12ce693eb67b58010a5fa06beb81fe249a231fb4
1357 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 Jakob Wennberg
|
12ce693eb6 |
feat(mcp): make search-only read tools reachable, and put the payload ceiling into reverse (#1976)
* feat(api): surface the registry's worked examples in the OpenAPI spec and generated skill EndpointDefinition.example is required and every one of the 125 v1 endpoints populates example.response, but generateOpenApiSpec() never emitted it. The examples reached only the docs markdown builder, so /api/v1/openapi.json carried none and the generated skills/accounted-api had zero json blocks in all 12 reference files: every agent reading the spec or installing the skill got schemas with no concrete body. Emit example on the application/json media types (request body and 200 response) and teach the portable renderOperationMd to print it as a fenced json block. 178 worked examples now reach the skill. SKILL.md is unchanged: the examples land in the on-demand reference files, not the entry file. Attached to JSON media types only, so a multipart body and a binary application/pdf response do not advertise an example they cannot send. Adds the one missing example.request (currency-revaluation) so the new exhaustive coverage assertions hold. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix(api): emit Retry-After on a v1 429 so the documented contract is real The published accounted-api skill has told agents to honor Retry-After on a 429 since it shipped, but no /api/v1 route ever sent one: the wrapper's auth failure path early-returns through v1ErrorResponseFromCode, whose finalize() set only X-Request-Id and Gnubok-Version. Unattended clients had nothing to pace against and had to back off blindly. 60 seconds is an exact upper bound rather than a guess: the rate limiter is a fixed one-minute tumbling window per key row and the limited branch does not slide it. The value moves into an exported constant next to that limiter, so the MCP server's hardcoded '60' now reads from the same place. Also corrects the withApiV1 doc comment, which claimed step 8 stamps X-RateLimit-Limit. It never did. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * test(mcp): guard the tools/list payload for the namespace new installs get The payload ratchet only ever serialized the gnubok_* projection. The accounted_* projection is inherently larger (every tool reference gains 3 chars, ~209 tokens across the default catalog) and CLAUDE.md points new MCP installs at exactly that namespace, so the payload a new user's client receives was never measured. It had already drifted ~90 tokens past the 63.4K ceiling while the guarded number sat comfortably under it. Measure both and assert on the larger. The ceiling moves to 63.6K to cover the real worst case; this buys no new catalog surface. A second test pins the direction of the delta so Math.max cannot silently stop describing reality. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * feat(mcp): make search-only read tools reachable, and put the payload ceiling into reverse DECISIONS.md records on 2026-08-26 that gnubok_reconcile_match had to be promoted back into the default catalog because "a search-only tool is uncallable on Claude.ai". That is a client-side limit, not a server one: the tools/call dispatcher has always resolved names against the whole tools array, and isDefaultCatalogTool gates only what tools/list shows. So catalogVisibility: 'search' was unusable as a payload lever for reads, and the ceiling could only ever go up. gnubok_call_tool gives such a client one visible name to forward through. It is a rewrite in the dispatcher rather than a forwarding wrapper: {tool, arguments} is rebound to the inner tool BEFORE resolution, so the scope check, unknown-argument guard, company routing, test-key write block, staging _meta and telemetry all apply to the real target instead of being bypassed. Reads only; a write must be named directly so its approval contract stays visible. Alongside it, gnubok_get_agent_briefing's outputSchema drops 7743 to 4565 chars. Four sub-schemas whose interiors were documentation rather than contract are condensed to a permissive object plus a fuller description; agent-briefing.test.ts already pins their runtime shape, so nothing is left unguarded. Net on the guarded (accounted) projection: 63 491 to 62 942 tokens, with the new tool included. The ceiling moves 63.6K DOWN to 63.1K, the first tightening in that ledger, and the note now says to demote a read before proposing a bump. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|
Jakob Wennberg
|
3447da027a |
feat(api): agent-substrate quick wins: worked examples in the spec, honest Retry-After, and a payload guard that covers the namespace new installs get (#1974)
* feat(api): surface the registry's worked examples in the OpenAPI spec and generated skill EndpointDefinition.example is required and every one of the 125 v1 endpoints populates example.response, but generateOpenApiSpec() never emitted it. The examples reached only the docs markdown builder, so /api/v1/openapi.json carried none and the generated skills/accounted-api had zero json blocks in all 12 reference files: every agent reading the spec or installing the skill got schemas with no concrete body. Emit example on the application/json media types (request body and 200 response) and teach the portable renderOperationMd to print it as a fenced json block. 178 worked examples now reach the skill. SKILL.md is unchanged: the examples land in the on-demand reference files, not the entry file. Attached to JSON media types only, so a multipart body and a binary application/pdf response do not advertise an example they cannot send. Adds the one missing example.request (currency-revaluation) so the new exhaustive coverage assertions hold. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix(api): emit Retry-After on a v1 429 so the documented contract is real The published accounted-api skill has told agents to honor Retry-After on a 429 since it shipped, but no /api/v1 route ever sent one: the wrapper's auth failure path early-returns through v1ErrorResponseFromCode, whose finalize() set only X-Request-Id and Gnubok-Version. Unattended clients had nothing to pace against and had to back off blindly. 60 seconds is an exact upper bound rather than a guess: the rate limiter is a fixed one-minute tumbling window per key row and the limited branch does not slide it. The value moves into an exported constant next to that limiter, so the MCP server's hardcoded '60' now reads from the same place. Also corrects the withApiV1 doc comment, which claimed step 8 stamps X-RateLimit-Limit. It never did. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * test(mcp): guard the tools/list payload for the namespace new installs get The payload ratchet only ever serialized the gnubok_* projection. The accounted_* projection is inherently larger (every tool reference gains 3 chars, ~209 tokens across the default catalog) and CLAUDE.md points new MCP installs at exactly that namespace, so the payload a new user's client receives was never measured. It had already drifted ~90 tokens past the 63.4K ceiling while the guarded number sat comfortably under it. Measure both and assert on the larger. The ceiling moves to 63.6K to cover the real worst case; this buys no new catalog surface. A second test pins the direction of the delta so Math.max cannot silently stop describing reality. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|
Jakob Wennberg
|
c0ecb34a2b |
fix(company): migration reset no longer blocks on existing vouchers, sequences, or invoices (#1977)
* fix(company): migration reset no longer blocks on existing vouchers, sequences, or invoices The 2026-08-18 eligibility rule stopped the archive-and-replace reset before the first journal entry, voucher sequence, or invoice. The reset deletes nothing: the source stays a write-closed, downloadable retention container, and the unchecked Radera foretag path already produced the same outcome without any of those guards, so the blockers only led owners into a dead end (Carrierstories, 2026-08-26). External-state blockers are unchanged. Closes #1916 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * docs(decisions): move the migration-reset entry to the end of the log Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Jakob Wennberg
|
dfed55cb6c |
feat(periods): undo klarmarkera so an externally closed year can be reopened (#1978)
markPeriodClosedExternally ("klarmarkera") closes and locks an imported
year without a closing entry, and nothing could reverse it: unlockPeriod
refuses closed periods and the SIE replace flow refuses closed or locked
years. An owner who klarmarkerade five imported years and then found the
prior-year SIE file was wrong had no way back (Forsslund Systems,
2026-08-27).
