Repository navigation
feat(ui): logic reads as sentences; XPath behind "code"; advanced panel in Logic/Display/Messages/Raw (#19) - #29
Open
PrjShrestha wants to merge 13 commits into
Open
PrjShrestha wants to merge 13 commits into
PrjShrestha wants to merge 13 commits into
Conversation
…lected" reopen as clauses (#15) Two defects met at ruleToClause in conditionReducer.ts: (a) clauseToRule emitted `${f}` and `not(${f})` as kind 'raw', so on reopen the parser handed back a raw rule, hydrateColumn fell to rawFallback and every control was disabled. They now have a real parser kind, `truthy` { field, negated }, that serializes to exactly the two spellings the builder has always written. Spacing-divergent forms (`${ f }`, `not( ${f} )`) stay raw via the self-check. (b) `${f} != ''` / `${f} = ''` parsed cleanly as kind 'answered' and were dropped at the same spot. They now map to the ref / not clause and carry the author's spelling as Clause.source, which clauseToRule re-emits while it still describes the clause. An unedited reopen writes the same bytes; `${f} != ''` is never rewritten to `${f}`. Pinned separately: a bare `${f}` fixture exercises only the truthy path, `${f} != ''` only the answered path. All four reducer tests and the three positive parser round-trip tests fail on bff69cb and pass here. Three UI consumers that switch exhaustively on Rule['kind'] gained a branch for the new kind (modal row, decisions prose, calc prose). Verified: shared tests 792 pass / 0 fail; typecheck clean; corpus sweep output byte-identical to master; a cell-level serialize(parse(x)) check over 4034 relevant/constraint/choice_filter cells in seven configs shows the same 37 pre-existing drifts as master and none new; 21 relevant cells now open as clauses that were raw before. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…uilder (#15) Two Playwright specs on a throwaway copy of the mini-config fixture: - write `${lmp_date}` and `not(${danger_signs})` through the strip, save, assert the bytes via the API, reload, and reopen both rows as clauses (no "hand-written" status, undo-last-clause present, insert enabled); - the fixture's existing `${lmp_date} != ''` relevant opens as a clause, re-inserts with zero edits as the same bytes, and survives a save. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
… function forms on `.` (#14) Real validation rules are written against the answer itself and real relevants use relative paths; the shared parser only knew `${field}` subjects, so 774 of the 777 constraints in seven real configs opened as plain text. This slice is shared-only and additive. New in shared/src/xlsform/operand.ts: a closed operand grammar (`.`, `${f}` / `../f` with the spelling kept, literals, whitelisted ODK + CHT calls, `+ - * div mod`, parens). On top of it, four additive rule kinds: - expr-comparison `. >= 0`, `string-length(.) <= 100`, `int(format-date(today(),'%Y')) + 57 >= int(.)`, `. > max(coalesce(${a}, 0), ...)`, `. <= today() - 30` - predicate `regex(., '...')`, `selected(., 'none')`, `not(...)` of those - not-group `not(selected(., 'none') and count-selected(.) > 1)` - always-true `true`, `true()`, `1` (text carried, never rewritten) Each carries the clause verbatim as `source`; the serializer re-emits it while it still parses to the same rule, so `.<=100` and `. <= 100` both open AND save back byte-identical, and only a rule the author changed gets canonical spacing. `../field` on the existing comparison / selected / answered / truthy kinds is a `refSpelling: 'relative'` flag, re-emitted exactly as written in either direction. The reducer attaches `source` to any hydrated clause whose canonical emission would differ, so a `../field` rule opens in the inline strip and saves back unchanged. Also: the self-check now runs on all-raw chains too. Splitting on the combinator rejoined `a and b` with one space (six distinct real FCHV / LMP constraints); such a chain is now one raw rule, byte-identical. Consumers that switch exhaustively on Rule['kind'] show the new kinds as the text the author wrote (modal row, decisions / calc prose); a change in the modal turns the rule into a raw fragment, as before. Measured on the seven analysis configs (777 constraint cells): constraint 682 / 777 open fully structured (3 before; 51 are placeholders) relevant 1839 / 3050 drift 0 / 3919 cells (serialize(parse(x)) === x on every cell) `../` share: 410 relative refs vs 4433 ${} refs; 329 cells use `../` Tests: operand grammar; serializer-exercising round trips from non-canonical fixtures for every form in the ticket; `../` hostile fixtures live for both halves (byte identity and opens-as-clause); the all-raw self-check; an "additive" guard that every pre-T9a fixture still yields its old kind; Playwright: seeded `.` / `../` cells survive open-and-save byte-identical, the modal opens a `.` constraint as rows, a `../` relevant opens in the strip. