mirror of
https://github.com/block/buzz.git
synced 2026-11-06 17:31:19 +00:00
237d794acb
## Summary Read-state blobs are NIP-44 encrypted, and the app decrypts them with pure-Dart ChaCha20 and HMAC-SHA256 on the UI isolate. Most of those decrypts repeat earlier work: - Every live subscription (each launch, resume and reconnect) replays all of the account's read-state slots, and `_handleIncomingEvent` decrypts each one again. The measured account has 27 slots of about 55 KB ciphertext each. - Before every publish, `_publishOnce` fetches the read-state snapshot and decrypts it, and decodes our own slots a second time to check they're readable. - The relay echoes each publish back, and our own blob is decrypted again. `ReadStateCrypto` now keeps a bounded memo from ciphertext to plaintext, one per conversation key. `decrypt` checks it first, and `encrypt` records the ciphertext it produced. NIP-44 payloads carry a random nonce and a MAC, so under one key the same ciphertext always decrypts to the same plaintext. Failed decrypts aren't cached and still throw every time. The memo is least recently used: a hit moves the entry to the newest end, so our own superseded blobs are evicted before live ones. It holds at most 8,000,000 code units across keys and values (about 8 MB, about 80 full slots). The first version used oldest-first eviction with a 2M bound, which saved nothing on the device: the 27 slots total about 2.6M, and a replay in slot order evicted each entry just before its next use. ## Measurements iPhone 12 mini, Profile/AOT, Tom's Block community. Each run resumes the app twice from the home screen and samples UI-isolate CPU for 15 s from each resume. | Build | `ReadStateCrypto.decrypt` per resume | | --- | --- | | `main`14037e3| 73 ms, 78 ms | | `main` + this PR + [#8965](https://github.com/block/buzz/pull/8965) | 9 ms, 4 ms | On the older base3d26ed7the first version of this PR measured 74/70 ms before and 73/71 ms after (no gain). An instrumented build showed 27 decrypts per resume with 0 memo hits, all on one `ReadStateCrypto` instance. With the LRU fix, the same base measured 7/3 ms. ## Testing - New tests: the same ciphertext twice runs the cipher once. A ciphertext from `encrypt` runs it zero times. Failed decrypts throw every time. The bound caps retained size. - Regression tests for the device finding: replaying 27 full-size slots twice decrypts 27 times, not 54, and a hit keeps a live entry ahead of superseded blobs under a small bound. Both fail on the first version's oldest-first memo. - Manager-level test: after the first sync, three publishes plus the echo and snapshot fetches add no decrypts (3 total, instead of 6 with the memo bypassed). - `just mobile-check mobile-test` passes (pre-push hook), after merging current `main`. - Automated device run in place of manual testing: the build above on the 12 mini, launched and resumed twice into a signed-in Block community, loading the channel list and syncing read state (snapshot fetch and live replay of all 27 slots). 🤖 Generated with [Claude Code](https://claude.com/claude-code) --------- Signed-off-by: Tom Brow <tomb@squareup.com> Co-authored-by: Tom Brow <tomb@squareup.com> Co-authored-by: studio <dc2fc481343672df31a689840c09299972e704dac592e0da6f18486646a7ee3e@buzz.block.builderlab.xyz> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>