Introduce app_asset_contents table as a read-through cache over S3 for
text-like asset files (e.g. .md skill documents). This eliminates N
individual S3 fetches during SkillBuilder builds — bulk_load pulls all
content in a single SQL query with S3 fallback on miss.
Key components:
- CachedContentAccessor: DB-first read / dual-write / S3 fallback
- AssetContentService: static DB operations (get, get_many, upsert, delete)
- should_mirror(): single source of truth for extension-based policy
- Alembic migration for app_asset_contents table
Modified callers:
- SkillBuilder uses accessor.bulk_load() instead of per-node S3 reads
- AppAssetService.get/update_file_content route through accessor
- delete_node cleans both DB cache and S3
- draft_app_assets_initializer uses should_mirror() instead of hardcoded .md
Refactor skill compilation around mergeable bundle patches so dynamic skill updates no longer require full rebuilds. Keep dependency closures accurate by recomputing affected nodes from direct dependency data.
Migrate ArtifactsSection to queryOptions + useQuery composition and derive\nfile tree/hasFiles locally from flat data. Remove the now-unused\nuseSandboxFilesTree helper and update related tests to mock the new\nqueryOptions-based flow.
Extract sandboxFilesTreeOptions and buildTreeFromFlatList from
useSandboxFilesTree so callers that need custom TanStack Query behavior
(e.g. refetchInterval) can compose at the call site instead of tunneling
options through the hook. Remove the thin useGetSandboxProviderList
wrapper in favor of inline oRPC queryOptions in the component.
Also remove redundant .input(type<unknown>()) from three no-input GET
contracts—oc already defaults TInputSchema to Schema<unknown, unknown>.
- move MCP availability context to block-selector/context and update imports
- preserve sandbox gating, parent-provider inheritance, and blockedBy semantics
- add context tests on top of refactor baseline cases
- regenerate and prune eslint suppressions
Harry