[arkavo-com]

Summary

Implements the vrm-author CLI proposed in docs/proposals/vrm-author-cli/: all 35 v1 commands, the native-anime-v1 template, GLB/VRM 1.0 export, spec+style and authoring-v1 QA with real Metal renders, provenance/rights ledger, Ed25519 delivery sidecar, JSON-RPC and MCP serve.

The template is a VRoid-lineage anime avatar (not a low-poly mannequin): 1024²/2048² feature textures via a deflate PNG encoder, lid UV isolation, blink/gaze/smile morphs covering all 18 VRM 1.0 preset expressions, fingered hands with hand-sigma skinning, 6.9-head proportions, garment shaping with trim bands and per-garment palettes, bob/long/ponytail hair via LayoutParams, cat-ears accessory, and 40 spring chains on the default bob.

Also lands the reworked verification contract (pinned oracles, evidence levels, per-command required evidence, staging), the acceptance-pack schema and runner (scripts/acceptance_run.py), and 35 acceptance packs pinned to the tree.

Targets

• VRMAuthorKit (Foundation + CryptoKit only), vrm-author, vrm-author-render (VRMMetalKit adapter), VRMAuthorKitTests

Verification

Local at fa80343 (fresh build, seed 42, native-anime-v1):

• project init → build → QA succeeded
• Style lint vs docs/style/profiles/vroid-lineage-anime.json: conforming — 15/15 must, 51/51 should, 9/9 may, zero failures
• QA suite authoring-v1: 22/22 checks pass (1 notApplicable, expected first-build baseline)
• Whole package: 2654 Swift tests green; acceptance harness 133 tests green; packs re-pinned

Follow-up at 5c52645 (CI Static checks):

• Dropped URL(fileURLWithPath: "/") / "/usr/bin/false" literals that failed .github/workflows/lint.yaml
• EvidenceRegistry.locate and StyleToolchain.ancestor now stop at filesystem root (pathComponents.count <= 1). FileManager.default.temporaryDirectory URLs walk "/" to "/.." instead of a fixed point, which hung the locator.
• DiscoveryPackTests 9/9, ConsumerImportTests.testReportParsingRejectsWrongConsumerAndSurfacesErrors, acceptance/packs/version.json, and scripts/repin_packs.py --check all pass. Packs re-pinned onto 11188c9.

Open under the contract

• acceptance/evidence.json has no evaluator receipts, so production capabilities / MCP tools/list expose no production-eligible tools until an independent-family evaluator admits evidence
• export verify / deliver stay incomplete until inspection records exist for the render scenarios
• visually-validated commands await the §5 judge

Scope cuts

• Ed25519 JSON sidecar, not COSE/JUMBF C2PA
• macOS-only determinism evidence (no Linux build)
• Remaining VRoid Studio gap is hand-painted art content. Clothing texture micro-detail is structurally blocked by per-loft overlapping UVs and would need a UV-atlas redesign.

[gitar-bot]

💡 Bug: Ledger self-match short-circuit skips no useful check but is fragile

ledgerConflicts matches a ledger asset's stored declaration against the incoming declaration by id and reports LEDGER_CONTRADICTION only when the content differs (ingredient != declaration); when they are equal it continues, skipping the modification/redistribution/commercial/usage-flag checks for that asset for this call. That's fine when the asset truly is the same declaration being re-resolved, but if two distinct ledger assets ever legitimately share a declaration id (e.g. two imports citing the same rights declaration), the second one's ingredient-term checks (prohibited modification, redistribution, commercial usage, etc.) are silently skipped instead of being evaluated. Consider matching on asset identity plus declaration equality separately from the ingredient-term checks so term checks always run regardless of the id coincidence.

Run term checks unconditionally; only skip the contradiction check itself when the ids coincide by design.:

for asset in ledger.assets {
    guard let ingredient = asset.declaration else { continue }
    let path = "/ledger/\(asset.id)/declaration"
    if ingredient.id == declaration.id, ingredient != declaration {
        conflicts.append(RightsConflict(code: "LEDGER_CONTRADICTION", ...))
    }
    // term checks always run, even when ingredient.id == declaration.id
    let terms = ingredient.meta
    ...
}

_{Was this helpful? React with 👍 / 👎}

[gitar-bot]

💡 Quality: exposedOperations() recomputes evidence capability on every tools/call

tools/call and tools/list both call exposedOperations(), which filters context.registry.ordered and, for each operation, calls context.evidenceRegistry.capability(for:toolInfo:policy:) — this repeats the full capability computation for every operation on every single tool invocation instead of caching it once per session/policy. For a large registry (35 v1 commands, more as they grow) this is wasted work on the hot path of every tools/call, though not currently large enough to be a real bottleneck.

Memoize exposedOperations() per ServeSession instance.:

// Cache the exposed operation list (and capability lookups) once per session
// instead of recomputing per request; invalidate only if policy/harness changes.

_{Was this helpful? React with 👍 / 👎}