scadalink-design/code-reviews/ExternalSystemGateway/findings.md

# Code Review — ExternalSystemGateway

| Field | Value |
|-------|-------|
| Module | `src/ScadaLink.ExternalSystemGateway` |
| Design doc | `docs/requirements/Component-ExternalSystemGateway.md` |
| Status | Reviewed |
| Last reviewed | 2026-05-17 |
| Reviewer | claude-agent |
| Commit reviewed | `39d737e` |
| Open findings | 0 |

## Summary

The External System Gateway is a small module (five source files plus options) that
implements the HTTP/REST client (`ExternalSystemClient`), the database access surface
(`DatabaseGateway`), and error classification (`ErrorClassifier`). The structure is
clean and the dual call-mode semantics broadly match the design doc. However, the
review surfaced several substantive problems that prevent the module from behaving as
designed. The most serious is that **no store-and-forward delivery handler is ever
registered** for the `ExternalSystem` or `CachedDbWrite` categories, so cached calls
and cached writes are buffered but can never actually be delivered on retry — a silent
data-loss path. Two further high-impact issues are that the **per-system call timeout
is never applied** to the HTTP client (the design's central error-handling guarantee
is absent), and that **`CachedCall` double-dispatches the HTTP request** because
`StoreAndForwardService.EnqueueAsync` itself re-attempts immediate delivery, breaking
the idempotency expectations. A cluster of medium issues concern resource leaks,
classification gaps (cancellation conflation), and the dropped `StoreAndForwardResult`.
Test coverage is thin — `CachedCall` transient/buffering paths and `DatabaseGateway`
are entirely untested. Themes: incomplete wiring against the S&F engine, and design-doc
requirements (timeout, retry settings) that are declared but not implemented.

#### Re-review 2026-05-17 (commit `39d737e`)

All fourteen prior findings remain `Resolved`; the resolutions for findings 001–014
were spot-checked against the current source and hold. The re-review walked the full
10-category checklist again and surfaced **three new findings**. The most serious
(`ExternalSystemGateway-015`, High) is a regression *introduced by* the
`ExternalSystemGateway-004` resolution: `CachedCall`/`CachedWrite` now pass a
per-system/per-connection `MaxRetries` of `0` through verbatim, but
`StoreAndForwardService.RetryMessageAsync` interprets a stored `MaxRetries` of `0` as
**retry forever**, not "never retry" — so the very `0` the ESG-004 fix claims to
"honour as never retry" actually produces an unbounded retry loop, and two ESG tests
assert the broken behaviour. `ExternalSystemGateway-016` (Medium) is that the
`ExternalSystemGateway-013` resolution used `ConfigureHttpClientDefaults`, which is a
**process-global** registration — it forces a `SocketsHttpHandler` (capped at the ESG
option) onto every `HttpClient` in the host, including the Notification Service's
OAuth2 token client, not just the gateway's per-system clients. `ExternalSystemGateway-017`
(Low) is a trailing-`?` URL nit when a GET method's parameters are all null. Theme:
both substantive findings are second-order defects in earlier fixes — the earlier
resolutions did not verify the downstream contract of the S&F engine they integrate
with.

## Checklist coverage

| # | Category | Examined | Notes |
|---|----------|----------|-------|
| 1 | Correctness & logic bugs | ☑ | Prior: URL edge cases, dropped S&F result, classification — 003, 006, 009. Re-review: `MaxRetries == 0` retry-forever vs never-retry contradiction (015); trailing-`?` URL nit (017). |
| 2 | Akka.NET conventions | ☑ | No actors in this module; `AddExternalSystemGatewayActors` is a no-op. Blocking-I/O isolation is delegated to Site Runtime. No issues found. |
| 3 | Concurrency & thread safety | ☑ | Services are stateless and DI-scoped; the S&F delivery handlers resolve in a fresh DI scope on the sweep thread. No findings raised. |
| 4 | Error handling & resilience | ☑ | Prior: handler registration, double-dispatch, timeout, cancellation — 001, 002, 003, 008. Re-review: the unbounded-retry consequence of finding 015 is also a resilience defect (recorded under category 1). |
| 5 | Security | ☑ | Error bodies now truncated (007). No new issues — auth secrets not logged, body capped. |
| 6 | Performance & resource management | ☑ | Prior: disposal and repository-scan findings 005, 010, 011 — all resolved and verified. No new issues. |
| 7 | Design-document adherence | ☑ | Prior: timeout, retry settings, logging — 002, 004, 012. Re-review: finding 015 is also a design-adherence gap (S&F retry contract); recorded under category 1. |
| 8 | Code organization & conventions | ☑ | Prior: `MaxConcurrentConnectionsPerSystem` wiring — 013. Re-review: that wiring uses process-global `ConfigureHttpClientDefaults`, leaking the ESG cap onto every host `HttpClient` — finding 016. |
| 9 | Testing coverage | ☑ | Coverage is broad after finding 014. Re-review note: the `ZeroMaxRetries...` tests assert the persisted column, not the sweep outcome, and so lock in the finding-015 defect. |
| 10 | Documentation & comments | ☑ | Inline comments at `ExternalSystemClient.cs:118-119` / `DatabaseGateway.cs:99-101` assert a "never retry" semantic that the code does not deliver — see finding 015. |

## Findings

### ExternalSystemGateway-001 — No S&F delivery handler registered; cached calls and writes can never be delivered

| | |
|--|--|
| Severity | Critical |
| Category | Error handling & resilience |
| Status | Resolved |
| Location | `src/ScadaLink.ExternalSystemGateway/ExternalSystemClient.cs:109`, `src/ScadaLink.ExternalSystemGateway/DatabaseGateway.cs:81` |

**Description**

`CachedCallAsync` and `CachedWriteAsync` enqueue messages under
`StoreAndForwardCategory.ExternalSystem` and `StoreAndForwardCategory.CachedDbWrite`.
`StoreAndForwardService.RegisterDeliveryHandler` is the only mechanism that lets the
S&F engine actually deliver a buffered message, and a repository-wide search shows it
is **never called for either category** anywhere in the codebase. Consequences:

1. On a transient failure, `EnqueueAsync` falls through to the "No handler registered
   — buffer for later" branch (`StoreAndForwardService.cs:163`) and the message is
   persisted.
2. During the retry sweep, `AttemptDeliveryAsync` (`StoreAndForwardService.cs:201`)
   logs `"No delivery handler for category {Category}"` and returns without ever
   removing or delivering the message.

The result is that every cached external call and cached DB write is silently
buffered forever and never delivered — a data-loss path for the exact "deferred
delivery is acceptable" use cases the design doc calls out (posting production data,
quality reports). The script also receives `WasBuffered: true` / a successful
`CachedWriteAsync` completion, so the failure is completely invisible.

