Re-ran all seven domain reviews at master f6eaa267 (reports rewritten in
place, each with a prior-finding status table): all 4 round-1 Criticals
verified closed; new top findings are the S7 connect-timeout OCE regression
(05/STAB-14), the ResilienceConfig operator-authorable brick (01/S-6), and
a batch of resilience-seam Mediums. 00-OVERALL.md carries the updated
maturity matrix + 12-item action list.
Adds R2-01..R2-12 design/implementation plans (one per action item, house
format + bite-sized TDD task breakdowns + co-located .tasks.json; 193 tasks,
~18-24 dev-days). STATUS.md updated: round-1 topology marked historical
(all merged+pushed), re-review findings table + plan pointers added.
Consume the published 0.2.0 gateway packages. M3 opc_quality gate checked
and CLEAR: OtOpcUa never writes HistorianHistoricalValue (uses only the
WriteLiveValues path, no opc_quality field), and every OpcQuality=0 is on
the read-side HistorianSample in test fixtures — unaffected by 0.2.0's
HistorianHistoricalValue.opc_quality proto3-optional change. Build clean;
gateway-driver 99/0 and runtime 364/0 tests pass.
- CapabilityInvoker: 3 methods that implement IDriverCapabilityInvoker now use
<inheritdoc /> instead of duplicating the interface docs (MissingInheritDoc).
- TwinCAT NativeRegistration ctor: added the 8 missing <param> tags (MissingParam).
- Stripped the arch-review bookkeeping IDs STAB-1 (IS7Plc, S7Driver) + STAB-2
(TwinCATDriver x4) from comments, keeping the descriptive text.
Left the other 86 flagged items untouched — they are checker false positives on
legitimate technical content: interface-implemented events / IEqualityComparer /
IDisposable members whose <inheritdoc/> IS correct (10 InheritDocMisused), and domain
terms the heuristic misreads as tracking IDs (IEEE-754, S7-1200/300/400/1500, PLC-5,
D2-260, S7/Melsec memory addresses, OPC UA Part 9, phase-N design-doc filename refs,
UNS level-N descriptions, and prose like 'the #1 bug source' / failover 'pick #3').
Flip the #12 rig-bug note from 'recommend fixing' to fixed: FOLLOWUP-10 + STATUS task #12
+ new branch topology row. central-1 verified booting clean with the committed compose.
The historian-gateway cutover made ServerHistorianOptions read Endpoint/ApiKey/UseTls,
but docker-dev compose still set Enabled=true with the retired Wonderware Host/Port/
SharedSecret keys → HistorianGatewayClientAdapter.Create built a Uri from an empty
Endpoint → unhandled UriFormatException crashed the Host at every bring-up, leaving
docker-dev unbootable against current code.
central-1 + central-2 now use the gateway-shape keys, DISABLED by default (the rig has no
historian gateway): Enabled=${OTOPCUA_HISTORIAN_ENABLED:-false}, Endpoint/ApiKey via env,
UseTls + AllowUntrustedServerCertificate. To enable, set OTOPCUA_HISTORIAN_ENABLED=true +
OTOPCUA_HISTORIAN_ENDPOINT + export ServerHistorian__ApiKey (never committed). Verified:
compose config valid, central-1 boots clean (no UriFormatException; OPC UA + AdminUI serving).
Found during the arch-review #10 live-gate rig session; branched off master (independent
of the archreview work).
Ran the #10 live gate: built the Host image from source (crit13 tip; Gitea feed answers
200 anonymously so restore worked), stood up the docker-dev rig, central-1 booted healthy
on the new invoker wiring. Prod DI binds the REAL DriverCapabilityInvokerFactory —
confirmed live + by a new deterministic guard (ResilienceInvokerFactoryRegistrationTests,
crit13 8a8b9ec5). The behavioral retry/breaker log-line was NOT observed: reading the
Modbus driver showed its cheap paths don't throw a wrapped exception (Discover offline,
Subscribe lazy, reads non-wrapped) — only an idempotent write-to-dead-endpoint or S7 does,
not worth the multi-dependency chase since the retry behavior is deterministically proven.
Also found a rig bug: docker-dev compose's retired Wonderware ServerHistorian keys
(Host/Port/SharedSecret, Enabled=true) crash the Host at bring-up (historian-gateway cutover
now reads Endpoint -> empty UriFormatException). Worked around with an override; recommend
fixing the committed compose. FOLLOWUP-10 + STATUS updated; crit13 topology head -> 8a8b9ec5.
