Repository navigation
learn(preflight, detection gaps, ledger): retire/shrink learned rules that current mxcli fixed (retest on v0.32 + MxBuild 11.8) - #824
Conversation
… that current mxcli fixed (retest on v0.32 + MxBuild 11.8) A retest of 108 learned rules and ledger entries on mxcli v0.32.0-374 against real MxBuild 11.8.0 (rungs 1-4) found 55 fixed, 19 changed and 1 misdiagnosed. - Fixed/changed rules move word for word to skills/learned-rules-archive.md (on demand, routed; still valid on an older mxcli). Live rows state the current behaviour with a version guard. - Wrong rules corrected: BUG-17 (not Mendix tokens), STOP #12 and STOP #1's ContentParams clause, STOP #32 (only $currentObject is refused), the DG2 column gotcha/register rows (refused at check now), microflow header text. - Ledger: 62 entries stamped Fixed/Changed/Still/Not reproduced; 6 new BUG-DRAFT candidates; retest record with every probe inlined in bug-logs/mxcli-v0.32.0-retest-2026-11-20.md. - #29/#30 kept (not reproduced on Mendix 11.8 only). Baseline 65,568 -> 64,737 words; skills/generated 38,589 -> 37,601. Co-Authored-By: Claude Opus 5.5 <[email protected]> Claude-Session: https://claude.ai/code/session_01HXpQSjXefSNicGGqP4aiSC
|
Review: MERGE-AFTER-FIX — STOP #12's "FIXED" claim is wrong for its own lead example:
Generated by Claude Code |
… one version guard; gate reads the archive - On v0.32.0-374 a ContentParams expression builds only when every attribute is written $currentObject/Attr. A bare attribute inside a function, toString(Status), still passes check and exec and fails mx check CE0117 (re-probed: data view and customContent, bare vs $currentObject/, real MxBuild 11.8.0). Corrected in the STOP #12 row, generated rules, archive, BUG-142 stamp, customcontent-binding stamp, STOP-12 scan message, retest record and CHANGELOG; a live detection-gaps row restored for the case. - Version guard defined once (archive header, preflight version line, record): v0.32.0-374 is 374 commits after the v0.32.0 release; v0.32.0 counts as older. Detection-gaps rows use the same wording. - learned-rules-archive routed to gate too (re-rendered). Baseline 65,568 -> 64,862 words; skills/generated 38,589 -> 37,930. Co-Authored-By: Claude Opus 5.5 <[email protected]> Claude-Session: https://claude.ai/code/session_01HXpQSjXefSNicGGqP4aiSC
|
Review: MERGE — re-review of head 4fed6f5 (fix commit after MERGE-AFTER-FIX): all three fixes landed at every listed location; STOP #12 now CHANGED with the
These now match the text at Generated by Claude Code |
…ams; exec callees before checking the workflow Behaviour A/B (6 tasks x 4 runs per arm, real mxcli v0.32.0-374 + mx check 11.8.0): both arms 24/24, but NEW caused more failed check rounds in two spots. - STOP #12: formatDecimal($currentObject/Amount, '0.00') passes check and fails CE0117 (4/4 NEW T1 runs); use toString($currentObject/Attr). - STOP #18: keep "exec the callee microflows before check --references of the workflow script" for every version (4/4 NEW T5 runs hit "not found"). Baseline 64,862 -> 64,892 words; skills/generated 37,930 -> 37,960. Co-Authored-By: Claude Opus 5.5 <[email protected]> Claude-Session: https://claude.ai/code/session_01HXpQSjXefSNicGGqP4aiSC
|
Review: MERGE — head ba82605 (addendum below); earlier re-review of head 4fed6f5 (fix commit after MERGE-AFTER-FIX): all three fixes landed at every listed location; STOP #12 now CHANGED with the
These now match the text at
Generated by Claude Code |
Direct lane
What changed and why (one paragraph):
A retest of 108 learned rules and ledger entries on mxcli
