Skip to content

Module updates that carry the app's edits are not routine; regressions get a case; cold-test fixes WM1 and CMM - #11

Merged
MendixMau merged 59 commits into
mainfrom
fix/signed-cards
Nov 17, 2026
Merged

MendixMau merged 59 commits into
mainfrom
fix/signed-cards

Conversation

@MendixMau

Copy link
Copy Markdown
Owner

The feature this PR is for

A Marketplace module the app edited is no longer counted as a routine update. The update replaces the whole module and the toolkit carries each edit back, so:

  1. No edits: routine, goes ahead under the plan sign-off.
  2. Edits carried over, screens the same on old and new: goes ahead. proposal.py puts these modules in their own step, "Module updates that carry the app's edits"; auto mode approves it with source carry-over (not routine), so the cockpit says which edits were carried.
  3. An edit not carried, or a screen differs: a decision with its case. A regression without a hand-written case file now gets one built from the run's own files (cockpitmatrix.auto_case): how many actions pass on old and fail on new, the old and new screenshot of the failing step, the carried screen edits as the first place to look, and the evidence lines.

New controls: proposal-steps-module-edits-carried-pass, cockpit-auto-case-pass. Docs: steps/heal.md, docs/feedback-ledger.md.

This PR carries the whole branch

fix/signed-cards has 56 commits main does not have ( 57 files changed, 3393 insertions(+), 171 deletions(-)), from cold tests WM1 and CMM on 2026-11-16 and 11-17: cockpit decision cases and per-action meaning, editroute (Marketplace pages the menu opens get the app layout in place), the 9-to-10 playbook (Postgres read route on 9.24), seed_snapshot, build and repair fixes (mx check crash, javac jar, copydoc, CE1613), widgetskew, crosscalls --own, and the feedback ledger. See the commit list.

Tests

  • Full control suite at 4a37eb8: 2648 passed, 0 failed, 60 skipped.
  • Customer names check: clean. Private names in the diff: 0 added.

🤖 Generated with Claude Code

MendixMau and others added 30 commits November 16, 2026 22:19
…n-off accepted leaves Blocked on you

Cold test WM1: after the toolkit moved to update-instead-of-keep, the signed PROPOSAL.md kept the old "keep" cards (the
signed path returned before re-rendering) and seven module cards stayed in Blocked on you while DECISIONS.md held them
CONFIRMED with the sign-off.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…uted update

REPLAY builds each element by a three-way merge of mxcli describe (installed package,
original app, updated app), sets the app's property values on the new version when the
vendor rewrote the lines around them, and holds any result mxcli check refuses. A module
the routed update keeps again is routed again (routes add up); the same reason twice holds.
Cold test WM1 Atlas_Core: 7 of 7 elements merge and pass check (was: replay refused, exit 13).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ber access; never replay a whole entity

Data test 2026-11-16 (WM1, real Postgres): save-edits wrote no file for the 4 associations the app's Email
module owns into Email_Connector; a re-created association lost its member access; a saved whole entity put
5 widened vendor attributes back to String(200). Without the original GUID, AppNote 0/3 and links 0/3 after
a green deploy; with it 3/3.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…arlier rows moved aside)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…runs

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…, reset on a rerun, score

- brainnotes.py: the company brain's practice notes for a run event, no model; --eval measures hit@k on a
  labelled set. Each route row carries its notes, the cockpit card shows them. No brain: nothing, exit 0.
- editroute reset, first thing in the update step: an earlier routing's lines leave the decision files.
  WM1 rerun: 'Atlas_Core # editroute REPLAY' forced Atlas_Core before routing and its edits were never merged.
- grep -c with no match no longer fails the update step.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…te step; brainnotes splits Atlas_Core and keeps version digits

After editroute reset emptied force-ui.txt, grep -v matched nothing and pipefail stopped the WM1 rerun at the force-ui read.
brainnotes hit@3 on the 19-situation labelled set: 15/19 -> 17/19.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…s pages; check with the mxcli exec runs

mxcli never writes a layout into a Marketplace module (any build). WM1 rerun: Atlas_Default and Atlas_TopBar were
rewritten to sit on the app's design-module layouts; the merged replay stopped the run. Now: a layout redirected to
another module's layout becomes ALTER PAGES SET LAYOUT (pages in Marketplace modules keep the vendor layout), a layout
the app added and nothing uses is not carried, any other edited layout is held for a person.
Collect checks with the mxcli exec runs: the container's (2026-10-13) refused what the local one passed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…t step's