**Recommendation**

Register delivery handlers for `StoreAndForwardCategory.ExternalSystem` and
`StoreAndForwardCategory.CachedDbWrite` during host/site startup. The `ExternalSystem`
handler should deserialize the payload, re-resolve the system/method, and re-invoke
`InvokeHttpAsync`, returning `true`/`false`/throwing per the transient-vs-permanent
contract `EnqueueAsync` expects. The `CachedDbWrite` handler should execute the SQL
against the named connection. Add an integration test that buffers a message and
verifies it is delivered by a retry sweep.

**Resolution**

Resolved 2026-05-16. Delivery handlers for `StoreAndForwardCategory.ExternalSystem` and
`CachedDbWrite` are now registered at site startup in `AkkaHostedService`, after
`StoreAndForwardService.StartAsync()`. Each handler resolves its consumer in a fresh DI
scope and calls a new `DeliverBufferedAsync`: `ExternalSystemClient.DeliverBufferedAsync`
re-resolves the system/method and re-invokes `InvokeHttpAsync`, and
`DatabaseGateway.DeliverBufferedAsync` executes the buffered SQL — each returning `true`
on success, `false` (park) when the target no longer exists or fails permanently, and
throwing on transient failure so the engine retries. `EnqueueAsync` gained an
`attemptImmediateDelivery` parameter; `CachedCallAsync` passes `false` so registering the
handler does not dispatch the request twice (the double-dispatch noted in
`ExternalSystemGateway-003`). Regression tests cover the success, target-removed and
transient-retry paths. Fixed by the commit whose message references
`ExternalSystemGateway-001`.

### ExternalSystemGateway-002 — Per-system call timeout is never applied to HTTP requests

| | |
|--|--|
| Severity | High |
| Category | Error handling & resilience |
| Status | Resolved |
| Location | `src/ScadaLink.ExternalSystemGateway/ExternalSystemClient.cs:130`, `src/ScadaLink.ExternalSystemGateway/ServiceCollectionExtensions.cs:13` |

**Description**

The design doc states each external system definition specifies a timeout that
"applies to all method calls on that system" and "applies to the HTTP request
round-trip", and `ExternalSystemGatewayOptions.DefaultHttpTimeout` exists as a
fallback. In practice no timeout is ever configured. `ServiceCollectionExtensions`
calls `services.AddHttpClient()` with no per-named-client configuration, and
`InvokeHttpAsync` calls `_httpClientFactory.CreateClient($"ExternalSystem_{system.Name}")`
without setting `client.Timeout` or passing a `CancellationToken` derived from a
timeout. `SendAsync` is therefore subject only to `HttpClient`'s default 100-second
timeout, regardless of the system definition or the configured `DefaultHttpTimeout`.
A slow or hung external system will block the calling Script Execution Actor far
longer than the operator configured, and the design's core error-handling guarantee
(timeout → transient classification) does not hold within the intended window.

There is also no `Timeout` field on `ExternalSystemDefinition` at all, so even a
correct implementation has nowhere to read the per-system value from — the entity is
missing the field the design requires.

**Recommendation**

Add a `Timeout` (TimeSpan) field to `ExternalSystemDefinition` and have
`InvokeHttpAsync` enforce it — either by setting `client.Timeout` via a typed/named
`HttpClient` registration, or by linking a `CancellationTokenSource` with the
per-system (or `DefaultHttpTimeout`) timeout to the supplied `cancellationToken`
before `SendAsync`. Ensure the resulting `TaskCanceledException`/`TimeoutException`
is classified as transient.

**Resolution**

Resolved 2026-05-16 (commit `<pending>`). `InvokeHttpAsync` now enforces a call
timeout: `ExternalSystemClient` takes an `IOptions<ExternalSystemGatewayOptions>` and
links a `CancellationTokenSource(DefaultHttpTimeout)` with the caller's token before
`SendAsync` and the response-body read, so the design's "timeout applies to the HTTP
request round-trip" guarantee now holds within the configured window (default 30s)
instead of `HttpClient`'s default 100s. A timeout is reclassified as a
`TransientExternalSystemException`; a caller-initiated cancellation is distinguished
from a timeout and propagated as `OperationCanceledException` rather than being
swallowed as transient. Regression tests:
`Call_SlowSystem_TimesOutAsTransientErrorWithinConfiguredWindow` and
`Call_CallerCancellation_IsNotMisreportedAsTimeout`.

Note (partial scope): the per-*system* `Timeout` field on `ExternalSystemDefinition`
remains unimplemented — adding it requires a change to `ScadaLink.Commons`, which is
outside this module's edit scope. Until that entity field exists, the configured
`DefaultHttpTimeout` is the effective per-call limit for every system. A follow-up
against the Commons module should add the `Timeout` field and have `InvokeHttpAsync`
prefer it over the default. This is a tracked follow-up, not a regression.

### ExternalSystemGateway-003 — `CachedCall` double-dispatches the HTTP request

| | |
|--|--|
| Severity | High |
| Category | Correctness & logic bugs |
| Status | Resolved |
| Location | `src/ScadaLink.ExternalSystemGateway/ExternalSystemClient.cs:84-117` |

**Description**

`CachedCallAsync` first calls `InvokeHttpAsync` directly (line 86). On a
`TransientExternalSystemException` it then calls `_storeAndForward.EnqueueAsync(...)`
(line 109). `StoreAndForwardService.EnqueueAsync` is **not** a pure enqueue — it
"Attempts immediate delivery" by invoking the registered delivery handler
(`StoreAndForwardService.cs:128-159`). If a delivery handler for the `ExternalSystem`
category is registered (as finding 001 recommends), the HTTP request will be executed
a **second time** synchronously inside `EnqueueAsync`, immediately after the first
attempt failed. For a transient failure that is actually a slow/overloaded system,
this doubles the load and — critically — if the original request did reach the
external system, the immediate retry produces a duplicate delivery before the script
even returns, worsening the idempotency hazard the design doc explicitly warns about.

**Recommendation**

Decide on one dispatch path. Either (a) have `CachedCall` not pre-invoke
`InvokeHttpAsync` and instead let `EnqueueAsync`'s immediate-delivery attempt be the
single first attempt (requires the handler to exist and to surface permanent vs
transient correctly); or (b) add an enqueue-only entry point to
`StoreAndForwardService` that skips the immediate-delivery attempt, and have
`CachedCall` use it after its own first attempt. Approach (a) is cleaner and removes
the duplicated logic.