AddOtOpcUaDriverFactories (the driver-node bootstrap) must bind the real
DriverCapabilityInvokerFactory — not the NullDriverCapabilityInvokerFactory
pass-through — and it must mint a real CapabilityInvoker per instance. This is the
deterministic DI half of the #10 live gate: a refactor dropping back to the
pass-through would make the whole retry/breaker/bulkhead pipeline inert in production
while every pass-through-defaulting unit test stays green. 3 tests (real factory bound,
real invoker minted, shared singletons), non-fixtured. Host.IntegrationTests.
The DriverInstance.ResilienceConfig column was authored in AdminUI, persisted to the
entity, and serialized into the deployment artifact by ConfigComposer — but the runtime
read path dropped it: DriverInstanceSpec didn't carry it and the invoker factory always
passed null, so every driver got tier defaults regardless of its configured overrides
(a silent dead-config gap — #10's residual sub-finding).
Read-path plumbing (write side was already complete):
- DriverInstanceSpec gains ResilienceConfig; DeploymentArtifact.TryReadSpec reads the column.
- IDriverCapabilityInvokerFactory.Create takes resilienceConfigJson; DriverHostActor.SpawnChild
threads spec.ResilienceConfig; the concrete factory parses it (ParseOrDefaults, layering on the
tier), logs any parse diagnostic (never throws), and builds the invoker with the merged options.
- Invalidate-on-change: the pipeline cache keys on (instance, host, capability) and ignores options
on a hit, so Create() now Invalidate()s the instance's cached pipelines first (no-op on first
spawn) — a respawn with changed options rebuilds them.
- DriverSpawnPlanner treats a ResilienceConfig change as a stop+respawn (the invoker/options are
bound to the child at spawn); a pure DriverConfig change stays an in-place delta (no reconnect).
- Host DI passes a logger to the factory for the parse diagnostic.
Verification (deterministic): factory Create applies a retryCount:0 override to actual execution
(control test proves tier default retries), invalidates the instance's cache on re-create (scoped —
sibling survives), malformed config logs+falls-back; planner respawns on ResilienceConfig change
(incl null→json) and stays delta on a pure config change; artifact parse carries/omits the column.
Core.Tests 243 (+5), Runtime.Tests 363 (+5), Host builds clean.
Akka.Cluster.Hosting's WithClustering enables an SBR downing provider by default
(applies SplitBrainResolverOption.Default when ClusterOptions.SplitBrainResolver is
null), which reads the pre-existing akka.conf keep-oldest block. So the cluster was
NOT running NoDowning before Critical 1 and hard-crash failover already worked — the
typed KeepOldestOption is reinforcing/explicit-in-code, not the sole activator.
Corrects the inaccurate 'HOCON inert / NoDowning / never fails over' framing in:
- ServiceCollectionExtensions.BuildClusterOptions XML comment
- akka.conf split-brain-resolver comment
- docs/Redundancy.md Split-brain section
- SplitBrainResolverActivationTests summary + assertion message (+ method rename)
No code revert (the typed option is correct belt-and-suspenders). Cluster.Tests 29/29.
Surfaced by the #9 hard-kill failover negative control.
Updates FOLLOWUP-10 + STATUS.md now that #10 is wired:
- The filed "thread CapabilityInvoker into the actor" was infeasible — Runtime is
deliberately Polly-free (refs Core.Abstractions, not Core). Shipped a seam mirroring
IDriverFactory (branch fix/archreview-crit10-wire-capability-invoker @ 62556c24).
- Triage corrections: only 6 of 7 sites are real (GenericDriverNodeManager is test-only
scaffolding, re-annotated); tier-defaults only (ResilienceConfig not in the artifact).
- Observability finding: the resilience-status tracker has no reader (Stream E.2/E.3
never shipped) → added retry/breaker logging as the interim surface + live-verify enabler.
- Verification: negative-control, recording-invoker routing, real-invoker actor-context,
analyzer + logging tests; clean solution build; Runtime 358 / Core 238 / Analyzers 32.
- Residual: live behavioral gate (rig recipe recorded) + the ResilienceConfig-in-artifact
sub-gap follow-up.