WM1 rerun 4: Email_Connector's kept reason named step6's idcheck (3 RED association removals, an XASSOC case);
editroute read step10, a refusal with no idcheck, said 'no RED rows read' and held it. The old Email_Connector
snippet then opened the new MxModelReflection select page (a Data Grid 2) from a reference selector: CE0620.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
WM1 rerun 5: Email.EmailNP_EmailTemplate (non-persistent owner) is an info row in idcheck; only RED rows were replayed,
so the app's data view and create activity lost it (CE1613). Score went 6/7 -> 7/7 with the previous fix.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
WM1 rerun 6: stepupdate's jar cleanup was refused (stale vendorlib) and the build repair ran without the flag, so the
old userlib pdfbox/fontbox 2.0.24 stayed next to vendorlib 3.0.7 and javac failed (Misc.java:643). With the flag the
7 SHADOWED jars resolve as leftovers of the updated modules.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…eline's microflow steps

WM1 rerun 7: Atlas_Core 4 brings Atlas_Core.Administrator/User; its layout's language selector calls
DS_Account_CurrentUser, allowed for those only. Nothing mapped them, mx check stayed clean, and the smoke walk failed
Administration.Account_Overview 'for security reasons'. modroles.py maps roles NEW in this run, of an updated module,
by name to the user roles; the guest role is never granted (GUEST line), no name match is UNMATCHED.
The after walk skipped microflow steps the baseline ran: 4 false REGRESSIONs.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
WM1 9.24 -> 10.24 gate: 4 open RED idcheck rows, all System (DataBrokerEntityMetadata, WorkflowActivity.
IsMigrationActivity): Mendix 10 drops them. FSM approved the same rows by hand after counting 0 rows. platform_drops.py
counts rows / non-null values on the baseline snapshot and approves only on 0, with the count as the reason; any other
count or a failed query stays open (KEEP). App and Marketplace modules are never touched.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
WM1: counts_of('status 404 -> 200') read the HTTP codes as row counts 404 -> 200, so three new /rest-doc/ API docs of
updated modules were classed real-regression with 'rows or columns of the old data are missing'. They stay open
(likely-real: a person accepts a new surface or switches it off) with a reason that says what they are.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…r.txt

awk on a missing ledger exits 2 and set -e ended step_diff (7 dry-run controls: pipeline-diff-*, pipeline-ladder-*,
pipeline-dry-diff-gate-pass). A run without a baseline snapshot would have died the same way.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Stubs are prepended with an unquoted env PATH=$X:$PATH; VS Code's 'Application Support/.../copilot-chat' entry split
it and 11 controls failed exit 127 on a healthy toolkit. Full suite: 56 -> 35 failed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- one spelling of the temp folder (no '/T//', /private/var resolved): gate-config-relative-*, mxrun-*
- TZ=UTC: fixtures are stamped in UTC (journeys-archive-keep on a +08 machine)
- no bytecode under controls/ (controls-fixtures-not-ignored), BSD wc padding trimmed
- the timing fake clock answers date +%s without using a tick (b4ddb02 added the cockpit's running marker): timing-*,
  runlog-timing-columns, pipeline-autorunlog-ok-row
- GNU-only sed -i replaced by the existing sedi helper: runlog-backfill-*, release-impact-orphan-*

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…toolkit-pass)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… every role INVALID on --port 1)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- toolcheck-host-bsd-sed-catch: the nogsed PATH was made but never used; a host with Homebrew gsed never warned
- usercode-no-package-catch: /bin/false does not exist on macOS
- jarid-apply-old-sbom-refused: BSD touch has no -d DATE; touch -t, and a python utime for '+1 hour'
- remedy-apply-release-todo-pass: the release rung's number moved when the gate gained a rung