**Resolution**

Resolved 2026-05-16 (commit `<pending>`). Re-triage: this finding was already fixed in
the codebase as a side effect of the `ExternalSystemGateway-001` fix and is no longer
reproducible against the current source. `StoreAndForwardService.EnqueueAsync` gained an
`attemptImmediateDelivery` parameter (recommendation approach (b)), and
`CachedCallAsync` passes `attemptImmediateDelivery: false` after its own first HTTP
attempt — so `EnqueueAsync` buffers the message for the background retry sweep without
re-invoking the registered delivery handler, eliminating the duplicate dispatch. A
dedicated regression test, `CachedCall_TransientFailure_DoesNotImmediatelyRedispatchViaRegisteredHandler`,
was added in this module's test suite: it registers a counting delivery handler, drives
a `CachedCall` whose HTTP attempt fails transiently, and asserts the handler is invoked
zero times during enqueue. The test was verified to fail if `attemptImmediateDelivery`
is flipped back to `true`.

### ExternalSystemGateway-004 — System retry settings are not honoured for cached calls/writes

| | |
|--|--|
| Severity | Medium |
| Category | Design-document adherence |
| Status | Resolved |
| Location | `src/ScadaLink.ExternalSystemGateway/ExternalSystemClient.cs:114-115`, `src/ScadaLink.ExternalSystemGateway/DatabaseGateway.cs:86-87` |

**Description**

`CachedCallAsync` and `CachedWriteAsync` pass the definition's `MaxRetries` /
`RetryDelay` to `EnqueueAsync` only when they are non-default
(`MaxRetries > 0 ? ... : null`, `RetryDelay > TimeSpan.Zero ? ... : null`), otherwise
falling back to the S&F defaults. The site-side repository that supplies these
definitions, `SiteExternalSystemRepository.MapExternalSystem`
(`src/ScadaLink.SiteRuntime/Repositories/SiteExternalSystemRepository.cs:194`), never
reads `MaxRetries`/`RetryDelay` from SQLite at all — the constructed entities always
have `MaxRetries == 0` and `RetryDelay == TimeSpan.Zero`. As a result, at sites the
per-system retry settings the design doc requires are *always* discarded and the
global S&F defaults are silently used instead. The `> 0` guard in the ESG also makes
a legitimately-configured `MaxRetries` of 0 ("never retry") indistinguishable from
"unset", so an operator cannot express "do not retry".

**Recommendation**

Within this module, drop the `> 0` / `> Zero` guards and pass the definition values
through directly (or use nullable fields on the entity to distinguish "unset"). The
companion fix in `SiteExternalSystemRepository` to actually map the retry columns
should be tracked against the SiteRuntime module.

**Resolution**

Resolved 2026-05-16 (commit pending). `CachedCallAsync` and `CachedWriteAsync` now pass
the definition's `MaxRetries` to `EnqueueAsync` verbatim — the `> 0` guard is dropped, so
a legitimately-configured `MaxRetries` of 0 ("never retry") is honoured instead of being
collapsed to the S&F default. The `RetryDelay > TimeSpan.Zero` guard is deliberately
**kept**: `TimeSpan.Zero` is the entity default for an unconfigured field and a literal
zero-delay retry loop is not a valid configuration, so falling back to the S&F default
interval for an unset delay is correct (only `MaxRetries == 0` is a meaningful operator
choice). Regression test `CachedCall_TransientFailure_ZeroMaxRetriesIsHonouredNotTreatedAsUnset`
buffers a transient failure and asserts the buffered message carries `MaxRetries == 0`
rather than the S&F default; `CachedCall_TransientFailure_BuffersWithSystemRetrySettings`
additionally covers a non-default settings pass-through. The companion fix in
`SiteExternalSystemRepository.MapExternalSystem` to actually read the `MaxRetries` /
`RetryDelay` columns from SQLite remains a tracked follow-up against the SiteRuntime
module (outside this module's edit scope) — until then, sites still supply
`MaxRetries == 0`, which this fix now correctly honours as "never retry".

### ExternalSystemGateway-005 — `HttpRequestMessage` and `HttpResponseMessage` are not disposed

| | |
|--|--|
| Severity | Medium |
| Category | Performance & resource management |
| Status | Resolved |
| Location | `src/ScadaLink.ExternalSystemGateway/ExternalSystemClient.cs:133-167` |

**Description**

`InvokeHttpAsync` creates an `HttpRequestMessage` (line 133) and receives an
`HttpResponseMessage` from `SendAsync` (line 155); neither is wrapped in a `using` nor
explicitly disposed. Both are `IDisposable` and own resources (the request's
`StringContent`, the response's content stream). Under the per-invocation call volume
of a busy site this produces avoidable pressure on the finalizer queue and can hold
socket/stream resources longer than necessary. The success path reads the content but
never disposes the response; the error path likewise reads `errorBody` and then throws
without disposing.

**Recommendation**

Wrap the request in `using var request = ...` and the response in
`using var response = ...` (or call `Dispose()` in a `finally`). Ensure disposal still
occurs on the exception paths.

**Resolution**

Resolved 2026-05-16 (commit pending). `InvokeHttpAsync` now declares the request as
`using var request` and wraps all response handling in a `using (response)` block, so
both `IDisposable` instances (and the request's `StringContent` / the response content
stream) are released on the success path **and** on the permanent/transient
exception paths. Regression tests `Call_SuccessfulHttp_DisposesRequestAndResponse` and
`Call_PermanentFailure_StillDisposesRequestAndResponse` use a disposal-tracking
`HttpMessageHandler`/`HttpContent` and assert both the request and the response content
are disposed; both were verified to fail before the `using` wrappers were added.

### ExternalSystemGateway-006 — `BuildUrl` ignores path templates and appends a trailing slash for empty paths

| | |
|--|--|
| Severity | Medium — partially re-triaged: trailing-slash bug fixed; path-templating sub-issue is a design decision (see Resolution) |
| Category | Correctness & logic bugs |
| Status | Resolved |
| Location | `src/ScadaLink.ExternalSystemGateway/ExternalSystemClient.cs:180-196` |

**Description**

`BuildUrl` does `baseUrl.TrimEnd('/') + "/" + path.TrimStart('/')`. When `method.Path`
is empty (a method that targets the base URL itself), this still appends a `/`,
producing `https://host/api/` which some servers treat as a different resource than
`https://host/api`. More importantly, the design doc shows method paths as templates
like `/recipes/{id}`, but `BuildUrl` performs no placeholder substitution — a `{id}`
token is sent literally in the URL and the corresponding parameter is instead appended
as a query-string entry (for GET/DELETE) or placed in the JSON body (POST/PUT). Either
the design's path-template feature is unimplemented, or the doc is stale; in the
current code a method defined as `/recipes/{id}` will never produce a correct URL.

**Recommendation**

Decide whether path templating is in scope. If yes, implement `{name}` substitution
from `parameters` in `BuildUrl` and exclude substituted parameters from the query
string/body. If no, update the component design doc to remove the `/recipes/{id}`
example and state that paths are literal. Also avoid appending a trailing `/` when
`path` is empty.