The pass-through NullDriverCapabilityInvoker returns the call-site task directly, so it
never exercises the real invoker's internal `pipeline.ExecuteAsync(...).ConfigureAwait(false)`.
Adds a test-only Core reference to Runtime.Tests (the PRODUCTION Runtime assembly stays
Polly-free via the seam) and a discovery test that drives DriverInstanceActor through a real
CapabilityInvoker over a genuinely-yielding ITagDiscovery driver — proving the invoker's
internal ConfigureAwait(false) does NOT leak to the actor's own await, so the post-await
Context.Parent.Tell still runs (DiscoveredNodesReady arrives). Runtime.Tests 358, Core.Tests 238.
Remediates RESILIENCE-DISPATCH-GAP (surfaced by the 07/C-1 analyzer): the Phase 6.1
CapabilityInvoker resilience pipeline (retry / breaker / bulkhead / telemetry) was
constructed only in tests — the production dispatch layer called driver-capability
methods directly, bypassing it entirely.
The filed plan (thread CapabilityInvoker directly into DriverInstanceActor) is
INFEASIBLE: Runtime is deliberately Polly-free (references Core.Abstractions, not
Core — the same boundary IDriverFactory documents). So this introduces a seam,
mirroring IDriverFactory exactly:
- Core.Abstractions: IDriverCapabilityInvoker (+ NullDriverCapabilityInvoker
pass-through) and IDriverCapabilityInvokerFactory (+ null factory).
- Core: CapabilityInvoker now implements the interface; new
DriverCapabilityInvokerFactory builds a per-instance invoker over the
process-singleton pipeline builder + status tracker + tier resolver.
- Runtime: DriverInstanceActor takes an IDriverCapabilityInvoker (default =
pass-through) and routes all 6 dispatch sites (write, alarm-ack, subscribe,
unsubscribe, alarm-subscribe, discover) through it; per-host key resolved via
IPerCallHostResolver for single-ref calls, driver-instance key for bulk calls.
DriverHostActor builds + injects the real invoker per spawned driver.
- Host: DriverFactoryBootstrap registers the tracker, pipeline builder, and
concrete factory (needs DriverFactoryRegistry.GetTier). Runtime SCE resolves the
factory from DI like IDriverFactory; pass-through on nodes without it bound.
- Analyzer: IDriverCapabilityInvoker.ExecuteAsync/ExecuteWriteAsync are now
recognized wrapper homes (interface-typed invoker calls must suppress OTOPCUA0001
too). All 6 RESILIENCE-DISPATCH-GAP pragmas removed — the analyzer (an error in
Runtime) is now the standing regression guard.
Also adds retry / breaker-open / breaker-close LOGGING to the pipeline builder —
the operator-facing observability surface (the Admin /hosts reader, Phase 6.1
Stream E.2/E.3, was never built, so the pipeline otherwise runs silently).
Scope notes:
- Tier-DEFAULT policy only: DriverInstanceSpec (the deploy artifact) does not carry
the per-instance ResilienceConfig JSON, so overrides aren't applied yet (tracked
follow-up: plumb ResilienceConfig through the composer/artifact).
- GenericDriverNodeManager's call is NOT wired: that class is test-only scaffolding
(zero production references — the Server has its own address-space path); its
pragma is re-annotated accordingly, not left as a "tracked gap".
Verification (unit + analyzer; live behavioral gate still pending — see FOLLOWUP-10):
- Negative control: unwrapping any site fails the Runtime build with OTOPCUA0001.
- New Runtime guard (recording invoker): write routes via ExecuteWriteAsync with the
IPerCallHostResolver host; subscribe via ExecuteAsync — proves runtime routing.
- New analyzer test: the interface is a valid wrapper home.
- New pipeline-builder test: retry events are logged.
- Full solution builds clean (0 errors); Runtime.Tests 357, Core.Tests 238,
Analyzers.Tests 32 all green; pass-through default keeps existing dispatch tests
byte-for-byte unchanged.
The load-bearing integration half of Critical 1. Adds TwoNodeClusterHarness.HardKillNodeAAsync
(canonical in-process crash sim: shut down node A's remoting transport via
IRemoteActorRefProvider.Transport.Shutdown() — associations drop, heartbeats/gossip stop, NO
graceful Cluster.Leave, so node B's failure detector marks A Unreachable) + a
WaitForNodeBSoleDriverLeaderAsync helper (waits for a STABLE takeover: B sole Up member + driver
role-leader). HardKillFailoverTests crashes the oldest node and asserts the survivor takes over —
the exact scenario the graceful StopNodeBAsync path (Cluster.Leave, no downing decision) cannot
exercise.