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… to SIGPIPE)

grep -q quits at the first match; unzip then died of SIGPIPE, pipefail made the check false, and deployment/ (the build
output, 858 MB on DOCGEN) was unpacked after all, with no note. bt9-fetch-* failed 1 run in 3; after: 8 of 8. The
note control now shows the fetch log when it fails.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- replay skip: git log | grep -qxF could lose to SIGPIPE on a long log, and a module already replayed was replayed again
- journeys stop: grep | grep -vq under '!' could read a SIGPIPE as 'only SCREEN REVIEW lines' and stop 12 (a comparison
  job) for a run that had FAIL lines; it is a count now

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
editroute approves a vendor drop when nothing outside the module refers to it. No references is not no data: WM1's
MendixSSO.AuthRequest.Nonce held 5 values, the ID check let it through, only the data rung saw the loss. Each such
approval is now counted: 0 stays, data or a failed count comments the line out (HELD) and the gate stops for a person.
WM1 baseline dry run: 'HELD MendixSSO.AuthRequest.Nonce: 5 non-null values in upg_run_base'.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The replay's own changes are undone (the tree was clean before it; untracked files it made are removed, paths relative to
the app), the module goes to replay-held.txt, and the build, journeys and gate still run and give evidence. The gate
stops 13 'upgrade complete except held replays' until a person fixes the replay or acknowledges it in
replay-held-ack.txt. A dirty tree before the replay keeps the old immediate stop: nothing can be undone safely.
5 controls (replay_hold_control.sh): 4 catch the old pipeline.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
mxcli's update gives a module's documents new unit ids but keeps the identities of its entities, attributes and
associations, which the tables hang on. WM1: 28 of 28 MODULE_RECREATED lines had no GUID change, yet each was a warn
saying the module was deleted and imported. Now: info with the count kept; warn (NS-7) when an identity changed or the
module has none to compare. Fixture reid_module.py --guids makes the real delete and re-import.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
datamove.py check: per 'Src.Entity.Attr -> Dst.Entity.Attr' line in datamove.txt, the non-null values in the target
after must be at least those in the source on the baseline. MOVED / LOST / UNCOUNTED; exit 1 unless all MOVED.
Real DBs (read only): drt_base 2 nonces, drt_c (vendor dropped the column) count fails -> UNCOUNTED, refused.
Not wired into the pipeline yet: where the move runs (a deploy before the update, or SQL) is still open.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ALTER PAGES SET LAYOUT skips pages inside Marketplace modules, so they keep the vendor layout. WM1 gate 2026-11-17: the
Accounts menu opened Administration.Account_Overview on plain Atlas 4 (sidebar and header gone, the mx logo over the
title; screens REVIEW). nav_copies: per Marketplace page on the redirected layout that a navigation profile opens, copy it
into the module where the app already keeps copies of that vendor module's pages (same title; WM1:
DISW_DesignSystem_Extention.MyAccount_Edit), on the company layout, grants mapped by role name, the menu pointed at the
copy. No such module: HELD for a person. Tried on a copy of WM1's 10.24 app: check passed, exec created the page and the
grant, the navigation differs from before only in the Accounts line.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ne grouped round

WM1 cold test 2026-11-17: 16 judgment stops, one question each, most answered with the toolkit's recommendation.
Interview mode auto (UPG_INTERVIEW_MODE=auto or interview-mode.txt) takes the recommendations at propose like unattended,
and parks the stops nothing later depends on (journeys, idcheck, smoke walk, diff, gate: nj_park) in parked/<step>.md;
the run goes on. At the end bin/parked.py writes one NEEDS-JUDGMENT.md, grouped (modules and data, journeys, screens and
gate) plus the ASSUMED proposal lines, and the run exits with the first parked code. Build or start failures, a missing
plan or a replay that cannot be undone still stop at once. A step that runs again parks afresh; a run with a parked
item never writes the PASS summary. Rung 2 inherits the mode and lists rung 1's parked items too.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
WM1 rerun 2026-11-17: the Account_Overview copy was HELD because the container mxcli (78bdfa5) refuses the
APP TITLE/APP ICON clauses the local mxcli's describe of the navigation prints. An omitted clause keeps the stored
value, so the copy drops them; checked with the container mxcli against the WM1 model: Check passed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… stand-in and its proof

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MendixMau and others added 28 commits November 17, 2026 11:35
…tgres route, not a hand stand-in