**Resolution**

Resolved 2026-05-16 (commit pending). The **trailing-slash bug** is fixed: `BuildUrl`
now appends a `/`-joined path segment only when the method's path is non-empty after
trimming, so a method targeting the base URL itself produces `https://host/api` rather
than `https://host/api/`. Regression tests `Call_MethodWithEmptyPath_DoesNotAppendTrailingSlash`
and `Call_MethodWithPath_BuildsExpectedUrl` (asserting on the captured request URI)
cover the empty-path and normal-path cases; the empty-path test was verified to fail
before the fix.

Re-triage of the **path-templating sub-issue** (`{id}` placeholder substitution): this
is a genuine design decision, not a code bug, and it requires editing the component
design doc — both outside this module's edit scope (`src/`, `tests/`, this file only).
The current code treats method paths as literal strings and routes parameters to the
query string (GET/DELETE) or JSON body (POST/PUT); a method authored as `/recipes/{id}`
sends the `{id}` token verbatim. **Tracked follow-up / surfaced design question:** the
design owner must decide whether path templating is in scope — if yes, implement
`{name}` substitution in `BuildUrl` and exclude substituted params from the
query/body; if no, the `Component-ExternalSystemGateway.md` `/recipes/{id}` example must
be changed to a literal path. The trailing-slash defect (the concrete correctness bug
in this finding) is fully resolved.

### ExternalSystemGateway-007 — External error response bodies are echoed verbatim into script-visible error messages

| | |
|--|--|
| Severity | Medium |
| Category | Security |
| Status | Resolved |
| Location | `src/ScadaLink.ExternalSystemGateway/ExternalSystemClient.cs:167-177` |

**Description**

On a non-success HTTP response, the full response body is read into `errorBody` and
embedded verbatim into the exception message (`$"HTTP {code} from {name}: {errorBody}"`),
which then flows into `ExternalCallResult.ErrorMessage` and back to the calling script,
and into Site Event Logging. An external system error page can be arbitrarily large
(an HTML stack trace, a multi-megabyte body) and may contain sensitive detail. There
is no size cap, so a hostile or misbehaving endpoint can inflate every error log entry
and error string returned to scripts. There is also no content-type check before
treating the body as text.

**Recommendation**

Truncate `errorBody` to a bounded length (e.g. 1–2 KB) before embedding it, and
consider logging the full body separately at debug level rather than returning it to
the script. Optionally only include the body when the content type is textual.

**Resolution**

Resolved 2026-05-16 (commit pending). `InvokeHttpAsync` now truncates the external error
response body to `MaxErrorBodyChars` (2048) via a `Truncate` helper before embedding it
into the transient/permanent exception message — so a misbehaving or hostile endpoint
can no longer inflate every script-visible `ErrorMessage` and Site Event Logging entry
with a multi-megabyte body. When truncation occurs the message is suffixed with
`… [truncated, N chars total]` so the original size is still visible. Regression test
`Call_PermanentFailureWithHugeErrorBody_TruncatesErrorMessage` drives a 400 with a
500 KB body and asserts the resulting `ErrorMessage` is bounded (< 4096 chars); it was
verified to fail (500 040-char message) before the cap was added. Content-type
filtering was considered optional in the recommendation and was not implemented — the
size cap alone closes the inflation/disclosure vector.

### ExternalSystemGateway-008 — Cancellation is conflated with transient timeout failure

| | |
|--|--|
| Severity | Medium — re-triaged: root cause already fixed in current source (see Resolution) |
| Category | Error handling & resilience |
| Status | Resolved |
| Location | `src/ScadaLink.ExternalSystemGateway/ErrorClassifier.cs:24-30`, `src/ScadaLink.ExternalSystemGateway/ExternalSystemClient.cs:157-159` |

**Description**

`ErrorClassifier.IsTransient(Exception)` returns `true` for `TaskCanceledException`
and `OperationCanceledException`. `HttpClient.SendAsync` throws `TaskCanceledException`
both when its internal timeout elapses *and* when the supplied `CancellationToken` is
cancelled (e.g. the Script Execution Actor is stopped, or the actor system is shutting
down). Because `InvokeHttpAsync`'s `catch` filter treats all of these as transient, a
caller-initiated cancellation during a `CachedCall` will be misclassified as a
transient failure and the message will be buffered for retry — work the caller
explicitly asked to abandon. For a `Call`, a shutdown-time cancellation is reported to
the script as a "Transient error" rather than an `OperationCanceledException`.

**Recommendation**

In `InvokeHttpAsync`, check `cancellationToken.IsCancellationRequested` first and
rethrow `OperationCanceledException` (or let it propagate) before applying transient
classification. Only treat a cancellation as a timeout when the supplied token is
*not* the one that was cancelled.

**Resolution**

Resolved 2026-05-16 (commit pending). **Re-triage:** the root cause described — a
caller-initiated cancellation being misclassified as a transient failure — is **no
longer present in the current source** and is not reproducible. `InvokeHttpAsync`
already wraps both `SendAsync` and the response-body `ReadAsStringAsync` in ordered
`catch (OperationCanceledException) when (cancellationToken.IsCancellationRequested)`
filters that rethrow the caller's cancellation *before* the
`catch (Exception ex) when (ErrorClassifier.IsTransient(ex))` branch is ever reached
(this was added alongside the `ExternalSystemGateway-002` timeout fix). A caller-cancel
therefore propagates as `OperationCanceledException` and is never buffered; only the
gateway's own timeout token reclassifies as transient.

`ErrorClassifier.IsTransient(Exception)` does still return `true` for
`TaskCanceledException`/`OperationCanceledException`, but that is **correct and
intentional**: a `TaskCanceledException` raised by an HTTP timeout *is* a genuine
transient failure, and the only caller (`InvokeHttpAsync:238`) is unreachable for a
caller-cancellation because the two preceding `when`-filtered catches intercept it
first. The transient-vs-cancel decision is contextual (which token fired) and cannot
be made from the exception type alone, which is exactly why the call site does it.
No source change was required. A regression guard,
`CachedCall_CallerCancellation_IsNotBufferedAsTransient`, was added: it cancels the
caller token mid-`CachedCall` and asserts an `OperationCanceledException` is thrown and
the S&F buffer remains empty (the cancelled work is not retried). The existing
`Call_CallerCancellation_IsNotMisreportedAsTimeout` covers the synchronous `Call` path.

### ExternalSystemGateway-009 — `StoreAndForwardResult` from `EnqueueAsync` is discarded; permanent failures during buffering are swallowed

| | |
|--|--|
| Severity | Medium — re-triaged: root cause subsumed by the ExternalSystemGateway-003 dispatch redesign (see Resolution) |
| Category | Correctness & logic bugs |
| Status | Resolved |
| Location | `src/ScadaLink.ExternalSystemGateway/ExternalSystemClient.cs:109-117` |

**Description**

`CachedCallAsync` assigns the result of `_storeAndForward.EnqueueAsync(...)` to
`sfResult` and then never reads it — it unconditionally returns
`new ExternalCallResult(true, null, null, WasBuffered: true)`. `EnqueueAsync` can
return `Success == false` (a permanent failure encountered during its
immediate-delivery attempt — `StoreAndForwardService.cs:142`) or `Buffered == false`
(delivered immediately). In both cases the ESG still reports the call as buffered and
successful to the script. A permanent failure surfaced by the S&F immediate attempt is
therefore silently lost instead of being returned to the script as the design requires
("On permanent failure (HTTP 4xx), the error is returned synchronously").