v0.32.0-374-g591bcd73e(source build) against real MxBuildmx check11.8.0 (rungs 1–4:check,exec,describe/BSON read-back,mx check; every probe on a fresh blank Mendix 11.8.0 app) found 52 fixed, 22 changed, 1 misdiagnosed (after the review's correction of three ContentParams rows). Version guard, defined once:v0.32.0-374is the build 374 commits after the v0.32.0 release; v0.32.0 itself and every earlier release count as older. This PR is the argument the word-budget rule asks for before protected text goes (CLAUDE.md → "Always-loaded skill text", rule 1): no rule text is deleted. Every fixed or changed rule moves word for word to the new on-demandskills/learned-rules-archive.md(routed,mdl,gate,review, stages 5–6), because projects pin older binaries. The live rows now state the current behaviour in one line with a version guard (v0.32.0-374 refuses it at check/on mxcli < v0.32.0-374 …). Five rules were wrong on the current binary and are corrected: BUG-17 (misdiagnosis), STOP #12 plus STOP #1's ContentParams clause (on v0.32.0-374 an expression builds when every attribute is$currentObject/Attr; a bare attribute inside a function,toString(Status), still passescheck/execand fails CE0117 — so the row now teaches the$currentObject/form, and a live detection-gaps row covers the bare case), STOP #32 (over-broad), the DG2-column gotcha and register rows (wrong failure mode), and the microflow-header error text. The ledger gets aFixed in/Changed in/Still reproduces on/Not reproduced on v0.32.0-374stamp on 62 entries (none deleted;bin/bug-lookup.shfinds them as before, and BUG-17's heading now says SUPERSEDED, so--indexflags it closed), plus 6 newBUG-DRAFT-*candidates with inline MDL and verbatim errors. The full verdict tables and all 140 probes are inbug-logs/mxcli-v0.32.0-retest-2026-11-20.md. Two other files that made the same claims were reconciled (retesting-learned-rules.mdcompletion criterion):learned-datagrid-customcontent-binding.md(RETIRED stamp) andlearned-workflow-patterns.md§8 Warning 1 (a v0.32.0-374 row). TheSTOP-12mdl-scan rule stayswarnbut its message now says the CE1613 is for mxcli < v0.32.0-374, and it is no longer listed as ablockcandidate.Kept on purpose: #29 and #30 did not reproduce, but only on Mendix 11.8; they were measured on 11.12–11.13, so they stay with a note. STOP #1's main rule stays:
mx checkis 0 errors, but its symptom was on Studio Pro open (not available here).PARALLEL SPLIT,calculated by, andrest call bodyare fixed statically only, so the runtime caveats stay. TheDECISIONline now says what builds, but BUG-141 (restart safety) is still open.Evidence (mxcli v0.32.0-374, Mendix 11.8.0, 2026-11-20)
[%BeginOfToday%]quoting✗ [%BeginOfToday%] is not a Mendix token (mxbuild: CE0161 "Error(s) in XPath constraint.") [MDL-TOKEN01];[%BeginOfCurrentDay%]/[%EndOfCurrentDay%]in the samewhere:mx check0 errorstoString($currentObject/Status),$currentObject/Name,if $currentObject/… then … else …:mx check0 errors.toString(Status)(bare attribute in a function; data view, customContent, list spelling):Check passed!,Created page Shop.F12a_Bare, thenCE0117 | Page 'F12a_Bare' | Expression of text template parameter {1} of text 'txtA' | Error(s) in expression.(Order = $Order):Check passed!,mx check0 errors; only$currentObject:… cannot be stored — widget btn1 is not inside a data view … [MDL-PAGEARG01]Reference error: alter page Shop.M11_Page: the insertion point (Qty) is in no data container, so no entity is in scope …/Refusing to execute … Nothing was written.()line 3:36 a microflow header needs its parameter list in parentheses, even when it is emptyDECISION(BUG-76/108)decision outcome 'Draft' is not fully qualified … [MDL-WF03];'Shop.OrderStatus.Draft'+'':Created workflow,mx check0 errors[%CurrentUser/…%]✗ [%CurrentUser/Administration.Account_Language%] is not a Mendix token (mxbuild: CE0161 …) [MDL-TOKEN01]Created microflow: Shop.WF18_Callee/Created workflow: Shop.WF18,mx check0 errors (was CE0117)⚠ applying this script to the project removes 2 member(s) from entity Shop.Line … [MDL087](atcheck)✗ … sets design property "Form orientation", which is not defined for this widget type — mxbuild reports CE6083 … [MDL-WIDGET11]widget tb1 … exec stores it as Shop.Line_Order/Shop.Order_Customer/Name … [MDL-ATTRPATH01],mx check0 errorsOrder_Customer is an