The pipeline already reads data through UPG_PSQL (pgdata.js + pgmap.py), which works on the 9.24 database as on
10.24 (WM1 rerun: 97 entities, 111 associations mapped). The hand OQL-to-SQL stand-in from the CLI test covers only
the queries typed into it; outside the pipeline use the same route.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…gin are noted, not FAIL

WM1 rerun 2026-11-17: the gate card said Data compare: Failed on 4 rows, none a loss of the app's data:
MendixSSO 4.2 -> 10.6 dropped AuthRequest.Nonce, MxModelReflection 7 -> 9 rewrote its Module rows, and
System.User.changedDate/Password changed on MxAdmin alone (runapp.py create_admin_user at every start).
- diff.py --module-jumps: a column gone or values changed in an attribute of a module the run updated is PASS*
  with the jump named; rows lost stay FAIL in any module; twin-old evidence still wins over this reason
- snapshot.py records System.User figures without the local admin row (UPG_ADMIN_USER, default MxAdmin);
  diff.py notes a change on that row alone as the toolkit's own login
- gate.sh and heal pass module-jumps.txt when the run has one
On the WM1 databases (fresh snapshot of a copy of the frozen db vs after_r5, --db-only): FAIL 7 -> 0, PASS 355.
Controls: 5 new (3 fail on the old diff.py, 2 guards hold on both), diff group 31/31.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…snapshot.py, not 5 synthetic rows per entity

Cold run 2026-11-17 (Marketplace template, 9.24.2): the package carried 9,235 demo rows in 139 tables and
bin/seed_snapshot.py already existed (steps/baseline.md item 1 prescribes it), yet fetch said "hsql data present,
not replayed (no tool)" and baseline seeded seed_pg.py rows. Walks and journeys then judge near-empty pages.
- seed_demo: replays data-snapshot.zip into X after the first start; nothing loaded (UNMAPPED) -> seed_pg.py as before,
  with a note; the file contents go to WORK/demo-files
- runapp.py copies UPG_DATA_FILES into data/files at every start (runapp wipes the run folder)
- fetch no longer prints the TODO when a data-snapshot.zip is there
Controls: 4 new; 3 fail on the old pipeline (1 guard, no snapshot -> seed_pg.py, holds on both).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…low steps are walked whenever the plan has them

Cold run 2026-11-17 (Marketplace template, 9.24.2, Atlas 3), baseline walk on the OLD app:
- "Switch user" FAILed in all 5 roles ("still on Dashboard", 15 s each) while the screenshot shows the Select Demo User
  popup open: Atlas styles .mx-window position:fixed, whose offsetParent is always null. landedWhere now also accepts
  a window with client rects
- 13 of 23 steps were microflows and SKIPPED; two roles were judged on no page at all. The "SKIPPED microflow steps"
  item fired only when the walk was INVALID (exit > 1), so with exit 1 nothing said so. smoke_mf now passes
  --microflows whenever the plan has a microflow step (each walk runs on its own DB copy); UPG_SMOKE_MICROFLOWS=0
  turns it off and leaves a note
Controls: 3 new, 2 rewritten (awk extraction of the multi-line smoke_mf); 3 fail on the old code (popup, plan
microflow, opt-out), the guards hold on both. Group: 31/31.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…achine has none

Cold run 2026-11-17: on a machine with no git user.email, stepupdate.sh's plain `git commit` stopped the run in
update (post-convert) with "Author identity unknown" (exit 4). The pipeline's own commits already passed -c user.*;
13 step scripts commit with plain git. git_identity exports GIT_AUTHOR_*/GIT_COMMITTER_* (upgrade-pipeline) only when
no identity is configured; a configured one is kept.
Controls: 2 new; the no-identity one fails on the old pipeline, the kept-identity guard holds on both.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…b-prefix

The kept name was written only when no db-prefix existed. A rerun with --db <new> worked on the new database,
but the next resume without --db silently dropped back to the old one. Now an explicit --db is written over the
kept name; without --db the kept name stays. Rung 2 has its own work folder, so its --db <X>_r2 does not touch
rung 1's prefix.