**Recommendation**

Inspect `sfResult`: if `Success == false` return an error `ExternalCallResult`; set
`WasBuffered` from `sfResult.Buffered` rather than hard-coding `true`. (This finding is
partly subsumed by the dispatch redesign in finding 003.)

**Resolution**

Resolved 2026-05-16 (commit pending). **Re-triage:** the stated root cause — "a
permanent failure surfaced by `EnqueueAsync`'s immediate-delivery attempt is silently
lost" — **can no longer occur** in the current source, and the dead `sfResult` variable
the finding cites has already been removed. The `ExternalSystemGateway-003` fix changed
`CachedCallAsync` to call `EnqueueAsync` with `attemptImmediateDelivery: false`. With
that flag, `EnqueueAsync` never invokes the registered delivery handler: it skips the
immediate-delivery block entirely (so the `StoreAndForwardResult(false, …, …)`
permanent-failure return at `StoreAndForwardService.cs:147` is unreachable from this
caller) and unconditionally buffers, returning `Accepted: true, WasBuffered: true`
(`StoreAndForwardService.cs:180`). The `ExternalCallResult(true, null, null,
WasBuffered: true)` that `CachedCallAsync` returns is therefore now factually correct
in every reachable case — the message *is* buffered and there is no swallowed permanent
failure. Permanent HTTP 4xx failures are still surfaced synchronously, because
`CachedCallAsync` makes its own first HTTP attempt and catches
`PermanentExternalSystemException` *before* it ever reaches `EnqueueAsync`. No source
change was required beyond the `ExternalSystemGateway-003` redesign that already landed.
Coverage: `CachedCall_TransientFailure_BuffersWithSystemRetrySettings` asserts both
`result.WasBuffered == true` and that the message is genuinely present in the S&F buffer
(depth == 1), confirming the `WasBuffered: true` claim is not a lie; the existing
`CachedCall` permanent-failure path is exercised by `Call_Permanent400_ReturnsPermanentError`
semantics shared via `InvokeHttpAsync`.

### ExternalSystemGateway-010 — `GetConnectionAsync` leaks the `SqlConnection` when `OpenAsync` fails

| | |
|--|--|
| Severity | Medium |
| Category | Performance & resource management |
| Status | Resolved |
| Location | `src/ScadaLink.ExternalSystemGateway/DatabaseGateway.cs:48-50` |

**Description**

`GetConnectionAsync` constructs `new SqlConnection(...)` and calls `await
connection.OpenAsync(...)`. If `OpenAsync` throws (unreachable server, bad
credentials, cancellation) the just-created `SqlConnection` instance is never disposed
— the exception propagates and the local reference is lost. While an unopened
`SqlConnection` is lightweight, over many failing calls this is an avoidable leak. The
design doc says `Database.Connection()` failures return an error to the script; the
current code lets a raw `SqlException` escape, which is acceptable, but the leak is
not.

**Recommendation**

Wrap the open in a try/catch that disposes the connection before rethrowing:
`try { await connection.OpenAsync(ct); } catch { connection.Dispose(); throw; }`.

**Resolution**

Resolved 2026-05-16 (commit pending). `GetConnectionAsync` now wraps `OpenAsync` in a
`try/catch` that calls `await connection.DisposeAsync()` before rethrowing, so a failed
open (unreachable server, bad credentials, cancellation) no longer leaks the
`SqlConnection`. Connection creation was extracted into an `internal virtual
CreateConnection(string)` factory so the failure path is unit-testable. Regression test
`GetConnection_OpenFails_DisposesConnectionBeforeRethrowing` substitutes a `DbConnection`
whose `OpenAsync` always throws and asserts the connection is disposed when the
exception propagates; it was verified to fail before the `try/catch` was added.

### ExternalSystemGateway-011 — Every call performs a full repository scan of all systems and methods

| | |
|--|--|
| Severity | Low |
| Category | Performance & resource management |
| Status | Resolved |
| Location | `src/ScadaLink.ExternalSystemGateway/ExternalSystemClient.cs:360-374`, `src/ScadaLink.ExternalSystemGateway/DatabaseGateway.cs:169-176` |

**Description**

`ResolveSystemAndMethodAsync` calls `GetAllExternalSystemsAsync()` and then
`GetMethodsByExternalSystemIdAsync()` and filters in memory on every single call;
`ResolveConnectionAsync` calls `GetAllDatabaseConnectionsAsync()` and filters in memory
on every cached write / connection request. At sites this hits the SQLite repository,
and `SiteExternalSystemRepository` re-reads and re-parses the methods JSON each time.
For a hot script path this is unnecessary repeated I/O and allocation. Definitions only
change on deployment, so they are eminently cacheable.

**Recommendation**

Add an in-memory cache of system/method/connection definitions keyed by name,
invalidated on artifact deployment. Alternatively use a name-keyed repository lookup
rather than fetch-all-then-filter.

**Resolution**

Resolved 2026-05-16 (commit `<pending>`). A cross-module change was explicitly
authorized, so the **name-keyed repository lookup** recommendation was applied — the
cleaner of the two options, and one that avoids the staleness hazard a
deployment-invalidated cache would introduce.

Three name-keyed methods were added to `IExternalSystemRepository`
(`ScadaLink.Commons`): `GetExternalSystemByNameAsync(name)`,
`GetMethodByNameAsync(externalSystemId, methodName)` and
`GetDatabaseConnectionByNameAsync(name)`. The connection lookup belongs on the same
interface because database connection definitions are already part of
`IExternalSystemRepository` (alongside `GetAllDatabaseConnectionsAsync` /
`GetDatabaseConnectionByIdAsync`), so the existing repository organization was
followed rather than introducing a new interface.

Both implementers of the interface were updated:

- `ScadaLink.ConfigurationDatabase.ExternalSystemRepository` — all three are genuine
  server-side keyed queries (`FirstOrDefaultAsync(x => x.Name == name)`, the
  method lookup additionally scoped by `ExternalSystemDefinitionId`), matching the
  existing `GetMethodByNameAsync` / `GetListByNameAsync` / `GetSharedScriptByNameAsync`
  convention in the other Central repositories.
- `ScadaLink.SiteRuntime.SiteExternalSystemRepository` — `GetExternalSystemByNameAsync`
  and `GetDatabaseConnectionByNameAsync` are genuine single-row indexed SQLite queries
  (`WHERE name = @name`; both tables have `name` as the PRIMARY KEY).
  `GetMethodByNameAsync` resolves the named method from the parent system's
  `method_definitions` JSON column; this still requires resolving the parent system
  (one id→name scan), but the gateway's new call sequence performs **one** scan instead
  of the previous **two** (`GetAllExternalSystemsAsync` + `GetMethodsByExternalSystemIdAsync`,
  which itself scanned), so the site path is strictly better than before — noted as
  the one place a residual scan remains, bounded by the deployed system count.

