integreat/docs/plans/2026-06-27-001-fix-ssr-modal-regressions-plan.md

SSR Modal Regression Fixes — QA Findings & Resumable Task List

Status: ALL identified items FIXED + verified (2026-06-27) — BUG A, BUG C, BUG D, BUG B, BUG E/F, the CSS size root cause, and a subtle wizard fade-in (§3). The only thing intentionally left is the optional richer forward/back slide transition, which needs design sign-off (§3). Owner: Bryce Branch: integreat-execute-refactor Date: 2026-06-27 Context: Follow-up to the SSR rendering-modernization migration (2026-06-02-001-refactor-ssr-rendering-modernization-plan.md, Phases 2–11 complete). The migration to simplified route rendering + whole-form/wizard swaps introduced modal size, animation, and a set of HTTP 500 regressions. This doc records a full browser QA pass (agent-browser, 9 migrated modals) with root causes and fixes.

0. How to resume / reproduce

• App is running on the port in .http-port (was 42987); nREPL in nrepl-port (was 35689).
• Browser session: agent-browser --session integreat .... Log in via http://localhost:<port>/dev-login → "Continue to dashboard".
• Test at a realistic viewport: agent-browser --session integreat set viewport 1440 900 (a short viewport exaggerates modal overflow and gives false positives).
• Screenshots from this pass are in ./tmp/ (gitignored): 02..19-*.png. Raw running notes: ./tmp/findings.md.
• After rebuilding CSS, the browser caches the old output.css; bust it by swapping the <link> href (?v=<n>) or hard-reloading, or you'll keep seeing the old sizes.

1. Test matrix (9 migrated modals)

Modal	File	Type	Result
Transaction Edit	transaction/edit.clj	form	Opens OK; BUG C — whole-form swap 500
Transaction Bulk Code	transaction/bulk_code.clj	form	✅ Works
Sales Summary Edit	pos/sales_summaries.clj	form	✅ Works
Invoice Bulk Edit	invoices.clj	form	✅ Works (empty-selection 500 = BUG D)
Transaction Rule	admin/transaction_rules.clj	wizard	✅ Works
Invoice Pay	invoices.clj	wizard	✅ Works (narrow until CSS rebuild — see §2)
New Invoice	invoice/new_invoice_wizard.clj	wizard	BUG A — choose-client 500
Vendor	admin/vendors.clj	wizard	Size broken (§2) + BUG E inner overflow
Client	admin/clients.clj	wizard	Size broken (§2) + BUG F inner overflow

Cross-cutting: §2 (stale CSS → sizes), BUG B (footer EDN leak), animations (§3).

2. ★ ROOT CAUSE of broken modal SIZES — stale compiled Tailwind CSS

resources/public/output.css (committed, last built Jun 24) is missing the migration's newer arbitrary size classes. Tailwind only compiles classes present in source at build time, so any size value added/changed after Jun 24 isn't in the CSS and the element falls back to its w-full h-full sibling class → the modal balloons to full width.

• Present (these modals size correctly today): md:w-[750px] (New Invoice), md:w-[950px] (Transaction Edit / Bulk Code / Sales Summary), md:h-[600px], md:h-[650px], w-[850px] (Transaction Rule).
• Missing (these balloon / mis-size): md:w-[760px] (Vendor), md:w-[820px] (Client), md:h-[520px], md:h-[560px], w-[50em] (Invoice Pay).

Verified: rebuilding to a temp file generates all missing classes; after rebuild + cache-bust, the Vendor card went 1257px → 760×520 and Client → 820×560 (exactly as declared).

DEEPER ROOT CAUSE found while fixing (2026-06-27): tailwind.config.js content only scanned ./src/**/*.{cljs,clj,cljc} — it never scanned resources/templates/**/*.html (the 46 Selmer templates the migration introduced). So a naive rebuild drops every template-only class (e.g. md:w-[950px] / md:h-[650px], used only in templates/transaction-edit/edit-modal.html, which would have re-broken Transaction Edit / Bulk Code / Sales Summary). The durable fix is to add the templates glob to the config, then rebuild.

FIX — DONE & verified:

1. Added "./resources/templates/**/*.html" to tailwind.config.js content.
2. npx tailwindcss -i resources/input.css -o resources/public/output.css (kept unminified to match the committed file; add --minify only if the prod pipeline minifies).

• Verified: all modal size classes now present (md:w-[760px], md:w-[820px], md:h-[520px], md:h-[560px], w-[50em] and the template-sourced md:w-[950px]/md:h-[650px]). Class-diff vs the old CSS shows the only removed classes are orphaned (the deleted mm/* modal-stack forward/backward/group/transition/htmx-*:translate-x-2/3 animation set + the unused lg:w-[900px] sales size) — none still referenced in src/ or templates/. Vendor modal confirmed live at 760×520 (was 1257px).
• Still TODO: confirm the prod/CI build runs this tailwind step (so output.css can't drift again) — ideally wire it into lein build / buildspec.

