Authored Asset Validation¶

Your code is checked by a compiler, an analyzer and a test suite. Your assets are checked by nobody. A designer forgets a reference, an artist deletes a sprite, a serializer migration retires a field — and the project still compiles, the tests still pass, and the damage shows up as "the thing flickers" or "no decal appeared", which nobody files as a bug.

These checks read committed .unity, .prefab and .asset files as text and report what is wrong, with the line to open.

Commands¶

Every command lives under Tools > Wallstop Studios > Unity Helpers > Authored Assets.

Every command but the last only reads. The repair rewrites files and asks first.

The same checks, running continuously¶

Four of these five also ship as IValidationRule implementations, so they run a few milliseconds at a time through the Asset Validation engine — in the window, on import, and in CI — instead of only when somebody remembers the menu. They are the same code asked one asset at a time, so a finding here and a finding there are the same finding.

Stale serialized keys is the one that is not, and the reason is its instrument: the declared-key set is SerializedObject over a throwaway instance, so answering the question constructs a ScriptableObject or adds a component to a hidden GameObject for every type it meets. That is a consumer's own constructors and OnEnable running from inside an editor tick, which is exactly what a continuous engine must not do. It stays a command you invoke deliberately.

The severity each rule reports at, the rule ids, and what a rule says about an asset it could not read are all in Asset Validation.

What a scan could not read¶

Every report names the files it could not open, and each command prints that set when it is not empty. A read fails for reasons that are not going away: a permissions error, a file another process has locked, a file deleted between enumeration and the read, an .asset saved in binary serialization mode, an I/O error on a network drive. The animation check fails differently — the asset database names a path as carrying a clip and then hands back none — and reports it the same way.

An unreadable file is not a finding. It is a hole in the measurement, not a defect in the asset, and folding the two together would make a finding mean two things: a caller could no longer read "no findings" as "nothing is wrong". So TryScan takes a list of unreadable asset paths beside its findings, sorted and naming each file once, and a CI caller that asserts the list is empty fails when a scan could not see all of its subject.

The subject counts cannot catch this on their own. They catch a scan that read nothing — a moved root, a renamed backing field. One locked file in a project of four thousand still reports a large count and a clean result, which is the failure these checks exist to prevent turned on themselves.

Why the set prints without turning the log yellow¶

Expect entries here in a project with baked lighting, and expect them permanently. Unity writes LightingData.asset as binary whatever the serialization mode says — measured on two of two under ForceText — so the scan opens it, finds no Unity document in it, and correctly reports that it could not see inside.

That is worth saying and not worth warning about. A warning claims there is something to fix, and a permanently unactionable one is exactly what teaches people to stop reading warnings. So the set always prints, and the severity follows the findings. A gate that wants to fail on a coverage hole asserts the list, which is what the scan returns it for.

Why text, and why loading is the wrong instrument¶

Opening a scene mutates it. Every OnValidate in it runs — a collider that rebuilds its points from the sprite's physics shape is enough — so the act of inspecting dirties the scene, and closing it then prompts for a save. A prompt blocks the editor's main loop, which is fatal for anything driving the editor over a bridge. A gate must not mutate what it measures.

Text also reaches a baked scene without opening one. And for the stale-key question, loading asks the same serializer that dropped the data what the data is, and it answers "there is no such field" — indistinguishable from "the field is empty".

The one question text cannot answer is animation: a keyframe's guid can resolve perfectly while the object does not, because a sprite sheet re-imported as Single still has a .meta describing every slice the importer no longer produces. That check uses AnimationUtility.

AuthoredAssetYaml.TryResolveObjectReference reads both external and same-file references. Unity omits guid when one document points to another document in the same scene, prefab or asset. Pass the owning asset GUID and the method supplies it for a nonzero fileID; an empty reference keeps a null GUID. Invalid inline mappings and local references without an owning GUID return false with empty outputs.

Script bindings¶

A MonoBehaviour or ScriptableObject reaches a scene, a prefab or an asset through exactly one door: the MonoScript Unity builds for the file that declares it. Unity picks the class for a file by name and falls back silently when nothing matches, so a file declaring two types can bind the wrong one — and the type you care about then has no door at all. It still compiles, and AddComponent<T>() still constructs it, so every behavioral test passes. The gap appears only when somebody tries to author the thing, where it reads as "the component will not drag onto the prefab".

Two rules, and neither is redundant:

1. Every concrete MonoBehaviour and ScriptableObject resolves to a MonoScript — the symptom.
2. Every script asset is named after the type it binds — the cause, and what keeps rule 1 true.

Nested and open-generic component types are not excluded: neither can carry a MonoScript either, so excluding them licenses the same defect in a shape the check has stopped looking at. Abstract types are, because nothing can be an instance of one, and so are custom inspectors: Unity finds an inspector through [CustomEditor] on the type rather than through a saved reference to its script, so nothing an author saved names it. An EditorWindow stays in scope, because a saved window layout does name its script.