Controls: pipeline-db-arg-persisted-catch (fails on the old code: prefix stays upg_old) and
pipeline-db-kept-no-arg-pass (holds on both).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…outine work; journeys skip old-app failures

- cockpit drawer at half the screen; each Triage action shows "What it means" and the findings item by item
  (outside calls, module edits, local-edit decisions, React moves, convert, release rows), from the run's own files
- Test stage: what ran on the old and the new app, in plain words, with the per-test list
- proposal --unattended: routine work (section 2) is approved by the agent (source routine); only real decisions wait
- run_journeys --skip-baseline-fail (pipeline default): a journey that failed on the old app is SKIPPED, not walked again
- heal: both-fail is test setup; toolkit-gaps.md names the rule owner; diff/logdiff accept known platform differences
- controls: found_means, test_runs, skip_baseline and the heal gaps control

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…enshot, why, options with the catch)

- work/cases/<id>.json, written by the step that raises the decision; screenshots embedded, so the cockpit stays one file
- Resolve opens a wide dialog with the case; each option card says what it does and its catch; no case file: the plain question
- steps/heal.md: ask a person only on a deviation, and how to write the case
- control cockpit-case-file-pass

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… place when a copy is not safe

- No app module that holds copies, or a copy that would drop a widget mxcli cannot write (WM1 Accounts: the user grid):
  ALTER PAGE <page> { SET Layout = <company> } instead of holding it for a person. Only the layout line changes, every
  widget stays, the menu keeps the page. Checked on WM1's model: mxcli check passes, describe differs in one line
- mxcli check refusing the in-place change: still held for a person
- controls: no-home and unwritable-widget now set in place; a refused in-place change is held

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…o the seed_snapshot controls run on a Mac

pgtmp.sh looked only under /usr/lib/postgresql/*/bin. On a Mac with Homebrew Postgres every seed-snapshot Postgres
control SKIPped, and a suite reading 0 failed said nothing about the snapshot loader. It now falls back to the
initdb on PATH.

Control: pgtmp-path-binaries-catch (exit 3 on the old script, a cluster starts and stops on the new one).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…mapped as an object id

A Mendix 9.24 demo package's snapshot holds <entity>$pk tables for external (OData) entities; their "_id" is the
remote key, a Long whose high bits look like an entity short id. seed_snapshot tried to remap it, found no entity
for those bits, listed 5 tables as UNMAPPED and loaded nothing, so the baseline fell back to 5 synthetic rows per
entity. Columns other than id of a table whose entity has remote_primary_key are now kept as they are.

Controls: seed-snapshot-remote-key-catch and seed-snapshot-remote-key-kept-pass (both FAIL on the old code: exit 1
UNMAPPED, no row loaded; the id is remapped to the database's short id and _id is kept on the new code).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… committed, not kept as UNSAVED

The edit router forced Administration through --force-ui for its association replay (XASSOC). Its update had to
take the --no-baseline route, which saved 0 files, and stepupdate KEPT it as UNSAVED. But localedits listed no UI
edit in Administration, so nothing was lost. With Administration kept, MendixSSO 10.6.2 moved on without it and mx
check stopped the build on 2 CE0117 (repair row 18). The pipeline also never passed its localedits report to
stepupdate. Now it does, and a forced module whose localedits entry has ui [] is committed with a note. A module
whose localedits list UI edits is still kept, as before.

Controls: stepupdate-unsaved-no-ui-edits-catch and pipeline-dry-stepupdate-localedits-catch (both FAIL on the old
code), stepupdate-unsaved-ui-edits-listed-pass (holds on both).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…lists them

The text output of mx check names only the element: cold CMM stopped at build with four CE0117 "at Action button
'actionButton1'" (and 2, 4, 7), names that occur on many pages, so NEEDS-JUDGMENT gave no page to repair. mxcheck now
passes -j <work>/check-N.json after the mpr on mx 10 and later (mx 9 has no -j; before the mpr mx prints its usage, repair row 15), and the new bin/checkwhere.py appends one
"[where] CODE Module / Document: element" line per error location to the check log. The [error] lines and the summary
are unchanged, so checkdiff.py and the error count read the log as before. The stop report gets "Where they are".
In CMM it named Page 'Mix_New_WizardStep1' (desktop and phone): the buttons' dynamic classes expressions. repair.md
row 27 records that case (an association test that 10.24 rejects in dynamic classes only; remedy not verified).