Note: during QA I rebuilt output.css to verify, then reverted it so the working tree is clean for review. The verified rebuild is preserved at tmp/output-rebuilt.css; backup at tmp/output.css.bak.

3. Modal ANIMATIONS

The new wizard step cards swap via hx-target "this" hx-swap "outerHTML" but their modal-card-advanced carries no transition classes, so step→step and modal enter/leave do not animate:

• New Invoice steps (render-basic-details, render-accounts): card class is only md:w-[750px] md:h-[600px] w-full h-full — no htmx-swapping:/htmx-added: variants.
• next-steps-modal (new_invoice_wizard.clj ~678): only static scale-100 translate-x-0 opacity-100 — missing the htmx-swapping:/htmx-added: swap variants.
• By contrast, dialog/success-modal- and the Invoice Pay card (last-modal-step transition duration-150) keep transitions, so the intended pattern still exists to copy from.

FIX — DONE (forward/back slide restored) in the shared wizard2 engine. The user confirmed the old wizard had a directional slide forward/back between steps; the engine migration dropped it. Restored the original mechanism (read out of the deleted mm/multi_modal.clj) in one shared place:

• wizard2/step-slide-classes — the group-[.forward]/transition:htmx-* + group-[.backward]/… variants, now applied to the swapped wizard <form>.
• wizard2/transitioner — the #transitioner wrapper with the @htmx:after-request hook that reads the x-transition-type response header and toggles group/transition + forward|backward on itself (so the variants fire on the next swap). All 5 wizard configs' :open-response now wrap the form in wizard2/transitioner instead of a plain #transitioner div.
• wizard2/handle-step-submit now sets x-transition-type (forward on advance, backward on Back, none on a same-step validation re-render) + HX-reswap: outerHTML swap:0.16s (the swap delay that lets the slide-out play). Direction is computed from the step order (transition-type).
• The earlier per-card fade-in (interim) was removed in favor of this.
• CSS rebuilt so the group-[.forward]/transition:htmx-* variants (16 fwd + 16 back) are compiled.
• Applies to ALL wizards (new-invoice, vendor, client, pay, transaction-rule) since it lives in the engine. REPL-verified: open-wizard emits the transitioner, the form carries the slide classes, and submit responses carry x-transition-type + the HX-reswap swap delay.
• Live-verify caveat: the long-lived dev server froze its route table at startup (auto-ap.handler/match->handler-lookup is a def that merged the per-ns key->handler maps), so an nREPL reload of leaf namespaces does NOT reach the running router — a server refresh (stop jetty → tools.namespace refresh → user/start-http) or a fresh lein run is needed to see it live. (Documented hazard; see the QA memory.)

4. Bug list + fixes (prioritized task list)

P0 — functional 500s

BUG A — New Invoice: choosing a client → 500 from /invoice/new/due-date.

• File: src/clj/auto_ap/ssr/invoice/new_invoice_wizard.clj, fn due-date (and same latent bug in scheduled-payment-date).
• Exception: ClassCastException: org.joda.time.DateTime cannot be cast to java.util.Date.
• Cause: form-vendor-client decodes :invoice/date/:invoice/due to clj-time DateTime (via clj-date-schema), but due-date then calls (some-> date coerce/from-date) — coerce/from-date expects a java.util.Date. scheduled-payment-date has the same (some-> due coerce/from-date) but is masked because due is nil-guarded by a when.
• Repro: open New Invoice, pick any client (vendor still empty) → red "unexpected error" banner; network shows PUT /invoice/new/due-date = 500. REPL-confirmed the decoded date is already a DateTime and time/plus works on it directly.
• FIX — DONE & verified live: dropped the coerce/from-date calls — due-date now uses the decoded date/due DateTimes directly (removed the date (some-> date coerce/from-date) rebind and the due coerce); same for scheduled-payment-date. Live: selecting a client now fires PUT /invoice/new/due-date + /scheduled-payment-date → 200 (was 500).

BUG C — Transaction Edit: any whole-form swap (mode toggle, vendor change, add/remove row, $/% toggle) → 500 from /transaction2/edit-form-changed, whenever the txn has ≥1 autopay-invoice match. (HIGH severity — breaks the flagship modal's core swap doctrine.)

