mirror of
https://github.com/vitorpamplona/amethyst.git
synced 2026-10-06 11:48:24 +00:00
Claude
f5fc432e95
Two clusters from the survey of what was left in amethyst/, both cases of code that was shared-layer already and just happened to live in the app. The four Marmot stores go to commons/jvmAndroid/marmot, next to the EncryptedAppendLog they were already built on. Nothing in them was Android despite the prefix: no android.* import, no Context, just (rootDir: File, encryption: SecretEncryption) and the quartz interfaces. One call site, AccountCacheState. The KDoc lines claiming "Android implementation" and "AES/GCM via Android KeyStore" are reworded: SecretEncryption is expect/actual, AndroidKeyStore on Android and a key file on desktop, so the old wording is now only half true. Commons already referred to two of them by name in its own KDoc; those references are updated. They are Encrypted* rather than File*, and that is not cosmetic. cli/stores already declares FileMlsGroupStateStore, FileMarmotMessageStore, FileKeyPackageBundleStore and FilePublishObligationStore — deliberately UNENCRYPTED test-harness stores, in a module that depends on commons. Two same-named classes with opposite encryption semantics, one import away from each other, is how MLS state ends up written in plaintext. Encrypted* is also what these actually are, and reads correctly next to EncryptedAppendLog. Nip11CachedRetriever + Nip11Retriever + RetrieveResult go to commons/relays/nip11, and OnlineChecker to commons/service. RetrieveResult had to travel: the cache is typed on it and commons cannot import from amethyst. Nip11RetrieverTest travels too — it sat in the same package and used the class with no import line. LoadRelayInfo and RelaySupportsNip stay behind; both reach Amethyst.instance. android.util.LruCache -> androidx.collection.LruCache is not the pure import swap it looks like, and the compiler said so: androidx bounds V to Any, while android.util.LruCache is a Java platform type that accepted LruCache<NormalizedRelayUrl, RetrieveResult?>. Checked all four put() sites first — every one stores a concrete RetrieveResult, so the nullable argument never meant anything and get() still returns null on a miss, which is what the readers already branch on. Dropping it is behaviour-preserving. NotifyCoordinator does NOT move despite being grouped with the other two: android.util.LruCache really is its only platform import, but it takes accountForPubkey: (HexKey) -> Account?, and Account lives in amethyst/model. It needs that abstraction, not an import swap. 17 new tests. OnlineChecker's predicates decide whether the UI shows a player or an offline placeholder, so the five-minute TTL is pinned in both directions — a stale online entry must stop reading online, and a stale offline one must stop suppressing retries — along with resetIfOfflineToRetry dropping only failures, since evicting good entries would refetch every URL that already worked. Its suspend probe is left alone: it needs a real OkHttp round trip and commons has no MockWebServer, so there is no honest way to drive it. Same reason Nip11CachedRetriever gets nothing new here; its fetch and error-caching paths are all network. The Marmot tests cover what restart depends on: group state, sender ratchet and message log surviving a new store over the same directory, deletes removing both the state and the listing, and two account directories not seeing each other. Writing them found validation I had not noticed reading the code — group ids must be hex, which is a path-traversal guard — after a first pass using readable labels failed every test on it. That guard is pinned now too: "../escape" and friends are rejected rather than resolved to a path. Verified: :commons:jvmTest (2549, 0 failures), :amethyst:testPlayDebugUnitTest, :cli:compileKotlin, :commons:verifyKmpPurity, :commons:compileCommonMainKotlinMetadata, :amethyst:compilePlayDebugKotlin. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AXvKXakvup4inNFfAhhr4L