Live-verified in-process (~35s: acceptable-heartbeat-pause 10s + stable-after 15s + convergence).
Negative control confirmed the test BITES: forcing explicit NoDowning
(akka.cluster.downing-provider-class = "") makes it time out with the node stuck Up-but-Unreachable.
FINDING surfaced by the negative control: removing Critical 1's typed
ClusterOptions.SplitBrainResolver did NOT break failover, because Akka.Cluster.Hosting's
WithClustering applies SplitBrainResolverOption.Default when the option is null — enabling the SBR
downing provider that reads the akka.conf keep-oldest block (present on master BEFORE Critical 1).
So the cluster was NOT running NoDowning before Critical 1; the typed option reinforces the akka.conf
strategy rather than being the sole activator, and Critical 1's 'HOCON inert / NoDowning' premise is
inaccurate. The test guards the failover OUTCOME regardless of activation path. Comments corrected to
reflect this; flagged for review.
S-1: replace the hand-maintained 5-project unit-tests matrix (which silently dropped
Client's 388 tests, Analyzers' 31, and every driver + most Core suite) with ONE
whole-solution leg — dotnet test ZB.MOM.WW.OtOpcUa.slnx --filter
"Category!=E2E&Category!=LiveIntegration". Self-maintaining: a new *.Tests project is
covered automatically, matching CLAUDE.md's own guidance.
S-2: emit trx + add scripts/ci/assert-not-all-skipped.sh, a fail-on-skip gate that
turns 'green CI == everything skipped' into a red build. Wired strict (MIN_EXECUTED=1)
on the unit leg; report-only (MIN_EXECUTED=0) on the fixtureless integration leg so
its skip tally is VISIBLE without a false red — with a documented follow-up to start
the one public-image fixture (opc-plc) as a service and flip it strict.
S-4 (paired): the newly-CI'd Client.CLI.Tests had fixed-sleep startup races
(await Task.Delay(100/150) before cancelling a background command) that would flake
under CI load. Added SubscribeInvoked / SubscribeAlarmsInvoked readiness signals
(TaskCompletionSource) to FakeOpcUaClientService and replaced the 11 sleeps across
AlarmsCommandTests / SubscribeCommandTests / EventHandlerLifecycleTests with a
deterministic await-the-signal (10s timeout guard).
Verified: workflow YAML parses; skip-gate proven locally on a real trx (executed=31
=> OK), a synthetic all-skipped trx (executed=0 => exit 1 with diagnostic),
report-only mode (never fails), multi-file sum, and missing-file (exit 2);
Client.CLI.Tests 104/104 green after the deflake. (CI job execution itself is
verifiable only on push — nothing pushed.)
Inject the custom UnwrappedCapabilityCallAnalyzer as an OutputItemType=Analyzer
ProjectReference from Directory.Build.props (excluding the analyzer + its test
project) so OTOPCUA0001 runs on every src/ and tests/ compilation — it previously
enforced its CapabilityInvoker-wrapping rule against nothing but its own 31 unit
tests (the 'built-but-never-wired' failure mode).
Triage of the ~280 surfaced hits, three categories:
1. RESILIENCE-DISPATCH-GAP (7 sites, DriverInstanceActor x6 + GenericDriverNodeManager
x1): a REAL, previously-untracked gap the analyzer caught on first wiring — the
Phase 6.1 CapabilityInvoker resilience pipeline (retry/breaker/bulkhead/telemetry)
is constructed ONLY in tests and was never wired into the production dispatch
layer. Scoped per-site #pragma with a greppable RESILIENCE-DISPATCH-GAP marker
explicitly noting these are tracked-but-not-intentional, pending the dispatch-wiring
remediation (filed as a follow-up). Keeps the analyzer live everywhere else in
those projects so a NEW unwrapped call still fails the build.
2. Driver-INTERNAL self-calls (3 sites, AbCipAlarmProjection x2 + S7Driver x1):
a driver's own poll/ack path calling its own capability method. The invoker wraps
the driver from the dispatch layer OUTWARD; a driver re-wrapping its own internal
calls would double-wrap. Genuinely intentional — scoped #pragma with that rationale.