reopenExternallyClosedPeriod reverses the mark while the closed state still
comes from klarmarkera (closed_externally set, no closing entry), clears the
lock, writes the audit_log row, and emits period.unlocked. New route
POST /api/bookkeeping/fiscal-periods/[id]/reopen-external with envelope codes
PERIOD_REOPEN_NOT_CLOSED / PERIOD_REOPEN_NOT_EXTERNAL; "Öppna igen" action
and "Avslutat i tidigare program" chip in Settings > Bookkeeping > Fiscal
years; unlock and SIE replace refusals now point at that path.
Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
|
||
|
Mattsson
|
c981384a0e |
fix(transactions): keep the assistant usable through the whole booking flow (#1975)
* fix(transactions): keep the assistant usable through the whole booking flow Three gaps around the agent sheet during booking on Transaktioner: - The AgentTrigger bubble lived inside #dash-shell, so the inert that non-modal dialogs set on the shell made the assistant impossible to OPEN once a booking dialog was up. Moved outside the shell (it is position: fixed) and raised to z-[45]: above the DialogVeil (z-40) so it stays clickable, below dialog content (z-50). Under true modal dialogs Radix's body pointer-events lock keeps it dead as before. - Wide dialogs centered on the full viewport while the docked sheet (z-60) covered their right edge, hiding e.g. the Granska button. DialogContent now centers in the space left of --agent-dock-w, and the wide booking/review variants cap their width against it. - Step 1 of the flow (template picker, QuickReviewDialog) was still fully modal, so the assistant was dead there. Both are now non-modal with DialogVeil + a shared ref-counted useDashShellInert hook (also replacing the duplicated inert effects in TransactionBookingDialog and NewInvoiceDialog). Their Esc/veil-click dismissal is kept: they hold no half-filled form. Clicks in the agent sheet or its trigger never dismiss any dialog: data-agent-ui counts as inside, same mechanism as data-dialog-companion. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(ui): dialog centering reads a docked-only sheet width variable Skeptic review refuted the first cut: globals.css seeds --agent-dock-w at the 10px frame gutter on :root by design, so the 0px fallback in the new dialog centering never applied and every dialog sat 5px left of center (full-bleed dialogs clipped 5px off-screen). Introduce --agent-sheet-w, set inline by AgentSheetProvider only while the sheet is docked and removed otherwise, so the 0px fallback is real: sheet closed or floating renders byte-identical to the old left-[50%] and old max widths. Also cap the template picker and QuickReview's narrow variant, which could clip off-screen left on narrow desktops with the sheet docked. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Mattsson
|
99c94d467a |
fix(white-label): back-to-clients link points at the byra cockpit's home domain (#1973)
* fix(white-label): back-to-clients link points at the byra cockpit's home domain A byra member working a company homed on another host (e.g. a pre-byra company on canonical) got a relative /clients on the wrong host instead of their white-label cockpit. resolveCockpitHref mirrors WL-14's home rule: relative when the current host is the cockpit's home (brand domain, or canonical for a brandless byra), else an absolute URL there. Cross- host links render a plain <a> with a 'Hanteras via' hint; the hop lands on the brand host's login (per-host sessions, WL-01) and WL-14 then lands on /clients. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(white-label): show the cross-host cockpit hint as visible text CodeRabbit: title-only hints never surface on touch devices, so the expanded-sidebar and mobile external back-links now render the 'Hanteras via {domain}' line as small muted text under the label (same pattern as the switcher's foreign entries). Also corrects the comments claiming the no-company branch never renders the back-link: it can, on its cockpit/settings surfaces, where the relative fallback matches pre-change behavior. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Mattsson
|
a860c690ed |
feat(white-label): WL-14 cockpit landing for BankID and OAuth/magic-link logins (#1972)
* feat(white-label): WL-14 cockpit landing for BankID and OAuth/magic-link logins Byra staff logging in via BankID or the Google/magic-link callback on their brand domain landed on /select-company resp. / instead of the cockpit, because those two paths bypassed the WL-14 landing rule. - Extract the rule into resolveLandingDestination (lib/company/landing-server.ts) so server code can call it without an HTTP round-trip; /api/clients/landing becomes a thin wrapper. - Auth callback: with no explicit destination, AAL1 sessions resolve the landing from the request host, degrading to / on any failure (MFA-enrolled users already get the rule via /mfa/verify). - BankID login: byra staff on their brand host get /clients; everyone else keeps the deliberate /select-company picker byte-identically. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(white-label): address PR 1972 review findings - /api/clients/landing: requireAuth() directly instead of withRouteContext, which 4xxed byra staff without a company of their own (COMPANY_CONTEXT_MISSING) and silently sent the cockpit's primary persona to /select-company. MFA enforcement unchanged. - landing-server: log the byra membership query error before degrading to '/' so a persistent failure is distinguishable from no membership. - Deduplicate the clientWithTeamMembership test mock to file scope. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(white-label): paginate the byra membership query fetchAllRows per repo convention: PostgREST silently caps unpaginated selects at 1000 rows, which could hide a qualifying owner/admin membership. Errors still degrade to '/' with a log. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Mattsson