`ExternalSystemClient.ResolveSystemAndMethodAsync` and
`DatabaseGateway.ResolveConnectionAsync` were rewritten to call the keyed lookups; the
`GetAllExternalSystemsAsync` / `GetMethodsByExternalSystemIdAsync` /
`GetAllDatabaseConnectionsAsync` + in-memory `FirstOrDefault` filtering is gone from
both hot paths.

Regression tests (TDD — written first, verified failing/not-compiling before the
implementation, then confirmed green; one was additionally verified to fail when the
keyed query is deliberately broken):

- ConfigurationDatabase (`RepositoryCoverageTests`):
  `GetExternalSystemByName_ReturnsMatchingRow`,
  `GetExternalSystemByName_MissingName_ReturnsNull`,
  `GetMethodByName_ReturnsMethodScopedToParentSystem`,
  `GetMethodByName_MissingMethod_ReturnsNull`,
  `GetDatabaseConnectionByName_ReturnsMatchingRow`,
  `GetDatabaseConnectionByName_MissingName_ReturnsNull`.
- SiteRuntime (`SiteRepositoryTests`):
  `ExternalSystemRepository_GetByName_ReturnsMatchingDefinition`,
  `ExternalSystemRepository_GetByName_MissingName_ReturnsNull`,
  `ExternalSystemRepository_GetMethodByName_ResolvesScopedToSystem`,
  `DatabaseConnectionRepository_GetByName_ReturnsMatchingDefinition`.
- ExternalSystemGateway: the existing `ExternalSystemClientTests` /
  `DatabaseGatewayTests` resolution stubs were migrated to the keyed methods (via
  `StubResolution` / `StubConnection` helpers), so the full gateway suite now exercises
  and protects the keyed-lookup resolution path.

`dotnet build ScadaLink.slnx` is clean; `ScadaLink.ExternalSystemGateway.Tests` (54),
`ScadaLink.ConfigurationDatabase.Tests` (106), `ScadaLink.SiteRuntime.Tests` (196) and
`ScadaLink.Commons.Tests` (226) all pass.

### ExternalSystemGateway-012 — Permanent-failure logging requirement is not met; `_logger` is injected but unused

| | |
|--|--|
| Severity | Low |
| Category | Design-document adherence |
| Status | Resolved |
| Location | `src/ScadaLink.ExternalSystemGateway/ExternalSystemClient.cs:24,169-177`, `src/ScadaLink.ExternalSystemGateway/DatabaseGateway.cs:22` |

**Description**

The design doc states permanent failures are "Logged to Site Event Logging", but
`InvokeHttpAsync` performs no logging on the permanent-failure path. In fact the
injected `ILogger<ExternalSystemClient>` and `ILogger<DatabaseGateway>` fields are
never used at all in either class. Either the logging is expected to happen in the
caller (Script Execution Actor) — in which case the design doc is imprecise about
where — or it is missing. Separately, `IsTransient(HttpStatusCode)` treats any
non-success, non-(5xx/408/429) status as permanent without an explicit comment, which
is a reasonable default but undocumented.

**Recommendation**

Add a `_logger.LogWarning` on the permanent-failure path (and a debug log on
transient), or clarify in the design doc that Site Event Logging capture is the
caller's responsibility and remove the unused `_logger` fields. Add a comment in
`ErrorClassifier` documenting the "default to permanent" behaviour.

**Resolution**

Resolved 2026-05-16 (commit pending). Partial re-triage of the finding text: the
"`_logger` injected but unused" claim is **stale** — both loggers are already used (the
`DeliverBufferedAsync` retry-sweep handlers added by `ExternalSystemGateway-001` log
park/error events). The genuine remaining gap — `InvokeHttpAsync` performing no logging
on the HTTP-failure paths — is now fixed: `InvokeHttpAsync` emits a
`_logger.LogWarning` on the permanent-failure path (status code, system, method,
truncated error body) so a permanent failure is visible in Site Event Logging as the
design requires, and a `_logger.LogDebug` on the transient-failure path (transient
failures are normal retry/S&F operation and must not be noisy at warning level). A
documentation comment was added to `ErrorClassifier.IsTransient(HttpStatusCode)`
explaining the "any other non-success status defaults to permanent" behaviour and why
permanent is the safe default. Regression tests: `Call_PermanentFailure_LogsAWarning`
(asserts a warning carrying the system name is emitted; verified to fail before the
`LogWarning` was added) and `Call_TransientFailure_DoesNotLogAtWarningOrAbove` (guards
against over-logging transient failures).

### ExternalSystemGateway-013 — `MaxConcurrentConnectionsPerSystem` and `DefaultHttpTimeout` options are defined but never used

| | |
|--|--|
| Severity | Low |
| Category | Code organization & conventions |
| Status | Resolved |
| Location | `src/ScadaLink.ExternalSystemGateway/ExternalSystemGatewayOptions.cs:9,12`, `src/ScadaLink.ExternalSystemGateway/ServiceCollectionExtensions.cs:13` |

**Description**

`ExternalSystemGatewayOptions.MaxConcurrentConnectionsPerSystem` (default 10) and
`DefaultHttpTimeout` (default 30s) are bound from configuration but neither is read
anywhere. `AddHttpClient()` registers the default factory with no
`ConfigurePrimaryHttpMessageHandler`/`SocketsHttpHandler` `MaxConnectionsPerServer` and
no `Timeout`, so both options have no effect. An operator setting these values gets
them silently ignored — a misleading configuration surface (`DefaultHttpTimeout` is
also referenced by finding 002).

**Recommendation**

Either wire the options into a named/typed `HttpClient` registration (set
`MaxConnectionsPerServer` on the primary handler, set `Timeout`), or remove the unused
options to avoid implying behaviour that does not exist.

**Resolution**

Resolved 2026-05-16 (commit pending). Partial re-triage of the finding text:
`DefaultHttpTimeout` is **no longer unused** — it became the effective per-call HTTP
round-trip limit when `ExternalSystemGateway-002` was fixed (`InvokeHttpAsync` builds a
linked `CancellationTokenSource(DefaultHttpTimeout)`). The genuinely-unused option,
`MaxConcurrentConnectionsPerSystem`, is now wired in: `AddExternalSystemGateway` adds a
`ConfigureHttpClientDefaults` registration that supplies a `SocketsHttpHandler` whose
`MaxConnectionsPerServer` is bound from the option, so it applies to every per-system
named client (`ExternalSystem_{name}`) the gateway creates rather than being silently
ignored. Regression test
`ServiceWiringTests.MaxConcurrentConnectionsPerSystem_IsAppliedToTheNamedHttpClientPrimaryHandler`
builds the DI container, resolves the named client's handler chain via
`IHttpMessageHandlerFactory`, walks to the primary handler and asserts
`SocketsHttpHandler.MaxConnectionsPerServer` equals the configured value; it was verified
to fail before the wiring was added.

