[ZacxDev]

civitai/civitai#7491 (per-collection document schemas) merged as a6053a9a, and civitai.com serves its manifest schema: schema-drift is green against the vendored copy here.

What

civitai/civitai#7491 adds storage.collections.<name>.schema, a JSON Schema subset that every document body in the collection must match (the store refuses a violating write with 400 schema_violation). This brings civitai app validate along with it:

1. Re-vendor schema/app-block.manifest.schema.json from #7491 (byte-identical to what merged). Before this, the collection object's additionalProperties: false refused the schema key outright, so validate rejected every manifest using the feature. The new schema carries the keyword subset, each keyword's shape, the 256-character pattern bound and "a pattern needs maxLength ≤ 1000" ($defs.storageCollectionSchema). No new regex pattern keyword, so the gloss ledger is unchanged.

2. internal/validate/storageschema.go ports the server's parseCollectionSchema rules that the JSON Schema cannot express, as ERRORS with the server's text:

   • at most 16384 bytes of JSON.stringify output in UTF-8. This is counted, not computed with json.Marshal, which escapes <>& and U+2028/9 and writes -0.
   • at most 16 nested schemas (the message names the innermost one)
   • at most 16 patterns per collection schema, and at most 32 across all of the manifest's collection schemas. This counts every place a pattern appears, and only in collections that are otherwise valid.
   • minLength/minItems/minimum not above its max
   • a count keyword not above 2^53−1
   • no schema on a write: "none" collection. The JSON Schema refuses this too, but its message is only "false schema", so the server's text is added next to it (the same approach as buildCommandRe).

3. Patterns: a warning, not parity. The server compiles a pattern with new RegExp(p, "u"). Go's RE2 disagrees with that language in both directions, so it is not used. internal/validate/jsregexp.go is a conservative syntax checker for ECMAScript u mode: it reports only constructs u mode throws on, as a WARNING with the reason, and says nothing on valid syntax. Every row below was checked against Node 24.21 with u:

   	patterns	CLI
   refused by u mode	\- outside a class, \_, \:, \#, \@, \z, \Q…\E, lone { } ], a{, a{,3}, [[:alpha:]], (?i), (?s), \pL, (?P<n>x), unbalanced ( ) [, malformed \x \u \c \k \0, [\B]	warning
   valid in u mode	lookaround, \1, \k<n>, \u{1F600}, \cA, \p{L}, [\-], \d\w\s\b\B, \/, {2,3}	nothing

   It stays a warning because the server is authoritative. A pattern the checker says nothing about is not proven valid: it does not check backreferences to missing groups, group or property names, quantifier bounds, or class-range order, and the server's submit-time ReDoS check (recheck) has no local counterpart. The help text and the README say so.

Tests (red first)

• e14457b: tests only, RED.
• 7eb6439: the re-vendor alone. The accept arm goes green; the server-only rules stay red.
• 4a333e3: the Go mirror. The new tests are green, but the field-coverage ledger went red: it could not name a method finding-producer.
• 84e269d: fixes that ledger. With 24f7c55 it is RED only on the 32-pattern manifest cap.
• 894490c: the cap mirror plus a re-vendor at 64d6412.
• 834e301: the pattern table (both lists, literal expectations), tests only, RED against the RE2 check.
• 91794d0: the u-mode syntax checker replaces RE2.

Expected values are literals taken from the server's tests and from Node's Buffer.byteLength(JSON.stringify(...)). They were not derived from the Go. The field-coverage corpus and the field ledger gain the new findings.

Not covered offline

• pattern semantics past syntax (see item 3)
• ReDoS
• document validation itself. The CLI checks manifests, not stored documents.

🤖 Generated with Claude Code