[ako]

Syncs ako/mxcli:main into upstream: 12 commits since #1793 — mxcli test can run tests in the live app without building them, plus microflow, page, catalog and agent fixes. 3 upstream issues fixed.

mxcli test: run a test in the live app without building it

• New: --allow-call. run --local --test-endpoint --allow-call switches on an exec route in the test endpoint, and test --attach compiles each .test.mdl test into versioned steps the app runs with the Core API, every expression evaluated by Core.evaluateExpression: nothing is injected or built. Verified on 11.12.1: seven tests give identical verdicts and messages on both routes, 0.26 s against 3.2 s. Because the runtime evaluator returns empty for an unknown name, an undeclared variable or unknown enumeration value is refused (ERROR), never compared silently against empty; a model newer than the running build is refused. The route takes POST, needs the token and a loopback caller, and answers 403 without --allow-call; run warns when it is combined with --hub. Design from the review of feat: mxcli test call — JSON runtime executor against an already-running app #1789 (refs feat: mxcli test call — JSON runtime executor against an already-running app #1788); a spike measuring Core.evaluateExpression is included.
• Each test is routed on its own: an attach injects and builds only the tests the executor can't run, and the executor now runs count/sum/average/minimum/maximum over an attribute and retrieves over an association. A differential run found Mendix's average over an empty list is empty, not 0; the executor agrees.
• A test run holds the dev loop's watcher instead of making a run --local --watch loop rebuild and restart for the injection and its cleanup (which dropped every browser session). The hold (.mxcli/watch-hold.json) records the model digest; on release nothing is built if the model is the one the hold began from, and --attach now also restores the .mpr byte for byte.
• No phantom "change pending" under --test-endpoint (run status / run wait now ignore the build's own theme scaffold, as the watcher already did), and an endpoint left behind by a killed run is removed by the next mxcli run, with run and check -p warning while AfterStartupMicroflow still points at it.

Microflows

• A change inside one branch of a multi-branch case is spliced in place. create or modify was refused ("cannot replace the ExclusiveSplit … it has 9 paths, not two") whenever anything differed inside one branch of an enumeration case: the body diff matched if and loop shells but not a case's. Verified with mx check 0 errors and a re-run reporting Unchanged. Fixes exec refuses a create or modify that check passes: unchanged multi-path case treated as replaced (MDL-V1-REBUILD) #1778

Pages

• alter page … insert into <layoutgrid> { row { … } } appends rows. It failed with "row needs a name" and "not a simple container". Rows now get their own path, as tab pages do; BEFORE/AFTER (a row has no stored name) and mixing rows with widgets are refused. Fixes ALTER PAGE INSERT INTO a layout grid fails: a bare row needs a name and a layout grid is not accepted as a container #1791

Catalog and lint

• Sequence flows and activity positions for lint rules. CATALOG.FLOWS holds one row per sequence and annotation flow, with origin and destination, IsErrorHandler, CaseValue and the connection sides; flows_for() exposes it, and activities gain X, Y, Width and Height — so a rule can tie a log to its error handler, a note to its activity, or a branch to its case value. Fixes Starlark lint API: no sequence flows (error handler path, annotation link, split branches, canvas position) #1308
• Agent tools match ToolCount: AGENT_TOOLS no longer includes consumed MCP services (they're in AGENT_MCP_SERVICES), and CONSUMED_MCP_SERVICES gains Endpoint and AuthenticationMicroflow, with refs to both. Catalog schema 44.
• MDL-AGENT02/03 now fire for a new agent or MCP service written with create or modify (they skipped every create or modify, since an omitted key keeps the stored value — but a new document has none). Lint gains MPR014 (agent without a selected model) and MPR015 (consumed MCP service without an endpoint).

Docs

• A genai-runtime skill for the runtime side of Agents Kit 2: registering a Mendix Cloud GenAI key or OpenAI-compatible endpoint, seeding agents from JSON, Conversational UI chats, PgVector and Mendix Cloud knowledge bases, traces — with the gotchas no check sees (the shared encryption key, the vector extension only the UI's Test button creates, the indexing delay, StoreTraces off by default)