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…s hidden; choice labels in pickers and readback (#16) The "show this question when" dropdown listed about 230 rows of a real pregnancy form by name in sheet order, every *_note, r_* summary row, __* hidden output and plumbing calculate included, with no search, no labels and no grouping (UX review finding 2, P0). One component now serves every place a rule picks a field: the inline strip, the "build" modal, the calculation builder's field references, and the strip's value cell when the value is another question. shared/src/xlsform/fieldMeta.ts (pure, node-tested): per named row, the resolved label (first non-empty in locale order), the innermost section (group label or name, both `begin group` spellings), the field kind, and whether the row is technical: note, r_*, __*, hidden type/appearance. A calculate is NOT technical: the harvest calculate that re-exports ../inputs/contact/sex is the sanctioned way to reach a contact value. client/src/ui/SurveyFieldPicker.tsx: [search] [select] [show technical rows]. The select stays native and keeps `ref-chip-select`, so keyboard behaviour and every existing e2e `selectOption` keep working; options read "Label (name)", grouped by section as optgroups. Typing narrows by label OR name and turns the select into a visible list, so matches show on the first keystrokes and one click picks; picking clears the search. Technical rows are withheld until the toggle. The current value is always kept in the list so a saved selection is never stranded. Metadata comes from a FieldMetaContext provided once per form by FormEditor; callers keep passing the dependency-ordered name list, so the picker never offers a field the caller withheld. FormEditor: the v0.3 "Typical for this check" / "Other fields" partition and its "Show all fields" checkbox are gone (the review found the box read as unchecked while everything was shown); op-typicality survives as ordering inside each section and never hides a field. The value cell shows choice labels ("Vaginal bleeding (vaginal_bleeding)") and the readback chips use labels; the written cell keeps the name. The free-text value cell no longer says `value or ${other_field}`: an "another question" toggle swaps it for the same picker and writes `${name}`. Found on the way: the search box grew on focus, so a mousedown on "+ insert" blurred it, shrank it, and moved the button out from under the cursor before mouseup; the click was lost. No width change on focus now. Tests: fieldMeta node tests (labels, sections, technical reasons, search); Playwright field-picker.spec.ts (search by name and by label, technical toggle, labels in value picker and readback with the name on disk, "another question" value, the modal uses the same picker, and on the real 280-row geriatric form "lmp" lists a handful grouped by section, skipped when that config is absent). The three v0.3 picker specs are rewritten for the section grouping and the technical toggle. Open-and- save is unaffected: corpus sweep output byte-identical to master. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…scroll to and highlight the new row (#17) Clicking a type tile in the add-question picker committed immediately: no step for required, hint or validation, and the new row was appended after the hidden __* outputs, off-screen, with no scroll and no highlight (UX review finding 3, P0). QuestionTypePicker: a `configure` step between a question tile and the commit — required, a hint per visible language, and a Validation slot (constraint expression with the existing "✎ build" modal, plus a constraint_message per language; 9e replaces the slot with presets). Enter anywhere commits. "add without details" is the one-click behaviour for this question; "Always skip this step" remembers it (localStorage `cht-ui-builder.oneClickTiles`). Structural tiles, the lineage sentinel, hidden rows and edit-type reopens never get the step. A select's choices step leads into the configure step the same way. PickerCommit gains an optional `details` block; FormEditor writes `required = yes`, `hint::<loc>`, `constraint`, `constraint_message::<loc>` only for the values the author set, so an untouched step adds exactly the row the one-click flow added. Insert position: "+ Question" lands directly after the row the author is on (the last focused card, tracked by a focus-capture on the survey tab), via the new shared `insertIndexAfterRow` (surveyEdits.ts, node-tested). A begin-group row counts as "inside the group" (first child), any other row is followed by its new sibling, so every pair stays balanced. With no current row it falls back to `defaultInsertIndex` (before the trailing plumbing calculates). After commit the new row is scrolled into view, focused and flashed for 2.5 s; a row hidden in Simple mode flips the editor to Full first. The Validation slot's field list is the same dependency-ordered, unique-name, non-plumbing list the row card uses (`pickableFieldsBefore`, now one function), computed for the insert position. Tests: surveyEdits.insertAfter.test.ts (after a top-level row, inside a group, after a begin row, trailing-plumbing fallback, unknown id; each asserting structural balance). Playwright add-question-configure.spec.ts: age (integer) with required, hint, constraint and message lands right after lmp_date, flashed and in view, with those cells on disk and every other row's cells unchanged; "add without details"; a question added while on a row inside a group lands inside, after it, balanced; the remembered preference commits on the tile click. Suite: the build specs were written against one-click tiles, so the Playwright profile seeds the preference ON (storageState in playwright.config.ts, which also covers specs that bypass setup.ts); the new spec clears it. demo-1's row-order expectations now reflect insert-after-current-row. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…ssages beside the rule; never normalises (#18) Validation is where a program says what a correct answer is, and the tool could show 3 of 777 real rules. With `.` readable (9a) most real rules are a dozen shapes a non-developer recognises by name. This slice gives them a panel and replaces both constraint builders. shared/src/validation/presets.ts (pure, node-tested): the preset model (between, compare-value, compare-field, compare-today, days-from-today, months-from-today, after-all-of, text-length, allowed-chars, pattern, int-value, bs-year, choice-alone, count-selected, always-true, code), `parseValidation` (one parsed rule → one preset; two adjacent numeric bounds → "between"; an `or` chain, a grouped or unparseable cell → one code item), `emitPreset` (canonical XPath), `presetsFor` (the menu per question kind; today() for dates, now() for date-times), `suggestMessage` and `presetComplete`. THE RECOGNISER NEVER NORMALISES. Every item keeps `source`, the rule text as written; `serializeValidation` re-emits it while it still reads as the same preset and writes the canonical spelling only for an item the author changed. `. <= 100 and . >= 70` is displayed as "Between 70 and 100" and saved as `. <= 100 and . >= 70`; edit the maximum and that item becomes `. >= 70 and . <= 99` while its siblings keep their spelling. relevantParser.ts: ParsedExpression gains an optional `separators` (the text between rules exactly as written — ` and\n`, ` and `), set only when a join is not canonical and honoured by serializeRelevant while it still fits the rule count and spells the combinator. Real configs break ~90 chains across a newline or a double space; they were raw, now they open. Additive: canonical cells and consumer-built expressions carry no separators. client/src/ui/ValidationPanel.tsx: a list of sentence-shaped preset rows with inputs, "+ Add rule" per question kind (plus a plain expression), a code toggle, constraint_message per visible language with a suggested text that stops as soon as the author writes their own, and (row editor) the required checkbox with required_message per language beside it. `true` / `true()` / `1` show "This rule always passes: no validation". Incomplete presets stay on screen without writing a broken expression; only complete items reach the cell. The rule and a suggested message are written in ONE row update (two updates in a tick each started from the same stale row and the second won). Mounted in the row editor in place of the constraint expression field (constraint_message inputs leave the hints block; required_message leaves the raw overrides), and in the add-question configure step's Validation slot (9d) in place of the expression box + modal. The inline strip no longer offers the constraint column. Measured on the seven analysis configs (777 constraint cells): 575 open entirely as presets, 61 as presets plus a plain-text item, 51 are placeholders now labelled, 90 stay plain text (mixed and/or inside not(), curly quotes, decimal-date-time arithmetic). Parser-level: 730 of 777 structured. Drift 0 of 777 through parseValidation → serializeValidation, and 0 of 3919 cells through the parser. A form holding every preset's canonical output (40 constraints) compiles with pyxform 4.5 xls2xform, the step cht-conf runs (Docker is not running on this machine; cht-conf itself is installed). Tests: presets.test.ts (every round trip from a non-canonical fixture through the serializer; edited-item canonicalisation; stale source ignored; every catalogue entry emits and parses back to itself; separators), relevantParser separators tests, Playwright validation-panel.spec.ts (age with Between 0 and 20 in the picker → sheet → reopened preset; reverse-order between, tight `.<=today()` and `true()` displayed as presets and saved byte-identical, then one edited item canonical; text length, date not in the future, select-many choice alone, required message, cells asserted on disk). Two earlier specs updated for the panel replacing the constraint box and modal. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…nd constraint match at runtime on a live CHT (#20) The June ticket (#9) had two acceptance criteria nobody had run. four-builders.spec.ts — acceptance 3: in one journey on a scratch copy of the fixture, the inline strip writes a relevant (`selected(${chair_rise}, 'pass')`) and a choice_filter (`selected(${danger_signs}, 'vaginal_bleeding')` on a later row — the picker only offers earlier fields), the Validation panel writes a constraint (`. >= 0 and . <= 20` with its message), and the calculation builder writes `${gravidity} + 1` on a fresh calculate row; every cell is read back through the API after a UI save. live-instance-check.spec.ts — acceptance 4: on a renamed copy of the fixture form, author by picking only (gravidity shows when chair_rise includes "pass"; accepts 0..20 with a message), set the sheet's form_id through the builder's API, deploy that one form with cht-conf 6.5.0 inside the cht-ui-builder image (`--add-host` so the TLS name resolves to the host), then as the CHW on the local CHT 5.2 instance: the question is hidden, shows after "Pass", rejects 25 with the authored message and refuses to submit (no report), accepts 10 and the report is in CouchDB with fields.gravidity = "10". The report and the form docs are removed afterwards; the record id is printed. Skipped when the instance or Docker is not reachable. Run on 2026-10-01 against the poc_demo CHT 5.2.0 instance at https://127-0-0-1.local-ip.medicmobile.org:10445 (NSSD config): report 01a0f7da-9c2b-766d-90aa-3aabfd8dd6be, fields.gravidity=10. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…el in Logic/Display/Messages/Raw (#19) T9f, the last slice of the T9 epic (#9). The row's advanced panel is now four groups. Each logic column is one sentence-shaped editor: "Show this question when …", "Filter the choice list when …" (select rows only), and the Validation panel for the constraint. The readback is plain English from the first committed clause, using question and choice labels. The XPath cell and the "✎ build" modal sit behind a per-column "code" toggle; a rule the sentence editor cannot show opens with its XPath visible so the author always has something to edit. The collapsed row summarises its logic in words ("shows when Chair rise includes Pass · 1 validation rule · computed"). "Compute the value as…" shows only for calculate rows or rows that already have a calculation; a plain question gets "+ compute this value…". Bytes are unchanged: every write still flows through the same serializer, and a displayed-but-unedited rule is re-emitted in its original spelling (`../lmp_date != ''` stays `../lmp_date != ''`). Tests - client/tests/sentence-editor.spec.ts: cold-start journey (pick, read back, code toggle, collapsed summary, bytes on disk) and the `../` reopen + byte-identical save; an unparseable rule opens with its XPath. - Existing specs updated for the new layout: the strip is addressed by `.cond-strip-unified[data-column=…]`, "+ insert" is "Apply", and specs that reach the "✎ build" modal click the strip's "code" first. - geriatric-build 9 / helpers: the modal rule row's "string" checkbox is addressed by its name (the 9c picker added a second checkbox). Demos - client/tests/t9-demos/*.demo.spec.ts + playwright.demo.config.ts: one captioned, slow-motion recording per sub-issue (9a–9g), written to client/demo/t9/<ticket>.webm (gitignored). Ignored by the main config. Validation: shared build+test green; typecheck clean; lint clean on the touched files (pnpm lint fails on master already); corpus sweep unchanged against the master baseline; e2e: only the four pre-existing master failures remain (demo 1 & 4, geriatric-build 7 & 8). Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
PrjShrestha
added a commit
that referenced
this pull request
Oct 2, 2026
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…o/t9; recorder also writes the .mp4 Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…the recorder keeps the copy in sync Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Closes #19. Last slice of the T9 epic (#9). Stacked on #28 (9g) → #27 → #26 → #25 → #24 → #23; merge in that order and this PR's diff collapses to the 9f commit.
What changed
The row's advanced panel is four groups: Logic, Display, Messages, Raw.
Bytes are unchanged. Every write still goes through the same serializer; a displayed-but-unedited rule is re-emitted in its original spelling (
../lmp_date != ''stays../lmp_date != '',.<=100stays.<=100).Tests
client/tests/sentence-editor.spec.ts— cold-start journey (pick, read back, code toggle, collapsed summary, bytes on disk);../fieldrelevants reopen as sentences and save byte-identical; an unparseable rule opens with its XPath..cond-strip-unified[data-column=…], "+ insert" → "Apply", specs reaching "✎ build" click "code" first). geriatric-build 9 and the geriatric helper address the modal's "string" checkbox by name (the 9c picker added a second checkbox to the rule row).Demo recordings
client/tests/t9-demos/+playwright.demo.config.ts: one captioned, slow-motion recording per sub-issue (9a–9g) written toclient/demo/t9/<ticket>.webm(gitignored; not part of the suite).Validation
pnpm --filter @cht-ui/shared build && pnpm --filter @cht-ui/shared testgreen;pnpm typecheckclean; eslint clean on touched files (pnpm lintfails on master already).node scripts/corpus-sweep.mjs— identical to the master baseline (pre-existing extraSheets / contact-summary drift only).🤖 Generated with Claude Code