3. Wire-level test suites + manual-testing CLIs (12 projects): invoke drivers directly
by design — the analyzer's own documented intentional case. Project-level NoWarn
with a comment.
Verified: full solution build green, 0 OTOPCUA0001 hits; analyzer's 31 tests pass;
negative control — dropping one dispatch-gap pragma re-fires OTOPCUA0001 and fails
the Runtime build, proving the analyzer is genuinely live tree-wide, not disabled.
Add a mechanical, reflection-driven guard that every method of every forwarding
interface DeferredAddressSpaceSink implements (IOpcUaAddressSpaceSink +
ISurgicalAddressSpaceSink) actually reaches the inner sink, plus the sibling
DeferredServiceLevelPublisher. A DispatchProxy recording inner auto-implements
future interface members with zero maintenance, so a newly-added sink capability
that the wrapper forgets to forward now fails a test instead of shipping inert on
every driver-role host (the F10b / PR#423 failure mode). A guard-of-the-guard
asserts the wrapper implements exactly the known forwarding-interface set, so a
brand-new capability *interface* trips the test too.
This is the prerequisite gate for 03/P1 (surgical remove methods) and 02/P
(batched WriteValues): both add sink capability members that must not ship inert.
Verified: 3 new tests green (Commons.Tests 57/57); negative control (stubbing
UpdateFolderDisplayName to a non-forwarding no-op) fails with the exact
member-named diagnostic, proving the guard bites.
Native ADS notifications (the default subscribe mode) were stored as opaque handles
with no record of the symbol/type/interval/handler needed to replay. On a client swap
(EnsureConnectedAsync building a fresh client after a drop) the notifications were
silently orphaned — no Bad status, no error, pushes just stopped until redeploy.
Compounding: the IsConnected fast-path keys on AMS-port state, not wire liveness, and
a probe failure only transitioned state without recycling the dead client.
Fix:
- Store REPLAYABLE INTENT: NativeRegistration (symbol/type/bit/interval/onChange +
swappable live handle) hung off DeviceState.NativeRegistrations, populated by
SubscribeAsync via RegisterNotificationAsync (under ConnectGate).
- Split EnsureConnectedAsync into a gate wrapper + EnsureConnectedUnderGateAsync core;
when the core installs a NEW client it replays every stored intent onto it and swaps
the live handle (disposing the dead one). Register + replay both run under ConnectGate
so they can't race.
- Probe loop: on a wire-probe failure (false or throw) RecycleClientAsync disposes+nulls
the client so the next tick rebuilds + replays — closes the fast-path-keys-on-port-state
compounding bug.
No TwinCAT docker fixture exists (integration needs a real TC3 XAR), so the fake-client
unit tests are the authoritative coverage:
- 4 new guards in TwinCATReconnectReplayTests (replay-onto-fresh-client + push reaches
OnDataChange + old handle disposed; replay-all-tags; unsubscribe-after-reconnect stops
replaying; probe-failure recycles+rebuilds).
- Full TwinCAT unit suite 174/174 green; full solution builds 0 errors.
The S7 Plc was opened once in InitializeAsync and never re-opened: a transient
PLC reboot / network blip permanently killed the driver until redeploy. Introduce
an IS7Plc / IS7PlcFactory seam (mirrors TwinCAT's ITwinCATClientFactory) and a lazy
EnsureConnectedAsync that disposes a dead handle and re-opens a fresh connection on
the next data call. Reads/writes mark the handle dead on a connection-fatal fault
(socket drop / ErrorCode.ConnectionError) but NOT on a data-address error; the probe
loop routes through EnsureConnectedAsync as a backstop.
The seam also closes the S7 TEST-1 gap — the reconnect state machine is now unit-
testable without a live PLC (S7.Net.Plc is sealed with no in-process fake).
Verification:
- 4 deterministic unit guards (fatal→reopen, data-error→no-reopen, raw socket→reopen,
reopen-fail→degrade-then-recover); full S7 unit suite 234/234 green.
- LIVE end-to-end proof against real python-snap7 (S7_1500ReconnectTests, double-gated
on sim reachability + S7_RECONNECT_BOUNCE_CMD): bounced the container, driver observed
the outage and resumed Good reads with NO redeploy. Baseline smoke 3/3 still green.
Joseph Doherty