules for reading it, the same three ci-health states:

declared-but-absent is not a refusal and not a pass. The declaration names a file that is not on disk. It is reported on its own line because “the file is not there” and “the file is the wrong shape” are different facts, and a reader who cannot tell them apart fixes the wrong one.
A check that fires on every subject is wrong. The first run of check-theme-art refused all five roles for undeclared-destination, because docs/ is not a declared directory — and it should not be: a declared entry names a graph, and the site is the render target. The check was corrected, not the repository.
This one is general, and it has now been paid for twice. The rule is stated here because this is where it was first measured, not because it is about art. Second instance, 2026-09-20: the first run of check-escaped-markup found the landing-page defect it was written for and two pages that are simply correct — translation-support.html documents the path <section>/<block>/… and md-authoring.html quotes <table class="md-table">, both inside <code>, which is how a document shows markup to a reader. 19 findings; 6 were real. Blanking <code> and <pre> regions took it to 6/6, and the exclusion is tested in both directions, because a test that only checks the excluded case passes equally for a check that never fires at all.
So when you write a check and its first run lights up broadly, the first question is which of these are correct? — not how do I fix the repository? A check that fires on legitimate subjects is a check somebody switches off, and a switched-off check is worse than none because it reads as coverage.
Warnings do not accumulate into a refusal. Twelve heavy files are twelve reports, not a failure. Weight is ll11’s subject.

Adding a constraint

Name the incident. No incident, no constraint.
Decide refuse vs report by the line above — can the art do its job?
Write the remedy before the detection. If you cannot say what to do about it, the check is not ready.
Prove it fires, with a test that fails when the check is removed.
Check it against art already shipped. The aspect bands are deliberately wide because the shipped laptop crops run 1.50–1.78; a tighter band would have looked rigorous and refused work that is fine.

The structure this grows into

The owner chose the axis on 2026-09-20, as “1 + 2, then specialized to 3” — three levels rather than a choice between three options:

Pipeline stage is the primary division — intake → declaration → generation → contrast. This is how the failures actually arrived: each of the four rounds of “still not image” was one stage.
Producer vs consumer cross-cuts it, saying which direction a stage faces. It is what makes “a theme with no CSS rule falls back to an opaque surface and paints over its own art” a consumer fact rather than a generation detail.
Per artefact is the specialisation, last and narrowest — sticky, landing board, docs background, navbar avatar. Specialised rather than duplicated: the general rule lives in its stage and the artefact package carries only what differs.

The ordering matters more than the names. Artefact-first would duplicate intake and contrast into every artefact, which is the shape that put the scrim rule in one skill’s prose and let it be rediscovered rather than read.

Why this is one package today and not four

Measured before splitting: the corpus is three skills. Four stage packages over three skills would declare directories with nothing in them — the dh4f defect, where a consumer scans an empty tree and reports a clean run over it — and this repository’s own rule is to declare only what exists.

So the axis above is written down and the split waits for the content to reach it. The move that was worth making now is the one that fixes findability: a theming skill is no longer one of 107 general ones. When the stage packages earn their own directories, the order is already decided and nobody has to re-litigate it.

What is NOT here

The theming code — schemas/themes.ts, schemas/theme-art-intake.ts, scripts/check-theme-art.ts, scripts/gen-themes-css.ts, scripts/gen-avatars-css.ts — stays where it is. This is the kg graph: the instructions an agent follows, not the implementation. A skill and the code it describes are different node kinds, and moving code to sit beside its documentation is how a partition stops meaning anything.

Processes that run this skill

process	step(s) that name it
Ingestion subprocess — ingest a theme	Read the served stylesheet’s declarations; Read the guide’s own stated rules; Map values onto the shared palette ROLES; Record contradictions IN the source
L1 document ingestion — a document to the L1 source knowledge graph	Ingest the theme (calls a sub-process)