[erossignon]

Findings from stress-testing @node-i3x/mcp over stdio against a portal instance (about 450 calls). Before → after on the same instance:

Finding	Before	After
5 parallel explore {maxDepth:3}	all failed: i3X Gateway request timed out after 15000ms for /v1/objects/related	about 1.6 s each
i3x_connect with a wrong token	connected, then every call 401	auth_failed, previous connection kept
telemetry with 200 copies of one id	5.8 MB	29 KB
inspect with a 5000-char id	10.9 s, blocking the event loop	108 ms
history of an Object whose variables keep no history	success, 126 empty series, count 0	per-component 501 … not supported plus a note
monitor on a healthy asset	unused alarm sub-states (UncertainInitialValue) reported as Bad	reported as GoodNoData (needs the server fix deployed)

Causes and fixes:

• Explore: it sent one /objects/related request per node through an unbounded Promise.all, and one timeout aborted the whole walk. It now walks breadth-first with one bulk request per batch of each level, under a node budget (maxNodes, default 500), and records failures on the node. maxDepth: 0 now means unlimited, as in the value tools; it used to silently become 2.
• Connect: it verified against /v1/info, which is public. It now also reads /v1/namespaces, and keeps the current token when re-connecting to the same URL without one.
• Size: elementIds are deduplicated and capped at 100. Tool results are capped at 100 000 characters (I3X_MCP_MAX_RESULT_CHARS) with a notice saying how to narrow the query. Suggestions skip fuzzy matching for queries over 128 characters.
• Core, history: _readComponentsHistory turned a component read failure into { values: [] }. The component now carries responseDetail (501), an additive field.
• Core, quality: any non-zero OPC UA status became Bad. Quality now follows severity: Good subcodes stay Good, Uncertain becomes Uncertain (or GoodNoData when the value is null, per the spec's null rules). The pseudo-session connector no longer emits status names such as UncertainInitialValue as the quality.

Tests: packages/mcp/tests/contract-stress.test.ts (real rest-server through inject()), plus core mapping and history, pseudo-session history quality and monitor summary tests. All 19 new tests fail on main's sources.

Caveats:

• Quality changes server-wide: Good subcodes that were reported Bad are now Good, and Uncertain statuses are now Uncertain/GoodNoData. One existing test asserted that status 1 (Good severity, info bits only) is Bad; it was changed to 0x80000000.
• Component history failures are still all reported as 501 "not supported", whatever the underlying error, as element-level failures already were.
• The before/after numbers were measured against a portal running the previous server. The two core fixes are covered by tests only until that server is redeployed.
• The pre-push hook's demo-embedded smoke test is still flaky under load: it failed the first two pushes of this branch, then passed.