Controls: checkwhere-document-catch, pipeline-mxcheck-where-catch and pipeline-dry-check-json-catch fail on the old
code; checkwhere-checkdiff-count-pass and pipeline-mxcheck-9-no-json-pass guard the count and the mx 9 call.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Cold CMM: login-MaterialHandler and login-MaterialManager FAILed on the healthy 9.24 app: the home page
Dashboard_MaterialHandler rendered in full, but its guard ".mx-name-scrollContainer1" never appeared. landing_widget
took the scroll container (inside the page's data view; the data view itself was skipped because the layout also has
a dataView1) as the outermost content widget. A Forms$ScrollContainer now counts as NON_RENDERING: it is never a
guard, so the layout grid inside it is taken, a page with nothing else stays a todo, and a kept impact.json that
landed on one is re-picked because guard_names leaves it out.

Controls (fixture models/guard6): gen-journeys-guard-scrollcontainer-catch, -only-catch and -names-catch fail on the
old code.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…onclusion and where it landed

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…s get a case automatically

- proposal.py: a module with local edits in localedits.json moves from the routine "modules" step to its own
  "module-edits" step; auto mode approves it with source carry-over (steps.CARRY), not routine
- cockpitmatrix.py: found detail for module-edits; cases() builds a case for a regression without a case file
  (auto_case: counts, old and new shot of the failing step, carried screen edits, evidence)
- controls: proposal-steps-module-edits-carried-pass, cockpit-auto-case-pass
- heal.md: the three cases; feedback ledger rows updated

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Cold CMM: DataWidgets was kept at its version and got a plan card "themesource
edits" asking a person to restyle the hero image, for one change: the package
holds themesource/datawidgets/.DS_Store and the app does not. themefiles.py
now skips .DS_Store, ._* (AppleDouble), Thumbs.db and desktop.ini on both
sides; a deleted real package file is still FILES-EDITED.

Controls: localedits-os-junk-files-pass, localedits-os-junk-real-removed-catch
(red on the old code: 2 failed; green: 2 passed).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Cold CMM: every cockpit render said "the page contradicts the run" for the
"Release notes (check)" and "<M> themesource edits (theme)" cards
(ask-no-evidence), because evidence.py only looked up module names. Each kind
now has its own source: release-impact.out for the release-note check, the
file rows of the module in localedits.json for a theme card. cockpitpage
passes the plan item's kind.

Controls: evidence-check-release-notes-catch, evidence-theme-files-catch,
evidence-theme-other-module-pass, cockpitpage-evidence-passes-kind-pass
(red on the old code: 4 failed; green: 4 passed).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Cold CMM: the 4 CE0117 at the dynamic classes of Mix_New_WizardStep1 vanished
in run 5 while 3 CE1613 were open, and came back unchanged in run 7 once
those were repaired (the pages are byte-identical in both runs). So mx check
reports no CE0117 while CE1613 errors stand: a check that lost its CE0117
next to new CE1613 has not fixed them.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…e update

Cold CMM: Administration 2.1.2 -> 4.5.0 dropped UserRole_Select and
RetrieveTimeZones, which two own employee pages used. Nothing said so
until mx check showed 3 CE1613 at the build. crosscalls only judged
kept callers, and the pipeline never ran it.

- crosscalls.py --own: the app's own modules are callers too, judged
  against the plan version package at any major (OWN / GONE lines, exit 1)
- the update step writes owngone.out before the updates and notes it;
  the build stop quotes it when mx check has CE1613
- steps/repair.md row 28: the remedy (copy the old document, or keep
  the module); a judgment item, mxcli cannot author selectors
- 7 controls: red 5 failed on the old code, green 21/21 with the group

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Cold CMM: mx check died loading the model (KeyNotFoundException from a copied
page with a stale pointer) and printed no summary. open_errors counted 0, so
the run went on to mxbuild, which failed on the same load. All 227 finished
checks in the lab logs print "The app contains: N errors."; the 6 without it
are all crashes.

