UI fix:
- Close-only modals showing a redundant second close button now use
hideCancel: ReceivedInvoices paid-detail modal (was two identical "Zavřít"
buttons) and PlanCellModal "Den je součástí rozsahu".
- Add modal-duplicate-close test enforcing close-only modals set hideCancel.
Lint: cleared all 68 warnings → 0.
- preserve-caught-error: attach { cause } in ai.service / exchange-rates.
- no-require-imports: package.json version read via fs (APP_VERSION) instead
of require(), avoiding a rootDir-expanding static JSON import.
- react-hooks/exhaustive-deps (11): ref-in-cleanup copies, derived-value
useMemo wrapping, PlanGrid field extraction, stable nextKey useCallback,
AuthContext documented cycle-break.
- no-explicit-any (53): precise route param/Prisma types, generic enrich*()
preserving payload shape, minimal vite module type, frontend body/query-key
types, SystemInfo for Settings.
Refactor (test enablement): shift-form types moved to dependency-free
shiftFormTypes.ts so the print-HTML builders are unit-testable without the
component graph; characterization test pins their output.
Gates: 649 tests pass, tsc -b clean, lint 0. Verified touched flows live
via Playwright (PlanWork CRUD + optimistic cache, warehouse form keys,
Settings system info, invoice detail).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- SPZ matching normalized on both sides (4SY7039 finds '4SY 7039') in
list_vehicles and the list_trips vehicle filter; JS-side name matching
diacritic-folded to keep utf8mb4_unicode_ci parity; ICO typed with spaces
matches the space-less stored value (suppliers + customers)
- get_top_customers: whole-table per-customer counts of invoices (sans
drafts) / orders / projects, top 20 + customers_with_records, optional
year, per-type permission re-checks — answers 'pro koho pracujeme nejvic'
exactly instead of refusing
- find_employee, find_supplier and stock search now return total_matching
(stock search also gained its missing deterministic orderBy); system
prompt teaches the model to answer counts from total_matching and use
aggregation tools instead of summing 20-row list pages
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The system prompt now lists the data areas the caller's permissions filtered
out, so the model tells the user explicitly they lack rights to that module
(and can ask the admin) instead of guessing the feature does not exist or
pointing them at a module they cannot open. Tool-less users get the same
explicit phrasing.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A failed call followed by a successful retry showed two identical chips
(one orange). Retries are loop mechanics: the trace now carries one entry
per tool with ok = delivered-at-least-once, for live chips and persisted
history alike (also keeps meta.tools under its 30-entry cap).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- find_employee: name -> user_id with route-parity population scoping
(attendance perms see plan users, trips perms see drivers sans usernames,
full directory only with users.view; trips.history unlocks nothing)
- get_attendance_summary: day-range detail with worked hours (shared
calcWorkedHours, break subtracted) + totals; attendance.manage also
unlocks the tool
- list_trips: kniha jizd with route-identical non-manager self-scoping and
the stats-endpoint km coalesce; bare month defaults to current year
- get_work_plan: resolved plan grid rows; off-grid users need
attendance.manage (grid-route parity)
- executeTool flags handler {error} results not-ok (consistent denial chips)
- system prompt: resolve employee names via find_employee first (gated on
the tool being present)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Remove single-thread getChatHistory/appendChatMessages/clearChatHistory and the
three /history routes; add listConversations, createConversation, getConversation-
Messages, appendConversationMessages, renameConversation, deleteConversation with
ownership checks and auto-title; wire six /conversations[/:id] routes.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The amount on a received invoice is the GROSS total the user pays (VAT
included), and VAT is the portion contained within it — not a net base with
VAT added on top. Switch the formula from amount*rate/100 to the
VAT-inclusive amount*rate/(100+rate) via a shared, tested vatFromGross()
helper used by both the create (manual upload + AI import share this
endpoint) and edit paths. Rate 0 → no VAT.
- AI extraction now asks Claude for the total INCLUDING VAT (was "základ bez
DPH"), reinforced in the structured-output schema description.
- Label the amount field "Částka s DPH" in the AI review card and the manual
upload/edit forms so the gross convention is explicit.
- Add unit tests with the user-confirmed figures (22 542,91 @ 21% → 3 912,41,
base 18 630,50) plus other rates and the no-VAT case.
Existing rows keep their stored (gross) amounts; their vat_amount recomputes
on next edit. A one-shot backfill can correct historical vat_amount if wanted.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Server-side, per-user chat history (the thread now survives reload):
- new ai_chat_messages table + migration; getChatHistory/appendChatMessages/
clearChatHistory; GET/POST/DELETE /api/admin/ai/history (ai.use-guarded);
service + HTTP tests (isolation, ordering, clear, perm, validation).
- DashAssistant loads the thread on mount and persists each turn; "Vymazat"
clears it. gcTime:0 so each mount re-reads the DB (no stale-cache resurrection);
composer gated until history settles + seed flag flipped on submit (no
seed/submit race clobbering a freshly-sent turn).
UX + correctness fixes from review/feedback:
- review cards moved OUT of the scrollable thread into a "Faktury k potvrzení"
section so Uložit is always reachable; chat auto-scroll no longer fires on
field edits.
- chat bubble text colour set on the Typography (GlobalStyles pins `p` to
text.secondary, which was overriding the inherited bubble colour → unreadable).
- never clear input/staged PDFs on a failed submit (rollback + keep), so a
budget error can't wipe staged invoices; roll back the optimistic user turn.
- stable attachment ids (no crypto.randomUUID — undefined over plain HTTP).
- editable Datum splatnosti + Popis fields; client-side extraction summary.
Security: DashProfile clipboard copy now uses an execCommand fallback (works
over plain HTTP) and only shows the success toast when the copy actually
succeeded — previously it falsely claimed to copy 2FA codes / the TOTP secret.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