### ExternalSystemGateway-014 — Cached-call buffering path and `DatabaseGateway` are untested

| | |
|--|--|
| Severity | Low |
| Category | Testing coverage |
| Status | Resolved |
| Location | `tests/ScadaLink.ExternalSystemGateway.Tests/ExternalSystemClientTests.cs:1`, `tests/ScadaLink.ExternalSystemGateway.Tests/DatabaseGatewayTests.cs` |

**Description**

`ExternalSystemClientTests` covers system/method not-found, success, transient 500 and
permanent 400 for `CallAsync`, plus `CachedCall` not-found and success. It does **not**
cover: the `CachedCall` transient-failure → S&F buffering branch (the most
behaviour-rich path, including the `_storeAndForward == null` fallback and `WasBuffered`
semantics), the `CachedCall` permanent-failure branch, connection-exception
classification (`HttpRequestException` thrown by the handler), `BuildUrl` query-string
construction, and `ApplyAuth` for the apikey/basic variants. There is **no test file
for `DatabaseGateway`** at all — `GetConnectionAsync` not-found, `CachedWriteAsync`
not-found, and the `_storeAndForward == null` guard are entirely uncovered. The
`MockHttpMessageHandler` also does not assert request URL/headers/body, so auth and
URL construction are unverified.

**Recommendation**

Add tests for the `CachedCall` transient/buffering paths (with a substituted S&F
service), `DatabaseGateway` not-found and null-S&F guards, and `BuildUrl`/`ApplyAuth`
by asserting on the captured `HttpRequestMessage` in the mock handler.

**Resolution**

Resolved 2026-05-16 (commit pending). Partial re-triage of the finding text: a number
of the listed gaps were **already closed** by tests added when findings 001–010 were
fixed — `DatabaseGatewayTests.cs` now exists (not-found, null-S&F guard, buffered-write
delivery), and the `CachedCall` transient/buffering and permanent paths are covered by
the `ExternalSystemGateway-001/003/004/008` regression tests. The remaining genuine
coverage gaps are now closed with new tests:

- `Call_GetWithParameters_AppendsEscapedQueryString` — `BuildUrl` GET query-string
  construction, asserting on the captured request URI and that an `&` inside a value is
  percent-escaped.
- `Call_PostWithParameters_SendsJsonBody` — POST JSON-body construction.
- `Call_ApiKeyAuthWithDefaultHeader_SendsXApiKeyHeader`,
  `Call_ApiKeyAuthWithCustomHeader_SendsNamedHeader`,
  `Call_BasicAuth_SendsBase64AuthorizationHeader` — `ApplyAuth` for all three auth
  variants, asserting on the captured request headers.
- `Call_ConnectionError_IsClassifiedAsTransient` — a connection-level
  `HttpRequestException` is classified transient.
- `CachedWrite_BuffersTheWriteWithConnectionRetrySettings` and
  `CachedWrite_ZeroMaxRetriesIsHonouredNotTreatedAsUnset` — the `DatabaseGateway`
  `CachedWrite` happy-path enqueue against a real S&F service.

The shared `RequestCapturingHandler` test helper was extended to capture request
headers and body so URL/auth/body construction is now verified, not just status codes.
These are new-coverage tests against already-correct behaviour, so they pass on the
current source; the `BuildUrl` and `ApplyAuth` paths they exercise are now protected
against regression.

### ExternalSystemGateway-015 — `MaxRetries == 0` is buffered as "retry forever", contradicting the ExternalSystemGateway-004 "never retry" claim

| | |
|--|--|
| Severity | High |
| Category | Correctness & logic bugs |
| Status | Resolved |
| Location | `src/ScadaLink.ExternalSystemGateway/ExternalSystemClient.cs:120-127`, `src/ScadaLink.ExternalSystemGateway/DatabaseGateway.cs:102-108` |

**Description**

The `ExternalSystemGateway-004` fix removed the `MaxRetries > 0 ? ... : null` guard so
that `CachedCallAsync` and `CachedWriteAsync` now pass the definition's `MaxRetries`
to `StoreAndForwardService.EnqueueAsync` verbatim. The stated rationale (and the inline
comments at `ExternalSystemClient.cs:118-119` / `DatabaseGateway.cs:99-101`, plus the
tests `CachedCall_TransientFailure_ZeroMaxRetriesIsHonouredNotTreatedAsUnset` and
`CachedWrite_ZeroMaxRetriesIsHonouredNotTreatedAsUnset`) is that a configured
`MaxRetries` of `0` means **"never retry"**.

That is the opposite of what the Store-and-Forward engine actually does with the
value. `EnqueueAsync` stores the passed `maxRetries` directly into
`StoreAndForwardMessage.MaxRetries` (`StoreAndForwardService.cs:139`), whose own XML
doc states **"Maximum retry-sweep attempts before parking (0 = no limit)"**
(`StoreAndForwardMessage.cs:30`). The retry sweep enforces it as
`if (message.MaxRetries > 0 && message.RetryCount >= message.MaxRetries)`
(`StoreAndForwardService.cs:285`) — when `MaxRetries == 0` that guard is false on
every sweep, so the message is **never parked and is retried forever**.

Consequences for a `CachedCall`/`CachedWrite` against a system or connection
configured with `MaxRetries = 0`:

1. A transiently-failing message that the operator intended never to retry is instead
   retried on every sweep indefinitely, accumulating in the buffer and repeatedly
   re-dispatching the request — the exact unbounded-retry / duplicate-delivery hazard
   the idempotency note in the design doc warns about.
2. The two ESG regression tests cited above assert `max_retries == 0` is *stored* and
   describe that as "honoured" — they verify the persisted column, never the resulting
   sweep behaviour, so they lock in the broken semantics.
3. Because the SiteRuntime repository still always supplies `MaxRetries == 0` (the
   open companion gap noted in `ExternalSystemGateway-004`), **every** cached call and
   cached write at every site currently buffers as retry-forever. Before the ESG-004
   fix the `> 0` guard sent `null`, so the S&F `DefaultMaxRetries` (a bounded value)
   applied — i.e. the ESG-004 fix turned a bounded retry into an unbounded one for the
   common path.

