Drove the native 2023 R2 client through a committed non-streamed (historical backfill) write to a
sandbox tag, with the IL-rewritten managed gRPC client dumping every byte[] payload. Read the value
back over gRPC = end-to-end validated.
Key discovery: the M3 historical write does NOT use AddNonStreamValues/TransactionService (the
roadmap's assumption from the static decompile). The native client routes it over
HistoryService.AddStreamValues with an "ON" storage-sample buffer (structurally the AddS2 "OS"
family), plus EnsureTags for registration:
AddStreamValues.values (56B): "ON"(0x4E4F) + u16 count=1 + u32 totalLen + u16 payloadLen +
16B tag GUID + FILETIME(sample) + u16 quality=192 + u32 type/desc + FILETIME(received) +
8B double value.
EnsureTags.tagInfos (144B): the analog CTagMetadata the SDK's EnsureTagAsync already builds
(0x4E marker ... fe 00 trailer).
Tooling that produced it:
- instrument-grpc-nonstream now instruments EVERY byte[]-input method on every Grpc*Client
(45 methods) so the real wire path surfaces regardless of assumptions.
- harness pre-loads the instrumented GrpcClient by identity (LoadFrom context reuses an
already-loaded assembly before sibling-probing, so the rewrite wins over the bin original);
capture-write sequence fixed to Begin -> Add -> SendValues -> End (End-before-Send = err 160
InvalidBatchId); GetTagInfoByName(cache:false) + resync wait resolves the server key; cache
gate (D2's 129) does NOT block the primed 2023 R2 client.
Buffers captured to gitignored artifacts/. Next: build the "ON" AddStreamValues serializer in
src/ (adapt the existing AddS2 "OS" serializer) + EnsureTags + ship AddHistoricalValuesAsync.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01B6mcaT2PjRFKcogzp9UkfC
Joseph Doherty
9bcfffb365