Psyclone modules can now drive Amazon Bedrock and Google Gemini directly — bringing native, low-latency generative-AI inference into the real-time pub/sub fabric for building sub-second AI applications.
Verified TLS/HTTPS networking with real handshake test coverage
The SSL build now links OpenSSL 3 and ships a network_https unit test that performs a real TLS handshake and encrypted HTTP round-trips over loopback with a runtime-generated certificate — with a clear certificate-verification hardening path on the roadmap.
Self-checking function + performance harness with tail-latency percentiles
The psytest suite was rebuilt into a context-chained set of stages exercising context switching, time sync, guaranteed & best-effort messaging, signals, and catalog retrieval — each stage emitting a PASS/FAIL + perf line with a final aggregated verdict. It now reports full p50/p90/p95/p99/p99.9 tail-latency percentiles, not just averages.
New unified unit-test framework + automated regression tracking
A ground-up CMSDK test framework: one PASS/FAIL line per object, per-test process isolation with timeouts (a crash in one test can’t take down the suite), self-healing shared-memory cleanup, and timestamped performance JSON with %-deltas for automated regression detection. 33 tests, all passing.
Leaner hot path: eliminated per-message allocations & syscalls
The message-routing fast path was rebuilt: a 36KB per-message ring copy replaced with a circular write index, a per-message heap allocation and no-match leak removed from routing, header receive reduced to a single allocation, and six redundant syscalls stripped from every send/read cycle.
UDP pub/sub latency cut ~35%
The same socket-mode and buffered-read optimisations lifted best-effort UDP messaging to 155,000 messages/second (up from ~107k) with latency down to about 6 microseconds (from 9.3us) — reinforcing Psyclone’s sub-10us real-time profile.
TCP messaging roughly 2x faster
By keeping sockets non-blocking instead of flipping mode around every send (which cost four syscalls per message and raced the receive thread), TCP throughput more than doubled — from ~42k to ~90k msg/s — and per-message latency roughly halved, from 23.8us to ~11us.
Concurrency & correctness hardening across the engine
Eliminated a class of silent data races: std::map::operator[] on read paths was inserting empty entries under a shared read lock (routing races + unbounded map growth). Fixed across SubscriptionEngine, Node, Monitoring, ServiceManager, InterSystemManager and the Catalogs, plus mutex-leak fixes and a NetworkChannel shutdown deadlock.
Fixed indefinite teardown hangs (timeout-emulation bug)
Two macOS timeout primitives treated an absolute deadline as a relative countdown, silently turning bounded waits into infinite hangs under contention — the root cause of intermittent 30-second teardown stalls. Now fixed and soak-validated.
Fixed a critical macOS shutdown-hang (thread-liveness UB)
Root-caused a heap-corrupting bug affecting the whole product: a destructive pthread_join was being used as a liveness probe, reaping threads so a later authoritative join operated on freed handles. Replaced with non-destructive pthread_kill(h,0) probing — eliminating intermittent shutdown hangs.