|
b30c71086e |
feat(byra): gate the automatic cockpit landing to owner/admin (#1970)
* feat(byra): gate the automatic cockpit landing to owner/admin Plain byra members now land like regular users; owner/admin keep the cockpit landing at both decision sites (post-login /api/clients/landing and the '/' bounce). The middleware zero-company steer stays ungated: a member with zero companies has nowhere else to land. Cockpit access itself is unchanged (nav + /clients remain membership-based). Supersedes the 2026-08-05 all-members widening (DECISIONS.md). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(byra): keep byra members out of the first-run wizard on auto-landing Skeptic finding: a member whose auto-resolved active company is onboarding-incomplete (e.g. mid migration-reset, which repoints active_company_id itself) fell through the new role gate into /onboarding, a dead end for role member (WL-15 refuses client creation). Byra members without a picked-company cookie now go to /byra at the onboarding check, restoring the pre-gate shield. Also pins the role column into the landing route's select assertion so dropping it can't pass the mocked tests silently. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Jakob Wennberg
|
23776337c8 |
feat(onboarding): the hosted checklist closes with Connect to Claude (#1971)
Founder call 2026-08-27: the strategic split is Claude-first onboarding for brand-new businesses (no migration friction) and hosted onboarding for prior-system users, whose closing CTA now hands them to Claude with everything already connected: the MCP onboarding skill then opens with reconciliation findings and the Att göra-list instead of setup steps (shipped in #1967). The final checklist step 'Bygg din bokföringsassistent' becomes 'Anslut till Claude': the button opens claude.ai's Add-custom-connector dialog prefilled (connectorName from the brand, connectorUrl built from the page origin so self-hosted and white-label domains link to themselves). The in-app assistant calibration stays reachable from the assistant page. sv + en strings added. Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Mattsson
|
5e9cd6f761 |
fix(enable-banking): unstarve the daily bank sync cron (#1969)
* fix(enable-banking): unstarve the daily bank sync cron The sync cron self-limited to 50s (no maxDuration export, so the route ran under the 60s platform default) and processed ~17 connections per day against 123 entitled active connections: any given connection only got an automatic sync every 4-7 days, and users bridged the gap by clicking 'Synka' manually, which pushed them to the back of the queue. - export maxDuration = 300 (Vercel Pro ceiling the code always assumed) - sync loop budget 50s -> 230s; health probe gets the 280s leftover - connections sync in concurrent waves of 4 with per-connection error isolation preserved - MAX_CONNECTIONS_PER_RUN 50 -> 300 (safety cap only; one run now covers the whole entitled queue) Cadence stays once daily at 05:00 UTC by design; users who want more can sync manually. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(enable-banking): serialize same-company connections within sync waves Skeptic refutation: the post-sync unattended reconciliation sweep is company-scoped, so two connections of one company syncing concurrently run two identical whole-company sweeps whose unlinked-GL-line snapshots race; both can claim the same journal entry for different bank transactions, leaving the GL short while every surface shows reconciled. Waves now fan out over company groups instead of raw connections: one company's connections sync sequentially inside a single wave slot, unrelated companies still run 4-wide. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(enable-banking): re-check the sync time budget inside company groups Review finding (PR Reviewer Guide): the budget was only checked between waves, so one company with many connections could run past 230s inside a single wave and eat the health-probe and teardown margin. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Mattsson
|
98aff7c577 |
fix(transactions): keep booking dialog open while typing in the agent sheet (#1966)
* fix(transactions): keep booking dialog open while typing in the agent sheet
TransactionBookingDialog was a default modal Radix dialog; the agent
sheet is a fixed z-[60] panel portaled outside its DOM, so a click in
the assistant's text field counted as an outside interaction and
dismissed the dialog, losing the half-filled booking.
Apply the established non-modal convention (NewInvoiceDialog /
NewJournalEntryDialog): modal={false} + DialogVeil, with escape and
outside-dismiss prevented. Closing is explicit via the header X.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(transactions): restore page modality via inert while booking dialog is open
Skeptic review caught that the non-modal convention was ported
incompletely: NewInvoiceDialog pairs modal={false} with an inert effect
on #dash-shell. Without it the background page stayed keyboard-, AT-,
and (mobile nav) tap-reachable behind the veil, so a stray Shift+Tab
plus Enter could navigate away and destroy a half-filled booking.
Also corrects the comment that cited NewJournalEntryDialog as non-modal
precedent (it is plain modal).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
|
||
|
Mattsson
|
ff88e3de05 |
fix(bookkeeping): remove false 1580 tax-receivable label, let company account names win (#1968)
The hardcoded ACCOUNT_DESCRIPTIONS entry labeled 1580 'Fordran for skatt' with a Skatteverket explanation. That is wrong on both counts: tax receivables are 1640 (skattefordringar) / 1650 (momsfordran), and 1580 was traditionally 'Fordringar for kontokort och kuponger', which BAS has since moved to 1686 (why 1580 is excluded from our BAS 2026 catalog). Reported by a user who books card/Swish settlements there. Also flip AccountNumber display precedence to the company's own account_name over the hardcoded reference name: chart rows are user-editable data and must not be visually overridden by our copy. The tooltip keeps showing the BAS reference info. Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Jakob Wennberg
|
dae0aac26a |
fix(mcp): reconcile_match callable + two-phase efterkontroll + instant value for web-onboarded users (#1967)
* docs(plugin): starter prompt nudges the connector's onboarding guide Several E2E runs composed the first reply from tool descriptions before the skill loaded (questionnaire instead of the guided round). Four words in the published prompt point the agent at the guide from message one; the rest of the prompt stays memory-first and universal. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(mcp): reconcile_match into the default catalog + two-phase efterkontroll + instant value for web-onboarded users E2E #12 (the smoothest run yet) surfaced three things: 1. gnubok_reconcile_match was search-only, so Claude.ai could not call it: the agent misread the uncallable tool as a missing reconciliation:write scope and punted the entire dont-double-book matching step to the web app. Promoted to the default catalog (the standing rule: skill-instructed tools must be callable); ceiling 63K to 63.4K documented. 2. Skill: efterkontroll split into two phases per founder direction. Step 3b runs DIRECTLY after the import commits: verify the trial balance, sanity-read the content, and PREPARE the chart (create 1630/8423/8314/6992 when the SIE lacks them) so the bank and Skatteverket connections land in a book that is ready for them. Step 4b after the connections: match SIE-covered bank rows with reconcile_match (never re-categorize), reconcile skattekontot to the öre, names, underlag, in a prioritized numbered list. 3. Skill: users who onboarded via the web app and connect the MCP with everything already set up get the reconciliation pass and Att göra-list as the FIRST reply instead of setup steps: immediate value regardless of onboarding path. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Jakob Wennberg
|
b160f7290d |
docs(plugin): starter prompt nudges the connector's onboarding guide (#1965)
Several E2E runs composed the first reply from tool descriptions before the skill loaded (questionnaire instead of the guided round). Four words in the published prompt point the agent at the guide from message one; the rest of the prompt stays memory-first and universal. Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Jakob Wennberg
|
c2f055b903 |
fix(bank): 90-day lookback is the default: long PSD2 requests make some banks kill the session (#1964)
E2E #11: the account picker's fiscal-year default sent a 365-day request for a 405-day-old fiscal year; Swedbank answered by TERMINATING the session: zero transactions, connection expired, no error surfaced (initial_sync_requested_from 2025-08-26, returned min/max null). 90 days worked. - Default lookback mode is now 'fast' (90 days), labeled rekommenderas on a first connect; the fiscal-year option stays but carries an inline 'vissa banker avbryter kopplingen' note when its span exceeds 90 days, and the long-range helper names the real failure mode + the SIE/CSV path for older history. - SIE drop card: stage + arm in ONE click (three clicks was one too many): after the verdict, the single button reads 'Bokför: oåterkalleligt (BFL 5 kap 5 §)' and the deliberate click commits with confirmed=true. - Skill: never re-ask an answered question; when a connect card rendered, do not paste the URL as text too. Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Jakob Wennberg
|
53c4d0d194 |
feat(mcp): approve the SIE import in the drop card + post-import efterkontroll in the skill (#1963)
E2E #10 (the first fully successful drop-card run) left the staged import stranded until the user prodded the agent, and the agent then improvised an excellent post-import audit (skattekonto reconciliation, missing 1630, over-stated payroll liabilities). Codify both: 1. The drop card now carries the approval: after staging, the button becomes 'Godkänn bokföringen' with a two-click BFL confirmation (confirmed=true armed on the deliberate second click, exactly the pending-operations widget pattern), then 'Bokfört' + a ui/updateContext pointing the agent at trial balance and voucher-gap follow-ups. 2. Skill: when the user writes after a card was shown, the FIRST call is list_pending_operations (an empty ledger does not mean the file never arrived); new Step 4b 'efterkontroll' codifies the audit pass (trial balance vs SIE, skattekonto vs 1630 with 8423/8314/6992 for ränta/avgifter, auto-created bank account names, underlag coverage, voucher gaps); memory-first extended with memory-back (save orgnr, bank, fiscal year, moms period after creation so the next conversation needs zero questions). Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Mattsson
|
84e6b1eb17 |
fix(ui): cap DialogContent grid track so nowrap children cannot widen modals (#1962)
DialogContent is display:grid with an implicit auto column, and the auto track sizes to the widest child's min-content. Chrome counts nowrap text (truncate, whitespace-nowrap) at its full width in that calculation even though it truncates at layout time, so one long description (e.g. a candidate row in MatchVerifikationPicker) widened the track past the dialog, stretched every sibling, and clipped the right edge behind a horizontal scrollbar. grid-cols-[minmax(0,1fr)] caps the track at the content box, hardening every modal at once; callers can still override via className (tailwind-merge resolves the conflict in their favor). Verified with a headless-Chrome replica of the reported dialog: scrollWidth 696 vs clientWidth 510 before, 510/510 after. Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Jakob Wennberg
|
ff84c8f316 |
fix(mcp): widget-bearing tools must be read-only: Claude.ai drops write-annotated interactive tools (#1961)
The SIE drop card kept flapping into Claude.ai's Interactive-tools list and vanishing, and the agent could never call it: every hypothesis (scope filter, tool-count cap, stale cache, schema shape) was eliminated against live data until one discriminator remained: all surviving widget tools carry readOnlyHint true and gnubok_create_sie_upload was the only one annotated as a write. Claude.ai accepts always-render widgets only on read-only tools and silently drops the tool otherwise. readOnlyHint is now true, which is also honest: the tool only mints a short-lived upload URL; the actual write is the staged gnubok_import_sie: the exact receipt_matcher shape (read-only widget tool, writes via separate approval-gated tools). A guard test pins the invariant for every future widget tool. Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Mattsson
|
9bad3aee06 |
fix(build): bound Turbopack build memory so Vercel's standard container stops OOM-killing builds (#1958)
Since the white-label merge the production build dies with SIGKILL roughly every other run: the compile phase outgrew the standard build machine. turbopackMemoryEviction 'full' evicts finished tasks to the persistent FS cache after every snapshot, bounding the compile working set; cpus 2 caps the static-generation worker pool (default cores-1 full Node processes). The two govern disjoint build phases. Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Jakob Wennberg
|
d8244ecaff |
feat(mcp): connect_migration: one-click card into the previous-system wizard (#1960)
* feat(mcp): gnubok_connect_migration: one-click connect card into the previous-system wizard 'Jag hade Fortnox' now gets the same feel as Skatteverket: the tool returns the migration-wizard link for the named provider and renders the connect-card widget (new migration branch: 'Hämta från Fortnox', button opens the wizard that logs into the old system and fetches all fiscal years plus invoices, customers, suppliers and documents). For visma/bokio (no API export) the instructions order the SIE drop card first and this card as the complement. Scope companies:read; skill step 3 points at the tool instead of raw wizard links; ceiling 62.4K to 63K documented. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(mcp): 5-minute freshness hint on tools/list, widgets and prompts: the catalog changes with every deploy The stateless-client CacheableResult hint on tools/list, resources/read (widget HTML) and prompts/list was 1 hour. Claude.ai honors it, so for up to an hour after a deploy the connector served a pre-deploy catalog: a freshly shipped tool flapped in and out of the tool list depending on which fetch hit the client cache, and two E2E runs dead-ended on 'tool does not exist' for a tool that was live server-side. These payloads are static only within one deploy; 5 minutes bounds the stale window. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Jakob Wennberg
|
5715dd3def |
docs(mcp): onboarding skill branches on the named previous system (#1959)
'Jag hade Fortnox' now has a real answer instead of one generic SIE instruction: API-connected systems (Fortnox/BL/Briox/Wint) get two offered paths (full migration wizard with invoices/customers/documents vs quick SIE drop, recommended by need), Visma/Bokio get the file-first path with the wizard as a complement after, unknown systems get the universal SIE export ask. Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Jakob Wennberg
|
0f7625535b |
feat(mcp): drag-and-drop SIE import card: exact bytes via tools/call, no model in the byte path (#1957)
E2E #6: the flow ordered SIE-first correctly, but the agent never discovered gnubok_create_sie_upload, ran a local preflight, and sent the user to the web wizard again; it also rendered a duplicate generic bank card before the Swedbank-specific one. 1. New sie-drop widget (ui://sie-drop/app.html), rendered definition-level by gnubok_create_sie_upload: the user drags the .se/.sie file onto the card, the widget reads the EXACT bytes (FileReader), computes sha256 (WebCrypto), calls gnubok_sie_preflight via tools/call with file_content_base64 + sha256, shows the verdict, and on Importera stages gnubok_import_sie with the preflight's mappings. No network from the iframe, no model reproduction: byte path goes through the host bridge only, narrated into chat via ui/updateContext. 2. The inline size cap now applies only WITHOUT sha256: a hash-verified payload is byte-exact by proof, so the widget's 100 KB+ base64 passes while unhashed model-retyped content stays refused. 3. Discovery + ordering fixes: create_sie_upload/preflight/import descriptions name the card path explicitly; create_company's history_note points at the card; connect_bank description says pass bank on the FIRST call when the user has named it (the duplicate generic card came from a bare call followed by the nudged retry). Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Mattsson
|
fdb5f6f891 |
feat(white-label): byra white-label infrastructure: brands, cockpit, home domains, branded email (#1956)
* feat(white-label): brand and team-kind foundation
- brands table: one white-label identity per byra team (unique mutable
domain, row presence = live, email sender identity, hex color CHECKs)
- teams.kind ('personal'|'byra'): ops-only kind changes, deterministic
ensure_user_team (personal team only), AFTER UPDATE role re-sync so a
demoted consultant loses admin in client books immediately
- resolveBrandByHost/resolveBrandForCompany with 60s TTL cache, derived
chrome tone and WCAG contrast gate; no brand row = default appearance
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(white-label): per-request brand theming, wordmark slot and source footer
- root layout resolves the brand from the Host header and injects a
server-rendered style block (light + dark), font pair classes and a
BrandProvider/useBranding context; default hosts render byte-identically
- BrandWordmark logo slot, host-aware manifest and favicon,
images.remotePatterns for Supabase Storage logos
- curated font menu mechanism (font_key -> variable pair, preload:false
for non-default entries)
- AGPL source-code footer link on login and public pages, both brands
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(white-label): byra team invites, member management and team billing
- team invites unfrozen behind a kind gate (byra teams only, owner/admin
invite); members route handles multi-team membership; members/[id]
unfrozen with last-owner protection; invite management UI in settings
- billing/status learns team-scoped grants and the settings page shows a
read-only "part of the byra agreement" state instead of the upgrade pitch
- 30-day trial suppressed for companies created under a byra team
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(white-label): brand-aware outbound mail, auth email hook and public invoice branding
- every outbound mail is sent in the brand of the company it concerns:
getSenderForCompany/getBaseUrlForCompany chain (verified brand domain,
"via Accounted" fallback, canonical default) wired into invites,
payslips, invoice deliveries and reminders
- Supabase Send Email hook endpoint (signature-verified with node:crypto,
dormant until configured) renders auth mail per brand via redirect origin
- public invoice pages carry the company's brand mark
- snapshot suite per template class guards against wrong-brand mail
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(white-label): byra cockpit, home-domain rule and tab guard
- Klienter route: five urgency-sorted columns (company, unbooked, inbox,
next deadline via the status engine, last booked) for byra team members,
who land there after login on their home domain
- soft switch straight into a client and back; blocking two-exit tab
guard against writes to the wrong active company
- client company creation admin-gated at the DB level (a created company
is +1 on the byra invoice), bound to the byra team, no trial
- home-domain rule in the UI: switcher partitions companies by host,
signpost page for companies homed elsewhere
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(white-label): brand-aware app name across UI strings
- 24 message keys per locale converted to the {appName} ICU parameter,
27 call sites pass the active brand name (useBranding client-side,
getRequestAppName server-side)
- 6 hardcoded JSX literals swept; statutory filing and API identity
surfaces deliberately keep the Accounted name
- 34 new i18n keys for the cockpit, team invites, billing state, tab
guard, signpost and source footer (sv/en parity verified)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* docs(white-label): domain glossary and decision log entries
- CONTEXT.md: the white-label ubiquitous language (brand, byra team,
home domain, signpost, umbrella subdomain, brand color, cockpit)
- DECISIONS.md entries from the build waves
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(white-label): lean byra cockpit sidebar with company-mode back link
Byra team members now get a two-mode sidebar: on cockpit routes (/clients
and the new /byra pages) only Hem, Klienter, Automationer and Nyckeltal
show; entering a client company brings back the full company sidebar with
a pinned back-to-clients link (expanded, rail and mobile). New pages: /byra
home with client count, needs-action count and per-client urgent deadlines
reusing the fetchClientOverview aggregation, plus designed empty states for
/byra/automations and /byra/kpi. Signpost gate allows the byra routes;
non-byra users are unaffected.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(white-label): cockpit shows no active company and keeps lean sidebar under settings
In cockpit mode the bottom user widget no longer shows the active company
subline or the company-switcher flyout: the cockpit sits above the
companies and clients are entered through the Klienter list. The settings
modal previously flipped the sidebar to the full company nav behind it
because the pathname becomes /settings/*; the sidebar now keeps the mode
of the surface underneath.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(white-label): keep company picker in cockpit with nothing selected
The cockpit user menu gets the company-switcher flyout back, but neutral:
the row reads "Valj bolag", no company carries the check mark or active
styling, and picking any company (including the technically-active one)
enters it with a full navigation. Company mode is unchanged.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* chore(db): renumber white-label migrations past main and add byra settings scope
Renumber 20260801100000-120000 to 20260804110000-113000: main already
carries applied versions up to 20260803231000, and Supabase branching
refuses local migrations stamped before the remote head (the repo rule
from 5932632f5: keep new versions strictly newest). Comment references
updated in the pg tests, route docs and onboarding precheck.
Also ships the byra settings scope: settings opened from the cockpit
(?ctx=byra, honored only for byra team members) show account-level
sections only (Konto, Medlemmar och roller), hide company-scoped
sections and the company kicker, and the team section is registered in
SETTINGS_SECTIONS so Medlemmar och roller renders inside the settings
window. The cockpit user menu drops Abonnemang and carries the scope on
its links; section switches preserve it.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(byra): cross-client nyckeltal view in the cockpit
Period presets and company chips in the URL, summary tiles, merged
monthly income/expense chart and a sortable per-client KPI table.
Numbers come from the existing get_kpi_report_aggregates RPC per
client (no new migrations); calendar months are the cross-client
axis since clients can have different fiscal years.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(white-label): byra self-service brand logo and app name
New Varumarke settings section (byra scope, owner/admin): logo
upload/remove and an editable app name; domain stays read-only.
brands has no write RLS by design, so writes go through
/api/byra/brand routes with the service client behind an explicit
owner/admin team check. Files land in logos/byra/{teamId}/. The
expanded sidebar shows the brand app name beside the logo.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(white-label): route root layout through the shared brand resolver
app/layout.tsx carried a private copy of resolveRequestBrand, so it
and lib/branding/request-brand.ts could drift. The layout now uses
the shared function, which also gains a BRAND_DEV_DOMAIN override:
on literal localhost hosts only, resolve that brand so branding is
testable in local dev. Real domains are unaffected even if the
variable leaks into a deployment.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(byra): automations roadmap teaser and cockpit i18n strings
The Automationer tab now previews the planned automation set
(Monday briefing, deadline watch, rule-driven bookkeeping,
connection watch, monthly checklist, report delivery) instead of a
bare empty state. Bundles the sv/en strings for the whole cockpit
wave (nyckeltal, varumarke, automations) and the decision-log
entries.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(white-label): byra owners/admins land in the cockpit, not an auto-picked company
After login "/" resolved the first-membership fallback and opened a client
company nobody chose, and the top-left brand mark always linked back to it.
Byra owners/admins now home to /byra: the logo links there always, and "/"
redirects there unless a company was explicitly picked this browser session.
The middleware writes the fallback company back to user_preferences, so the
DB cannot tell picked from auto-picked; setActiveCompany stamps a session
cookie (gnubok-company-picked) on every explicit switch instead. The byra
check on "/" reuses the layout's team_members query via a request-cached
helper, so it costs no extra round trip. Byra members and regular users are
unchanged.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* refactor(white-label): drop brand color theming, keep monochrome everywhere
White-label is logo + app name + domain only (founder call): the
layout no longer injects brand color CSS variables, stamps
data-brand or colors the browser chrome. buildBrandVarsCss, its
WCAG gate and the brand_color/chrome_color columns stay dormant
for a future opt-in.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(db): arm SIE RPC statement_timeout via pgrst.db_pre_request hook
ALTER FUNCTION ... SET statement_timeout (20260629160100, 20260721144311)
never re-arms the running statement's timer, so large SIE imports still
died at the role default 8s. The pre-request hook runs as its own
statement before the main query, so set_config there is what the main
statement's timer is armed with. Scoped by request path to the three SIE
RPCs; every other request keeps 8s.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(byra): drop the 'what's coming' tail from the automations intro
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(white-label): byra owners/admins with zero companies land in the empty cockpit
Both no-company gates (Edge middleware and the dashboard layout) sent
every company-less user to the onboarding wizard, which forced a fresh
byra owner to create a personal company before ever seeing the cockpit.
Byra owners/admins now pass through to cockpit routes (/byra, /clients,
/companies/new, /settings, /api) and are steered to /byra elsewhere.
Plain byra members and regular users keep the onboarding redirect.
The membership lookup runs only in the rare no-company state, so the
middleware hot path is untouched.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(white-label): auth wordmark shows the brand logo alone
Byra logos usually carry their own name, so logo + app name text on the
login/register hero read as a duplicate. Branded hosts with an uploaded
logo now render the logo only, with the app name as the image's alt
text. Hosts without a logo keep the text wordmark unchanged.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(white-label): per-brand favicon via brands.favicon_url
Branded hosts used logo_url as the tab icon, which squashes wide byra
lockups at 16px. New optional brands.favicon_url holds a square mark;
the root layout prefers it and falls back to logo_url as before.
Migration applied to staging (idempotent DDL).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(auth): wire the villkor and integritetspolicy footer links
Both auth pages shipped with href="#" placeholders. Villkor now points
at the platform terms on the marketing site (accounted.se/terms; the
terms are the platform's even on branded byra hosts) and
integritetspolicy at the in-app /privacy page, host-relative so it
resolves on every branded domain. Both open in a new tab so the auth
form state survives.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(settings): styled popup for the team role dropdowns
The byra team panel's role pickers (member rows + invite form) were
native selects, so the opened list rendered as the unstylable OS menu.
Swapped to the Radix Select with the popup styled like every other
overlay; the trigger keeps the flat quiet SettingsSelect look.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(email): branded sender shows the brand name alone, no via-platform
Byra invite mail read "Willem via Accounted" in the From display name.
The tier-2 fallback (brand on the platform address) now renders just the
brand name; the platform stays visible in the actual From address until
the brand verifies its own sender domain (tier 1, unchanged).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(white-label): byra landing applies to every team member, not only owners/admins
An invited byra consultant (role member) still landed in an auto-picked
client company after signup. The cockpit landing rules ("/" redirect,
brand-mark home link, and both no-company gates) now key on byra team
MEMBERSHIP instead of the owner/admin role: anyone with cockpit access
homes to /byra. Regular users unchanged.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(email): branded team invite names the byra, not "ett team pa <platform>"
Subject, headline, body and text variant now read "Du har blivit
inbjuden till <Byra>" (brand casing kept) when the team has a brand.
Brandless teams keep the platform phrasing byte-identical.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(white-label): sidebar keeps cockpit mode after refresh on settings
The sidebar's cockpit/company decision on /settings/* rested on React
state remembering the surface underneath, which a hard reload wipes: a
byra user refreshing settings opened from the cockpit got the full
company nav and read it as landing in a client company. The ?ctx=byra
marker already in the URL survives reloads, so the sidebar now honors
it as the cockpit signal alongside the in-session memory.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(white-label): hide the active-company chip in byra-scoped settings
The full-page settings header (the hard-refresh fallback surface) showed
the ActiveCompanyBadge even under ?ctx=byra, so a byra user read the
auto-active client as "the company I am in". The chip now follows the
same byra-scope rule as the modal's kicker.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(white-label): tab guard no longer fires in the tab that initiated the switch
BroadcastChannel delivers the company-switch broadcast to every listener in
the same tab too, so the cockpit tab raised its own WL-09 "switched in
another tab" dialog over the hard navigation into the clicked client.
performCompanySwitch now marks the switch as self-initiated; CompanyTabSync
suppresses only the dialog for that observation (stray writes still get
their 409) and clears the marker on bfcache restore so back-navigation
regains the full guard.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(settings): styled popups for every settings dropdown
SettingsSelect rendered a native <select>, whose OS listbox cannot be
styled and clashes with the panel (same problem the team-panel role
dropdowns had). It now renders through Radix Select with the flat
dashed-underline trigger, keeping the native prop surface so all 13 call
sites work unchanged: value/defaultValue, onChange(e.target.value),
<option> children, and a hidden input that carries `name` into
SettingsFormWrapper's FormData read and raises the bubbling input event
its dirty tracking listens for. Empty-string option values map onto a
sentinel at the Radix boundary. The backup form's boxed fiscal-year
select moves to the shadcn Select with a placeholder.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(white-label): home-domain affinity redirect in middleware
Every signed-in user now homes on a domain: byra team members on their
brand's domain, everyone else on the platform app URL, except a byra's
client users, whose home is the byra domain their companies live under.
On any other product host the request redirects to the home domain's
root, where the user meets the RIGHT branded login (sessions are
per-domain by design). localhost, direct *.vercel.app hosts and IP
hosts are exempt; a 15-minute host-scoped cookie caches the "this is
home" verdict so the hot path costs zero extra queries; lookup failures
fail open. Complements the WL-01 signpost, which keeps handling
per-company homing inside a domain.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(white-label): render hero brand logo at 64px on auth pages
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(white-label): shareable invite link and re-send for byra team invites
A failed invite mail previously surfaced only as a toast description while
the invitation quietly waited for a mail that never arrived (the Arbore
case). The inviter now always has a recovery path:
- persistent share-link line after invite create/re-send: ochre attn line
with a copy action when the mail did not go out, quiet muted line with
the same action when it did
- POST /api/team/invite/[id] re-sends a pending invitation with a fresh
token and expiry (same byra-only owner/admin gates as DELETE)
- brand mail sending extracted to lib/email/send-team-invite.ts, shared
by create and re-send so the two paths cannot drift
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(white-label): sidebar shows uploaded brand logo alone, no app-name label
Byra logos usually carry their own name, so logo + text in the expanded
sidebar read as a duplicate (same founder call as BrandWordmark,
2026-08-05). The app-name label now renders only for branded hosts
without an uploaded logo.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(white-label): close the four skeptic refutations before merge
- trial seed: migration 130300 now carries the seven-key PAID body from
20260818170000 plus the byra guard, instead of silently reverting it;
pg test pins the full key set against PAID_CAPABILITIES
- byra gate: new migration 130600 adds the owner/admin gate to
create_company_for_user (v1 API + MCP path), and both surfaces resolve
the default team personal-only, so a consultant's private company can
never attach to the byra team
- home-domain: byra staff who also have canonical-homed companies are no
longer redirected off the platform host; the signpost handles per-company
homing (5 new middleware tests)
- settings selects: the Radix popup renders optgroup group headers again
(ROT/RUT work-type picker)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test(schema): re-baseline unresolvable-expression ceiling after #1954 catch-up merge
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(white-label): pg-real rollback-safe assertions and deep-link-preserving affinity redirect
The byra company-creation pg test asserted persisted rows through the pool
after withUserContext, which always rolls back its transaction; the
assertions now run inside the transaction after RESET ROLE. The home-domain
affinity redirect carries the original path and query across the domain hop
(PR Agent finding), so invite links and deep links survive the correction.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
|
||
|
Jakob Wennberg
|
6f0f4d1eea |
fix(mcp): flow-order signals in tool results: ask the bank, import history first (#1955)