association: an expression names it qualified, $Order/Shop.Order_Customer/… — mxbuild rejects the expression with CE0117elseifin page expressionline 5:93 missing ELSE at 'elseif'modifyAltered enumeration: Shop.G03_Enum= nullthe constraint names "null", which is neither an attribute nor an association of it — mxbuild rejects the constraint (CE0161 …)Context/New/Casereservedattribute 'Context' is a reserved word (CE7247) [MDL021],enumeration value 'New' … [MDL010];Caseentity:Created entity: Shop.CasethenCE7247 The name 'Case' is a reserved word.Created microflow: Shop.MP11_Discard,mx check0 errors✗ Attribute Note of Shop.Order is a String, but the value assigned to it is an Enumeration … [E009]; exec stillCreated microflow,mx checkCE0117 at the change and at thereturn(unflagged)ExposedToClient: false/OptimizedClient: 'Yes',mx check0 errors — Mendix 11.8 onlyThe full 108-row table, every probe's MDL and read-back command:
bug-logs/mxcli-v0.32.0-retest-2026-11-20.md.Impact
bin/lib/skill-routing.tsv(+1 row:learned-rules-archive, ondemand,mdl,gate,review, stages 5,6) and its re-render (README.md, ROUTING.md,agents/mdl-agent.md,agents/gate-agent.md,agents/review-agent.md,bin/gate-check.shstage map).bin/lib/mdl-scan-rules.tsv: STOP-12 message and header comment text only (severity unchanged,warn).skills/generated/*re-rendered bybin/mdl-rules.sh --render.bug-logs/mxcli-bugs.mdis edited by many PRs: conflicts are possible but local (stamp lines under headings, new entries at the end). No new tools, no new artifact reads.bin/render-routing.sh --check: Baseline 65,568 → 64,892 words (−676), budget 80,000 (first push: 64,737; the review fixes added 125 and the A/B fixes 30). Stage 5: total 60,263 → 59,587; gate 21,747 → 21,328; review 30,445 → 30,026; mdl 14,708 → 14,743. Stage 6: total 49,933 → 49,549; gate 20,764 → 20,345; review 24,977 → 24,558; mdl 9,603 → 9,638. Preloadedskills/generated/rules-*.md: 38,589 → 37,960 (−629). The retest report estimated ~1,400 baseline words; the difference is the rules kept on purpose (above), the STOP #12 case that turned out still live, and microflow-patterns, which is not baseline.mx check11.8.0. Primary score 24/24 OLD and 24/24 NEW, cost per run $0.457 vs $0.454. NEW removes forced workarounds: T1 ContentParams expression used 0/4 OLD vs 4/4 NEW; T3 extra nested data views for a 2-hop binding 4/4 OLD vs 1/4 NEW; OLD's quoted-DynamicClasses spelling drew 2 MDL-WIDGET33 refusals, NEW 0. OLD's DECISION and #32 STOPs stopped nobody. NEW cost more failed check rounds (25 vs 6) in two places: T1, 4/4 wroteformatDecimal($currentObject/Amount, '0.00')and needed 2–3mx checkrounds for CE0117 (1–2 rounds on 4fed6f5, n=2); T5, 4/4 checked the workflow before exec'ing its callees (not found). Both are now addressed in ba82605: STOP #12 names the format-string trap, STOP #18 keeps the exec-callees-first sequencing for every version. Headline: T1–T6 pass OLD 4/4 each, NEW 4/4 each (All 24/24 vs 24/24; NEW2 2/2); retries OLD 6, NEW 25 (T1 3→14, T5 0→6, T6 3→5). Caveats: ceiling effect (both 100%), n=4 so rework differences are directional; the fixes in ba82605 were not re-run (account usage cap).Time cost: none · n/a · can force rework: n · offset: removes detours (extra String attributes for ContentParams, an extra microflow for show-page params, split workflow scripts) that the old rows forced on v0.32 projectsField evidence — the retest itself: mxcli
v0.32.0-374-g591bcd73eagainst real MxBuild 11.8.0mx check, four probe apps built from one blank Mendix 11.8.0 app, 140 probe scripts (Appendix B of the record), each attributed by unique document name. Not available: Studio Pro's loader (rung 5) and the runtime (rung 6) — rules whose failure lived there were kept.Fixtures (scoped; master → branch, same worktree base