• File: src/clj/auto_ap/ssr/transaction/edit.clj.
• Exception: ClassCastException: PersistentVector cannot be cast to Named at edit.clj:849 (links-body* does (name (:action step-params))).
• Cause: the link-panels each render a hidden <input name="action">. The unpaid panel (line 690) and rule panel (line 730) add :form "" to exclude that hidden from form serialization. The autopay panel (line 661) is missing :form "", so its action="link-autopay-invoices" hidden is serialized alongside the main action hidden (line 862, value "manual"). Ring collapses the duplicate action params into a vector → (name vector) throws.
• Repro: edit a transaction whose "Link to autopay invoices" tab shows a badge count ≥1, click "Switch to advanced mode". REPL-confirmed: single action → 200; action as a vector → the exact 500. (Txns with 0 autopay matches render panel-empty* with no action-hidden and swap fine — hence intermittent.)
• FIX — DONE & verified live: added :form "" to the autopay panel's action-hidden (edit.clj:661) to match the unpaid/rule panels. Live: on a txn with an autopay match, "Switch to advanced mode" now POSTs edit-form-changed → 200 with a single action=manual param (the duplicate action=link-autopay-invoices is gone) and the mode toggles correctly.
  • (Optional hardening, not applied: make links-body* coerce a vector action via (some-> (:action step-params) (as-> a (if (coll? a) (last a) a)) name).)

P1 — sizing / layout (after §2 CSS rebuild)

§2 CSS rebuild — do first; fixes Vendor/Client/Invoice-Pay widths.

BUG E/F — Vendor + Client wizard step layout overflows the card — DONE & verified live.

• Actual root cause (found while fixing): step-card places the step timeline in a vertical left sidebar (grow-0 ... self-stretch hidden md:block), but vendor-timeline/client-timeline rendered the horizontal timeline/timeline component. A horizontal <ol> in a shrink-to-fit sidebar forced the sidebar to ~full card width, pushing the body off-screen → horizontal scroll, clipped fields, the "huge grey left region". (Vendor additionally pinned its bodies to a fixed w-[600px], compounding it.)
• FIX applied:
  • vendors.clj vendor-timeline + clients.clj client-timeline: use timeline/vertical-timeline + timeline/vertical-timeline-step (the components already existed) instead of the horizontal pair.
  • vendors.clj: the 5 step bodies' w-[600px] h-[350px] → w-full h-[350px] so the body fills the remaining width next to the now-narrow vertical timeline.
• Verified live: Vendor 760×520 and Client 820×560 with a proper vertical timeline, no horizontal overflow, fields fill the body; forward nav (Info→Terms) works, the vendor-name chip renders, and the validation-error re-render also lays out correctly.
• Note: edits made with the plain Edit tool (exact-string) — the clojure-mcp editors reformat the whole file against a stricter config than the project's lein cljfmt, producing large spurious diffs; both files pass lein cljfmt check.

P2 — robustness / cosmetic

BUG D — Invoice Bulk Edit with no selection → 500 — DONE & verified live. With no selection, selected->ids returns nil, which all-ids-not-locked fed straight into a Datomic :in $ [?i ...] query → "Unable to find data source" exception → global "Oh, drat!" toast.

• File: src/clj/auto_ap/ssr/invoices.clj — wrapped all-ids-not-locked's body in (when (seq all-ids) ...), so an empty/nil selection yields [] and the modal opens cleanly (also protects every other caller). Live: Bulk Edit with nothing selected now opens the modal, no toast, no 500.

BUG B — Footer leaks a raw Hiccup attr-map as text in the red "unexpected error" banner ({:x-show "unexpectedError", ...}). Pre-existing in master (NOT a migration regression), but visible whenever unexpectedError flips true (e.g. it showed during BUG A). Cause: modal-footer- calls (hx/alpine-appear {...}) twice — the 2nd return value lands in child position and renders as literal EDN.

• File: src/clj/auto_ap/ssr/components/dialog.clj ~line 59.
• FIX — DONE & verified: deleted the duplicate (hx/alpine-appear ...) line in modal-footer-. Rendering the footer to HTML now shows a single alpine-appear (the legit attrs) and no literal {:x-show ...} text child.

Cosmetic — over-tall empty modals. Invoice/Transaction Bulk Edit and similar use fixed md:h-[650px]; with little content they show a large empty lower region. Consider letting height hug content (cap with max-h) rather than a fixed height. Low priority.

5. Suggested order of work

1. §2 CSS rebuild (unblocks all width checks; commit output.css; verify prod build does this).
2. BUG A + BUG C (the two functional 500s; one-liners each, REPL-verified).
3. BUG E / BUG F (wizard step-body layout rework; needs visual iteration per step).
4. §3 animations (add swap-transition classes to wizard step cards).
5. BUG D, BUG B (robustness / cosmetic).
6. Re-run the full agent-browser pass (all 9 modals) + the Playwright e2e suite as the parity gate.

6. What was verified vs. inferred

• Verified in browser + REPL: BUG A (500 + exact exception + fix direction), BUG C (500 + exact exception + vector repro + line 661 cause), §2 CSS root cause (missing classes + rebuild fixes Vendor 1257→760 and Client→820), BUG D (500 on empty selection), Vendor/Client inner overflow (screenshots 16/17), and that Bulk Code / Sales Summary / Transaction Rule / Invoice Pay / Transaction Edit-open all render without errors.
• Inferred (static analysis, not yet visually A/B'd against master): §3 animation regression (step cards lack htmx-swapping:/htmx-added: classes) and BUG B being pre-existing.