**Recommendation**

Reconcile the ESG and S&F interpretations of `MaxRetries == 0` — they must agree.
Either: (a) treat the entity's `MaxRetries == 0` as "unset" and pass `null` so the
bounded S&F default applies (reverting to the pre-ESG-004 behaviour, and accepting
that "never retry" then needs a different representation such as a nullable field or a
`MaxRetries == 1` convention); or (b) if `0` genuinely must mean "never retry", add an
explicit no-retry path — e.g. do not enqueue at all on transient failure when
`MaxRetries == 0`, or introduce a distinct sentinel — and fix the
`StoreAndForwardMessage.MaxRetries` doc and `RetryMessageAsync` guard so `0` no longer
means "no limit". Update the two `ZeroMaxRetries...` tests to assert the *sweep*
outcome (parked / not retried), not just the stored column value.

**Resolution**

Resolved 2026-05-17. Root cause confirmed: the S&F engine treats a stored
`MaxRetries == 0` as "no limit / retry forever" (`StoreAndForwardMessage.MaxRetries`
doc "0 = no limit"; sweep guard `MaxRetries > 0 && RetryCount >= MaxRetries`), while
the entity's non-nullable `int MaxRetries` defaults to `0` — so passing it verbatim
buffered every cached call/write as an unbounded retry loop. Fix (ESG-side only,
recommendation (a)): `CachedCallAsync` and `CachedWriteAsync` now pass
`MaxRetries > 0 ? MaxRetries : null`, so an entity `0` is treated as "unset" and the
bounded S&F `DefaultMaxRetries` applies; the misleading "0 = never retry" inline
comments were corrected. The two `ZeroMaxRetries...` tests were rewritten to
`CachedCall_TransientFailure_ZeroMaxRetriesIsTreatedAsUnsetNotRetryForever` /
`CachedWrite_ZeroMaxRetriesIsTreatedAsUnsetNotRetryForever`, asserting the buffered
message carries the bounded default (99) and never `0`.

### ExternalSystemGateway-016 — `ConfigureHttpClientDefaults` applies the ESG connection cap to every `HttpClient` in the host process

| | |
|--|--|
| Severity | Medium |
| Category | Code organization & conventions |
| Status | Resolved |
| Location | `src/ScadaLink.ExternalSystemGateway/ServiceCollectionExtensions.cs:21-29` |

**Description**

The `ExternalSystemGateway-013` fix wires `MaxConcurrentConnectionsPerSystem` by
calling `services.ConfigureHttpClientDefaults(builder => builder.ConfigurePrimaryHttpMessageHandler(...))`.
The inline comment claims this "applies to the dynamically-named clients created by
`ExternalSystemClient`" — but `ConfigureHttpClientDefaults` is **process-global**: it
adds the configuration to *every* `HttpClient`/`IHttpClientFactory` client created
anywhere in the host, regardless of name.

The Host registers the External System Gateway alongside other components that also
use `IHttpClientFactory` — notably `ScadaLink.NotificationService` (`OAuth2TokenService`
and its `ServiceCollectionExtensions` call `AddHttpClient`). With the ESG registration
present, the OAuth2 token client (and any future `HttpClient` consumer in the host)
has its **primary handler replaced** by a `SocketsHttpHandler` whose
`MaxConnectionsPerServer` is the ESG's `MaxConcurrentConnectionsPerSystem`. That:

1. Silently caps unrelated clients at a value owned by, and named for, a different
   component — an operator tuning the ESG option unknowingly throttles Microsoft 365
   OAuth2 token acquisition.
2. Overrides/discards any primary-handler configuration those other components add for
   their own clients (e.g. a custom handler, proxy, or certificate settings).

This is a leaky, surprising side effect for what the option claims to be a per-ESG
setting; `ConfigureHttpClientDefaults` should not be used to express a single
component's policy.

**Recommendation**

Scope the handler configuration to the gateway's own clients only. The ESG already
creates per-system clients with a deterministic name pattern
(`ExternalSystem_{system.Name}`); register a typed/named `HttpClient` (or a small
factory abstraction) for that pattern and call `ConfigurePrimaryHttpMessageHandler`
on that registration instead of on the global defaults. If a name-pattern registration
is impractical, document the global effect explicitly and rename the option, but the
preferred fix is to stop using `ConfigureHttpClientDefaults`.

**Resolution**

Resolved 2026-05-17. Root cause confirmed: `ConfigureHttpClientDefaults` is
process-global and replaced the primary handler of every `IHttpClientFactory` client
in the host, leaking the ESG connection cap onto unrelated clients. Fix: the global
`ConfigureHttpClientDefaults` registration was replaced with an
`IConfigureNamedOptions<HttpClientFactoryOptions>` (`GatewayHttpClientConfigurator`)
that applies the `SocketsHttpHandler`/`MaxConnectionsPerServer` cap only to clients
whose name starts with `ExternalSystem_` (the gateway's own per-system clients), so
clients owned by other components keep their own (or the framework default) primary
handler. Regression test
`ServiceWiringTests.MaxConcurrentConnectionsPerSystem_IsNotAppliedToNonGatewayHttpClients`
asserts a non-gateway client does not inherit the cap while the gateway client still
does; it was verified to fail before the fix.

### ExternalSystemGateway-017 — `BuildUrl` appends a bare trailing `?` when a GET method's parameters are all null

| | |
|--|--|
| Severity | Low |
| Category | Correctness & logic bugs |
| Status | Resolved |
| Location | `src/ScadaLink.ExternalSystemGateway/ExternalSystemClient.cs:324-333` |

**Description**

In `BuildUrl`, the GET/DELETE query-string branch is entered when
`parameters != null && parameters.Count > 0`, but the projection then filters out
null-valued entries (`parameters.Where(p => p.Value != null)`). When a GET/DELETE
method is invoked with a non-empty parameter dictionary whose values are *all* null,
`queryString` is the empty string and the code still executes `url += "?" + queryString`,
producing a URL ending in a bare `?` (e.g. `https://host/api/resource?`). Most servers
tolerate a trailing `?`, but it is an unintended artifact, can defeat response caching
on some endpoints, and makes captured request URLs harder to read in logs.

**Recommendation**

Only append `"?" + queryString` when `queryString` is non-empty (compute the joined
string first and check it), so a method whose effective parameter set is empty
produces a clean URL identical to the no-parameters case.

**Resolution**

Resolved 2026-05-17. Root cause confirmed: `BuildUrl` appended `"?" + queryString`
whenever the GET/DELETE parameter dictionary was non-empty, even when every value
was null and `queryString` was the empty string, yielding a bare trailing `?`. Fix:
`BuildUrl` now appends `"?" + queryString` only when `queryString.Length > 0`, so a
method whose effective parameter set is empty produces a URL identical to the
no-parameters case. Regression test
`Call_GetWithAllNullParameters_DoesNotAppendTrailingQuestionMark` asserts the
captured request URI has no trailing `?`; it was verified to fail before the fix.