1a56c3a1; re-run after the review fixes with the same counts):tests/wave2/test-bug-lookup.sh bin/bug-lookup.sh: 28/0 → 28/0 (ranker floors unchanged: dev top-1 89%, held-out top-1 92%, top-3 100%)tests/wave2/test-mdl-rules.sh bin/mdl-rules.sh: 43/0 → 43/0tests/wave2/test-mdl-scan.sh: 53/0 → 53/0 (P: the preflight still marks exactly STOP-1/12/16/26)tests/wave2/bench-mdl-scan.sh: rc 0 → rc 0, same outputtests/wave2/test-context-pack.sh: 59/0 → 59/0tests/wave2/test-runbook-slice.sh bin/runbook-slice.sh(branch): 31/0by inspection:
test-file-bugs.sh's grep (only a human filesin detection gaps) andtest-exec-lock-watch.sh's ledger grep (alter page … { set Url = '…'; }) still matchbin/mdl-rules.sh --check: clean;bin/render-routing.sh --check: in sync;bin/check-scripts.sh: 145/145, 30/30;bin/check-portability.sh: clean (410 files); leak guard: clean;bin/check-no-private-citations.sh: clean;bin/check-pr-discipline.sh: cleanGrep proof that the wrong text is gone and the corrections landed (master → branch): STOP Cloud-dev skill, node-express-react layout config, and nine instrument fixes from one end-to-end migration #1
Do NOT use ContentParams: [{1} = toString("State")]1 → 0; unguarded STOP docs + doctor: "Where you run this" visual, Windows onboarding fixes, toolkit.env #12 ban 1 → 0 (now verbatim in the archive); STOP Four instruments that were installed and not working, and the obligations to notice #32 "any argument outside a data view" 1 → 0; DG2 gotcha "REPLACE … columnpassescheck, fails mxbuild" 1 → 0; register "Check passed!, exec applies it" 1 → 0; header "check errors point at returns/List of" 1 → 0 (now version-guarded); BUG-17MISDIAGNOSED0 → 1; ledger stamps 0 → 62; newBUG-DRAFT-*0 → 6No client data anywhere in the diff (leak guard clean; probe names are
Shop.*/ModB.*)Size cap: exceeded, on purpose — 22 files, +2,536/−125. 1,835 lines are the captured retest record (probes plus verdict tables), 279 are the archive (moved text, verbatim), 288 are ledger stamps and new entries, and 10 files are mechanical renders (routing,
skills/generated/). The hand-written diff to the live skill files is about 80 lines. Splitting would put the archive and the rows it replaces in different PRs, and the rows would cite a file that does not exist yet.Test tier: T1/T2 — scoped fixtures plus the mechanical checks above; the claims themselves were verified at the real
mx checkgate (rung 4)For a new skill: routing row added (
bin/lib/skill-routing.tsv+bin/render-routing.sh)CHANGELOG.mdline appended, credited to the retest runNew bug entries headed
## BUG-DRAFT-<slug>:Not changed, and why:
Words that could be retiredestimate counted Re-probe BUG-76 and BUG-121 on mxcli v0.21.0 — both still open, and BUG-76's diagnosis was wrong #29/Generalise the popup-feedback rule to any task-completing microflow #30 (Mendix-version-dependent), STOP Cloud-dev skill, node-express-react layout config, and nine instrument fixes from one end-to-end migration #1's main rule (Studio Pro loader untested), and the workflowPersistentIdrow (NOT-TESTABLE); all three stay.rename moduleleaves XPath,[Assoc = 'literal']= BUG-78, constants print secrets). They got aStill reproduces on v0.32.0-374stamp, not a duplicate entry. The new entries are the other six. The DG2 "available columns list omits the association column" detail in the report is not in any saved output, so it was left out of the collision entry.learned-workflow-patterns.mdprose ("Never script aDECISIONat all") predates its own v0.24 table row and still reads absolute; only the version table got the v0.32.0-374 row. A rewrite of that section is a separate change.MXTK_BASELINE_BUDGET_WORDSis untouched.Review fix probe (real MxBuild 11.8.0, fresh blank app + base model): four pages —
toString(Status)andtoString($currentObject/Status), each in a data view and in a customContent column.checkpassed all four, exec created all four;mx check -j: CE0117 onF12a_BareandF12c_CCBareonly, 0 errors on the two$currentObject/pages.Review: MERGE-AFTER-FIX — fixes pushed, awaiting re-review
https://claude.ai/code/session_01HXpQSjXefSNicGGqP4aiSC
Generated by Claude Code