- mxcheck appends "[error] [MXCHECK] mx check ended without a result" with
  the first ERROR line when the summary is missing; the build stops on it
- steps/repair.md row 29: what it means and how to undo the last change
- 2 controls: red 1 failed on the old code, green with the group

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…own module

Repair row 28 as a tool. Cold CMM: Administration 4.5.0 dropped
UserRole_Select and RetrieveTimeZones, used by two own Employee pages
(3 CE1613). mxcli cannot author or repoint those selectors. copydoc.py copies
the document from the same-version model before the update into an own
module, with fresh ids (pointers follow), its module roles mapped (--roles),
and repoints only own users (--retarget). Run 7: the 3 CE1613 gone.

The first version renewed the ids but not ControlBar.DefaultButtonPointer;
mx check died on load. The apply control checks the pointer: it fails on
that version.

- refusals: different Mendix versions, document or module missing, a
  Marketplace target, a dependency the app lacks, roles not given, MPR v2
- 7 controls: red 7 failed without the tool, 1 failed on the first version,
  green with the group; row 28 names the tool and the verified run

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ity too

Cold CMM run 8: after repair rows 27 and 28 mx check showed 5 CE2729, "No read
access to association 'Employee_Account' in entity 'Administration.Account'".
The XASSOC replay re-created EmployeeManagement.Employee_Account (Employee to
Administration.Account) and restored the owner's grants only; the original
model gave Account ReadOnly on it for Administrator and User, after the update
it was None.

collect now also reads the other entity's grants from the original model and
writes each one cut down to the association member (member_grant). mxcli's
grant merges members into the role's rule, so the vendor's own rights stay.

Control editroute-collect-assoc-other-side-catch: fails on the old code (1
failed, 11 passed), passes on the new. Repair row 27 rewritten with the real
cause ($currentObject in a nested data view is the inner object; $dataView1,
verified run 8); row 30 for CE2729 (verified run 9: 0 errors).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…acjar.py)

Cold CMM run 9, 10.24.28: mxbuild stopped in javac, "cannot find symbol
CookieHeaderNames.SameSite" in MendixSSO 10.6.2's OpenIDUtils. vendorlib has
netty-codec-http 4.2.17 with it; userlib/server-examples-0.4.1-SNAPSHOT.jar,
a fat jar, carries its own older Netty, and javac took that copy. jarclash
listed the clash as a note; the stop named only javafinal.py.

bin/repair/javacjar.py resolves each missing symbol's class through the
source file's imports, looks it up in every userlib/vendorlib jar (nested
too) and prints HAS / HIDDEN BY. --strip [--apply] writes the fat jar again
without the classes a vendorlib jar also carries. The mxbuild stop runs it
on "cannot find symbol" and quotes it when a class is hidden. Repair row 31.

Controls javacjar-hidden-catch, -not-hidden-pass, -strip-dry-pass,
-strip-apply-catch, pipeline-build-javacjar-catch, -not-hidden-pass:
5 failed on the old code, all pass on the new.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…te (widgetskew.py)

Cold CMM (9.24.2 -> 10.24.28): Atlas_Web_Content 3.9.0 replaced Datagrid 2.2.2 with
2.30.6 in widgets/, while DataWidgets, whose themesource carries the grid styles, was
kept (FILES-EDITED). Every grid page rendered unstyled rows; afterwalk passed 23/23
because the data was on the page, and the screens rung only said REVIEW.

bin/widgetskew.py joins .stepupdate/widgets.txt with kept.txt and the kept module's
cached target package: a widget that package ships and another step replaced is a
WIDGET-SKEW (exit 1). step_update parks on it (exit 13) unless widgetskew-ack.txt
accepts it. Repair row 32. Controls: widgetskew-catch, widgetskew-nothing-kept-pass,
pipeline-update-widgetskew-catch, pipeline-update-widgetskew-ack-pass.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@MendixMau

Copy link
Copy Markdown
Owner Author

CI: this PR makes the controls run on GitHub for the first time (main has failed at setup since #10: no bson, no zsh). Result on the Linux runner: 2624 passed, 10 failed, 67 skipped. The 10 (smokewalk-landedok, smoke-users-bad-entity, six stepupdate, remedy-keptdeps-file, cockpit-decide-cards) pass on macOS, where the full suite is 2648 passed, 0 failed. They are an open row in docs/feedback-ledger.md. Merging on the owner's instruction.

@MendixMau
MendixMau merged commit d64c7f2 into main Nov 17, 2026
1 check failed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant