Sync from ako/mxcli: microflow canvas layout authoring (#884), marketplace cache, dev container fix

[ako]

Title

Sync from ako/mxcli: microflow canvas layout authoring (#884), marketplace cache, dev container fix

Description

Sync of ako/mxcli:main — 7 non-merge commits since #885, 31 files, +2202 / −99. mendixlabs/main is fully contained in this branch (0 commits ahead), so the diff is additive and conflict-free.

Three threads: a feature set closing most of #884, two contained fixes, and two proposals.

---

1. Microflow canvas layout — upstream #884 (3 commits)

The reporter generates TOF-convention microflows from MDL and controls layout with @position/@anchor. Three of their four problems are fixed here. Every claim was checked against the code and against Mendix's own storage first; all four stand up, two turned out worse than filed, and one asked for something Mendix does not store.

Unknown annotations were silently dropped → MDL059

The grammar accepts any @name and the visitor's switch had no default arm. The reporter found this as "@size(600, 300) parses without error but produces no effect". The sharper case is a typo of an annotation that works:

@postion(999, 111)   -- check passed; the position was silently discarded

On a workflow whose whole point is scripted layout, that throws away the thing being authored and reports success. Now an error, enforced by exec as well as check.

Annotations are read reflectively: fifteen statement types carry them, and a hand-written type switch that skips the sixteenth reintroduces the silent drop one level up. Two tests parse the source to keep this honest — one asserts every *ActivityAnnotations field is named so the accessor reaches it, the other parses the visitor's own case labels and compares them against the validator's list in both directions. Checked against 407 real MDL blocks for false positives.

A hand-curved edge was flattened by any rewrite → @curve

The request asked for edge waypoints. Mendix stores none: a flow's shape is two bezier control vectors on its Microflows$BezierCurve line. So the syntax maps to what exists:

@curve(from: (40, -90), to: (-40, 90))

Both writers already emitted the vectors and the legacy parser already read them, but nothing could set them, so they defaulted to "0;0" and every rebuild flattened the curve — measured by patching "40;-90" into the stored line and re-running the same script unchanged. Data loss, not a missing feature. Needed no grammar change: name: (x, y) is already annotationParenValue.

The merge that closes a split was unpositionable → @merge

A statement's @position belongs to the split, so its end-if join had no annotation and routinely landed on a neighbouring activity.

@merge(900, 400)

All three split builders read through one mergePosition helper, so the override cannot be honoured at one split type and ignored at another.

MPR008 compared nodes on different canvases

The rule flattened a LoopedActivity's children — positioned relative to the container — into the same list as absolute canvas coordinates, and reported overlaps that cannot happen on screen. Verified in the stored BSON: an outer activity at 200;230 beside a loop child at 141;130. Now one plane per canvas.

Note for the reporter: their quoted pair does not reproduce. (141,130) vs (134,230) is dy=100 against activityBoxHeight = 60, so it never fired. The mechanism is real; that evidence was not.

The design constraint shared by all three

Both new annotations are applied at one choke point. Threading a curve alongside the anchor would mean editing all seven sites that create a flow, and threading the describer's splitMergeMap to emit @merge would mean editing ten-plus call sites of emitObjectAnnotations — the multi-site change that fails silently when one is missed. Instead: the curve is recorded against the activity and stamped in a single pass after the graph is built, and commonMergeAfter finds a split's merge by walking its branches using data already in scope. The walk is bounded and node-capped, because a retry loop makes the flow graph cyclic.

DESCRIBE emits both, which is what makes them real rather than write-only — the reporter's "DESCRIBE output is identical before and after manual adjustment" was the read half of the same bug. @merge was implemented and reverted from the first PR for exactly this reason, and re-landed only once the describe half existed.

2. Fixes (2 commits)

• Marketplace reference cache stored a whole blank Mendix app — 34 MB per entry, of which widgets/ (9.6 MB), themesource/ (6.4 MB), theme-cache/ (2.1 MB) and javascriptsource/ (1.6 MB) are never read. Both consumers take the .mpr and nothing beside it. Entries are now model-only.
• The generated dev container installed postgresql-client only — so mxcli run --local --ensure-db had no service to start and no superuser, while psql on PATH made the container look correctly provisioned. Adds the server package, asserted for both the docker and podman variants.

3. Proposals (2 commits, docs only)

• mxcli init --solution — two existing proposals assume a repo where several .mpr projects sit under one workspace, and nothing creates that shape: init discovers only the first .mpr and writes a dev container per project, so two projects yield two competing containers. Scoped deliberately to the repo-shape prerequisite, with the orchestration half left in warm-loop slice 5 as a non-goal.
• mxcli developer panel — analyze-runtime.md already covers joining logs, metrics, traces and the catalog; what it cannot supply is which spans belong to the thing the user just did. Proposes a panel that captures the interaction plus an in-process OTLP collector and an MCP surface. Deliberately avoids the unverified push API: the panel records, the agent pulls.

---

Verification

mxbuild 11.6.6, both engines (modelsdk and legacy), for the #884 work:

Check	Result
mx check on every fixture	0 errors
@curve on the annotated activity's flow	40;-90 / -40;90, 0;0 elsewhere
@merge	stored at 900;400 as authored
Re-executing DESCRIBE's own output	both preserved
MPR008 cross-canvas / same-canvas	silent / still reported
MDL corpus (407 blocks)	no false positives
unit suite, -tags integration	green

Two process notes worth carrying upstream. My first @curve tests called the functions directly; deleting both call sites left them green — they proved the functions worked and nothing about the wiring. Every annotation added here now has a test that fails when unwired, verified by removing each call site in turn. Separately, an intermediate "no MPR008 violations" reading turned out to be the scratch project having been reaped rather than the fix working, so that A/B was redone with a pre-fix binary against a rebuilt project.

Scope of my own testing: I authored and verified the #884 thread end-to-end. The marketplace cache, dev container fix and both proposals came from other sessions; they build and pass the suite in this branch, but their own measurements are theirs, not re-run by me.

Still open in #884

Problem 1, container Size. Diagnosis confirmed: measureStatementsSpan derives width from statement count, never from child positions, which is exactly the reporter's evidence of identical child positions producing different Size. Deriving it from a child bounding box is a layout-engine change affecting every generated flow's geometry, and wants its own proposal rather than a patch. Note that @size(600, 300) is now rejected rather than ignored, so adding it later means updating the known-annotation list — the drift test fails loudly if only one side changes, which is the intent.