Vylinejs stability notes

Vylinejs keeps LINEJS as its protocol and API base. These safeguards bound long-running queues and external input while keeping Talk/Square event order and E2EE key history intact.

Runtime limits

Area	Limit and overflow behavior
Talk/Square push events	At most 200 pending events per stream. Producers wait for consumption; queued events survive reader cancellation and stream restart. Shutdown rejects blocked producers so the server can replay events whose revision was not advanced.
PLANET control messages	256 queued messages. Overflow ends the call and reports control_queue_overflow.
PLANET and Andromeda audio/RTP	100 queued packets and 1 MiB per transport. Oldest stale packets are dropped; an oversized packet is discarded. Andromeda SIP waiters are removed on timeout and rejected on close.
PLANET video receive work	128 pending packets and 1 MiB per serial queue. Old queued video work is discarded first to keep playback current.
Decoded video frames	At most 30 queued frames; the oldest frame is discarded when full.
Agent I / LINE AI SSE	At most 1 MiB per event. Early iteration exit cancels the underlying reader.
Agent I conversation	At most 40 messages by default; maxHistoryMessages can lower or raise this cap. User text and accumulated assistant text are limited to 1 MiB each.
LINE AI JSON responses	At most 4 MiB; larger responses are canceled and rejected.
TMoreCompact	Varints are limited to 10 bytes/uint64; field bitmaps stop at supported field IDs below 30; declared strings, tables, collections, maps, and headers must fit the remaining input. Parser state resets on reuse.

E2EE key generations are retained for historical decryption. No automatic key eviction is performed.

LIFF send-token caching is scoped to account MID, LIFF ID, and recipient, with a 128-entry LRU bound to cap memory on long-running clients.

LINE VOOM ended on 2026-09-30. Its old client.voom, client.voomRest, and client.base.timeline APIs remain temporarily for source compatibility, are marked deprecated, and now fail immediately without issuing tokens or making network requests. Remove their use from applications; there is no replacement service announced. See LINE's official shutdown notice. The compatibility exports are planned for removal in a future major version. Generated protocol enums and fields remain where needed to decode historical LINE data.

FileStorage measurement

Run deno run -A scripts/bench-file-storage.ts to measure 30 reads and 5 writes against a temporary JSON file containing 50,000 keys. On Deno 2.7.14 / Windows, the original implementation was measured once; the optimized version uses the median of three runs:

Operation	Before cache	After cache
get average	39.56 ms	0.20 ms
set average	73.98 ms	20.75 ms

These are single-machine microbenchmarks, not production workload guarantees. Reads use an in-process snapshot and check file metadata for external edits; writes preserve the JSON format and atomically replace the file. On POSIX, construction and every write restrict the credential file to owner read/write (mode 0600), including files that were previously more permissive. For five updates in a 50,000-key file, five individual writes averaged 16.0 ms each (about 80 ms total); FileStorage.setMany() committed the same five updates in 15.9–16.5 ms across two runs. At 1,000 keys, five writes averaged 1.41 ms each and the batch took 2.00 ms. The batch preserves the JSON file and uses one atomic replacement; encrypted storage batches encryption before delegating to its backing store. Complete login credential writes and paired E2EE group-key records use the batch API when available; older duck-typed storage adapters retain their sequential-write fallback.

Storage migration previously called set() once per key. Migrating a MemoryStorage snapshot into FileStorage took 330.56 ms for 250 keys and 1,194.11 ms for 1,000 keys. After routing the three built-in migration methods through setMany(), it took 3.11 ms and 5.38–7.71 ms, respectively. A 10,000-key batch completed in 21.22 ms. Reproduce with deno run -A scripts/bench-storage-migrate.ts [keyCount].

Performance experiments

Talk Push TMoreCompact payload wrapping (Deno 2.7.14 / Windows, 10,000 conversions per run, five runs) now shares the incoming ArrayBuffer instead of copying each response. In a conversion-only microbenchmark, median time changed from 5.55 to 1.27 ms for 256-byte payloads, 17.49 to 0.94 ms for 4 KiB, and 89.60 to 0.69 ms for 16 KiB. The benchmark excludes protocol parsing and network work; reproduce with deno run -A scripts/bench-push-buffer-view.ts. Non-UTF-8 decoded fields still copy their own bytes so they do not pin the entire incoming frame chunk in memory.

Thrift serializer (Deno 2.7.14 / Windows, 10,000 compact-protocol messages, 1,000 warm-up calls, five measured runs): direct chunk-to-output copying removes the intermediate concatenated body and second full-body copy. One measured comparison was 309 ms before and 313 ms after, within run-to-run noise; no throughput improvement is claimed. This change targets peak buffering and copy count. Re-run deno run -A scripts/bench-thrift-encode.ts to compare.

PLANET decode experiment (same runtime, 20,000 representative INFO_REQ packets, 2,000 warm-up calls, five runs): the former outer-field parse plus message parse had an 86.44 ms median; the single message parse had a 77.04 ms median (10.9% less time). Debug-only structural decoding remains available when a debug hook is configured. Re-run deno run -A scripts/bench-planet-decode.ts to compare.

EVS3 reassembly (Deno 2.7.14 / Windows, 64,000-byte VP8 frame, 1,000-byte fragments, 500 frames per run, five runs): reusing the descriptor already parsed by Evs3Assembler.push() reduced the median from 59.0 ms to 54.8 ms (about 7%). The previous offset-caching experiment had no robust gain and was reverted. Reproduce with deno run -A scripts/bench-evs3.ts.

SRTP packet processing (Deno 2.7.14 / Windows, 1,000 160-byte RTP packets, encrypt plus decrypt, five measured runs): static crypto imports and synchronous internal AES-CM processing reduced a paired-run median from 53.5 ms to 44.0 ms (about 18%). Results vary by run; reproduce with deno run -A scripts/bench-srtp.ts 1000.

LEGY request encryption (Deno 2.7.14 / Windows, 1 KiB request bodies, 1,000 mocked requests per run, two warm-ups and five measured runs; empty mocked responses): removing intermediate copies during header/payload assembly and wrapping the encrypted bytes as a view brought two run medians from 116.1/111.9 ms to 97.3/97.1 ms (about 13–16%). The benchmark measures the request encryption path without network latency or response decryption. Reproduce with deno run -A scripts/bench-legy.ts 1000.

The SRTP packet path now passes typed-array views directly to Node crypto, removing per-packet Buffer.from() copies for keys, payloads, and HMAC input, plus the intermediate Buffer.concat() and output copy. Against the previous implementation, two 1,000-packet paired-run medians improved from 37.1/38.9 ms to 33.3/32.6 ms (about 10–16%). These local microbenchmarks are sensitive to runtime load; reproduce with deno run -A scripts/bench-srtp.ts 1000.

Compact text encoding (Deno 2.7.14 / Windows, 500 encodes per run, five runs) now measures UTF-8 length before choosing an encoding, then allocates only the selected representation. For 13,000 ASCII characters, the median fell from 56.8 ms to 22.6 ms; for 7,000 Japanese characters, from 70.4 ms to 25.0 ms; and for mixed text, from 70.0 ms to 33.6 ms. Reproduce with deno run -A scripts/bench-compact-text.ts.

Group audio mixing now validates decoded PCM in one pass and removes expired jitter chunks in place. The transport skips receive/RTP debug-record creation when no debug callback is configured. Andromeda removes timed-out SIP waiters and cancels them on close. These paths have regression coverage; no separate throughput claims are made for them.

Upstream issue status

• LINEJS #198: reply transaction handling and remote-release behavior have local synthetic UDP/SRTP regression coverage. The reported 60-second live result used an experimental JavaScript vendor; a separate live phone call using the TypeScript implementation has not been confirmed, so the issue remains open.
• LINEJS #211: generation-aware group-key lookup and multi-entry E2EE key handling are present in the current baseline. Regression cases pass locally, but the original reporter's account remains unverified. Preserve historical key material; do not force key registration or clear storage.
• The last issue review recorded here was 2026-10-29. Account and device confirmations require access to the affected account or a real phone pair; synthetic tests cannot establish those results.