Unfilled required fields¶

WNotNullAttribute draws a warning beside the field in the inspector and does nothing else, and a drawer needs somebody looking at the inspector. So the package tells an author "this must be assigned" and a build ships with the slot empty.

The subject set comes from the annotations through TypeCache, never from a hand-listed set of fields: a list has to be updated by the change that adds a field, and forgetting makes the check pass. "Unfilled" is the drawer's own answer rather than a second one — a reference is empty when it names no object, a string when it is blank — so the build and the inspector cannot disagree about the word.

A key the document does not carry at all is not reported: that is a different state — the asset predates the field — and reporting it would report every asset of a type the moment a field is added to it. The reported set is slots an author saw and left empty.

An annotated field the check cannot read is reported as a budget rather than skipped: a type with no MonoScript (an inline-only [Serializable] class) appears in no document, and a check that quietly cannot see part of its subject is the failure this exists to prevent.

AuthoredRequirementValidator.TryScan takes the attribute type, so a project's own "author must fill this" attribute gets the same treatment.

Serializable dictionaries¶

Two states an authored map can be in:

• Keys written, values absent. The asset records what the mapping is about and nothing about what it maps to. It still loads, the dictionary is empty, and every lookup misses.
• A key whose value names no object. The asset is well-formed, so TryGetValue returns true with a null and every caller has to remember a second check.

A dictionary whose value type is itself a collection stores its values in _boxedValues, because Unity drops an array whose element type is a collection. So "no _values" is not by itself a defect, and the check compares the keys against whichever array carries the values.

The dictionaries-inspected count is an output, so a caller can refuse a vacuous pass: a moved root or a renamed backing field otherwise turns the check green in silence.

Stale serialized keys¶

Unity keeps an unknown serialized key on load and writes it straight back out, so a field deleted years ago is still in the YAML today and reads exactly like a live one. Beyond diff noise, a retired object-reference key keeps another asset looking referenced — a guid search answers "this sprite is used by the level prefab" about a slot nothing reads.

The declared set is SerializedObject over a throwaway instance, never reflection over the type's fields: only Unity knows what its serializer accepted. Every [FormerlySerializedAs] alias counts as a live key, walking the base chain — judging by SerializedObject alone was measured reporting 565 aliases doing their job as orphans. A document whose script does not resolve is counted, not reported: that is a missing script, a different defect with its own signal.

The engine's own header keys are listed rather than discovered, because the probe cannot report them: a ScriptableObject asset is written with the full MonoBehaviour header — m_GameObject, m_Enabled, m_EditorHideFlags — while a SerializedObject over a ScriptableObject reports none of it. So is the references block a [SerializeReference] document ends with.

Findings are grouped by cause (Type::Key), because a migration retires a field once and every asset of that type inherits it.

The repair, and why it is a wrapper¶

There is no "delete this key" API. AssetDatabase.ForceReserializeAssets rewrites an asset from what it loaded, dropping every dead key. It is not safe unsupervised:

• An asset whose content lives in sub-objects can come back with them gone. A render profile measured going from twenty serialized documents to one, losing every volume component, while the rewrite reported success.
• Restoring the file is only half of an undo. The editor still holds the damaged object and the next save writes it straight back out; the other half is a forced synchronous re-import.
• Prefabs need the metadata option. With assets-only they are silently not rewritten at all.

So the repair rewrites one asset at a time, compares the non-null object count before and after, and undoes any rewrite that lowers it — or that throws — by writing the original bytes back and re-importing. Refusals are printed, and each one says which of the five things happened: the file could not be read, the rewrite lost objects, the rewrite threw, the undo itself failed, or nothing changed. Commit or stash first.

The rewrite is covered, not just the refusals: a fixture authors a plain asset, an asset whose content lives in sub-objects, and a prefab, leaves a key no field claims in each, repairs them, and asserts the outcome and the object count. It pins the prefab finding directly — the same prefab rewritten with assets only comes back byte-identical, stale key and all.

The two branches nothing authorable can trigger — a rewrite that loses content, and a rewrite that throws — reach production code through one seam each, and everything after the seam is the real thing: the comparison, the error log, the byte restore, and the forced re-import. A render profile loses content and a test cannot build one; nothing at all makes ForceReserializeAssets throw. The confirmation dialog stays anyway, and committing first is still the advice.

The reader underneath¶

AuthoredAssetYaml parses a committed file into the documents Unity wrote — each with its !u! class id, anchor, line range, and every key at any depth — and MonoScriptIndex resolves a document's m_Script guid to a type and back. They exist so the checks above share one parser rather than four, because four parsers are four chances for one of them to drift into reporting clean.

MonoScriptIndex's forward lookup is name-narrowed first, which is fast and correct while a script asset is named after the type it binds — the second script-binding rule above. When it is not, the search falls through to a full index rather than to null, because a resolver that silently answers "no such type" makes every "except its own file" exclusion match nothing, and a check can then pass for every type at once.

Related¶

• Asset Validation: the continuous, time-sliced validation engine
• Analyzers: WUH012 reports the code half of the same problem