E2E #5: the agent skipped the bank question (generic picker link) and went straight to connect cards for a company whose fiscal year started 13 months ago (history import should come first). The skill says both, but the agent follows tool results more reliably than skill prose, so the rules now live in the results themselves: - gnubok_lookup_company still_to_ask gains two flow questions: which bank (for the bank= deep link) and whether the bookkeeping lived in a previous system (SIE import BEFORE bank). - gnubok_create_company's confirmed result computes days of history in the fiscal period: past 90 days it emits history_note (PSD2 cannot reach it; run gnubok_sie_preflight first) and reorders the remaining- setup message to (1) SIE import, (2) bank, (3) Skatteverket. - gnubok_connect_bank called without bank= now opens its instructions with a nudge to ask for the bank and re-call for the direct consent link. Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Jakob Wennberg
|
cd9fb5b717 |
feat(mcp): byte-exact SIE upload path + brevity and memory-first onboarding (#1954)
From the fourth E2E run: the agent correctly refused to reproduce a 104 KB SIE file token by token (silent mid-verifikat truncation) and dead-ended to the web wizard, and its replies were walls of compliance prose. 1. gnubok_create_sie_upload: signed same-origin upload URL (reuses the pending-document infra; .se/.sie/.si only, 50 MB HTTP cap). gnubok_sie_preflight and gnubok_import_sie accept upload_id as the byte-exact source, plus optional sha256 (hex of the raw bytes) verified on the upload_id/base64 paths so truncation is DETECTED, never silent. Inline content above 120k chars is refused with a pointer to the upload flow. Scope bookkeeping:write (same intent as import_sie). 2. Skill: brevity rule (max ~8 short lines per reply, one warning per step, no legal essays), memory-first rule (check what is already known before asking the opening questions), the upload-first SIE step, and gnubok_explain_voucher_gap after import for skipped voucher numbers. 3. CONNECTORS.md starter prompt rewritten memory-first so it stays copy-paste ready without the user's own data in it. Plugin v1.2.2. tools/list ceiling 62K to 62.4K documented in the bench. Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Mattsson
|
dcb5d329a3 |
fix(skatteverket): show a banner when the connection is dead (#1925)
* fix(skatteverket): make a dead Skatteverket connection impossible to miss A company whose SKV session had expired got no usable signal anywhere. /skattekonto/saldo answers 200 for a token already flagged needs_reconsent (it keeps the stale snapshot visible on purpose), so the page looked healthy on load, and the reconnect AttnLine lived inside the `data` branch: a company that had never completed a sync had no snapshot, so clicking "Synkronisera nu" set the reconnect message and rendered absolutely nothing. The 401 branch returns before the toast, so the click was a silent no-op. On /transactions the same banner was gated on the 'skatteverket' source filter, which most users never select. Skattekonto page: - probe /status on load through the shared skvStatusNeedsReconnect predicate, so the prompt appears before the user clicks anything - render the reconnect line at page level, regardless of `data` - swap the "Synkronisera nu" header action for "Anslut igen": the sync cannot succeed while the session is dead, and offering it is what made the failure look like nothing happening - give the expired case its own StartCard instead of the onboarding one, which reads as "you never set this up" to someone who did - drop the "klicka pa Synkronisera nu" hint while a reconnect is pending Transactions page: show the reconnect line on every source filter. Max one attn line per page is preserved: reconnect outranks the imported-rows line and the saldo shortfall, the latter because a shortfall computed from a stale snapshot must not outrank "the data is stale". Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(skatteverket): show a banner when the connection is dead Cut back to a banner and nothing else, after three skeptics refuted the first cut. /skattekonto: the reconnect line lived inside the `data` branch of the saldo section, so a company with no snapshot (never completed a sync) saw nothing at all when the sync it just clicked died on an expired session. The 401 path returns before the toast, so the click was a silent no-op. Move the line to page level and probe /status on load so it also appears without clicking, since /skattekonto/saldo answers 200 for a token flagged needs_reconsent and nothing else in the payload reveals a dead session. /transactions: show the same line on every source filter. Nothing is hidden or removed while the flag is true: the saldo tiles, the Kronofogden line, "Synkronisera nu", the empty state and the shortfall warning all stay exactly as they were. Prod has 151 of 162 connected companies with every token dead, so a flag that hides anything hides it from nearly everyone, and the shortfall warning's action is the page's only route to the bankgiro and OCR. Also: a healthy probe clears a message left by an earlier failed sync (previously only syncNow's success path did, so a transient 401 kept the banner up until a remount), and the existing visibility refetch now covers the reconnect state so the banner does not survive the consent that fixed it. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(skatteverket): guard status probe against stale responses A fire-and-forget /status probe from an earlier reload could resolve after a later one and overwrite needsReconnect with an outdated result, restoring the reconnect banner right after a successful sync. Each reload now bumps a probe sequence and a result is applied only while it is still the latest. Also puts the 2026-08-26 decision record in the required [date] decision: why format. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Jakob Wennberg
|
4af7469523 |
feat(onboarding): minimal input: orgnr + moms period is the whole ask (#1952)
* feat(onboarding): minimal input: orgnr + moms period is the whole ask
Two fixes from the third E2E attempt (2026-08-26):
1. accounting_method is now optional in CompanySetupSchema and defaults by
form in planCompanySetup: aktiebolag = accrual (the norm), enskild
firma = cash (the common small-EF choice; legal under 3 MSEK, BFL 4
kap 4 paragraf). The plan flags the default (resolved.accountingMethodDefaulted)
and gnubok_create_company's preview carries accounting_method_defaulted
so the readback names it and the user overrides in the same 'ja'.
Never silent: the preview is the checkpoint. Applies to the MCP tool
and POST /api/v1/companies (additive; response shows the resolved
value). The lookup tool's still_to_ask no longer lists it.
2. The agent refused a real orgnr because the user said 'nytt bolag' and
the registry showed an established company ('Stopp. Numret matchar
inte ett nytt bolag'): lookup instructions now state that an
established company with F-skatt/VAT is the NORMAL case (new = new to
Accounted) and the orgnr is never second-guessed for looking
established.
Skill + plugin (v1.2.1) updated; API skill regenerated; DECISIONS.md entry.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(onboarding): surface the kontantmetod 3-MSEK condition on the defaulted cash method
Compliance-review finding on #1952: the EF cash default carries a legal
eligibility condition (turnover normally under 3 MSEK, BFL 4 kap 4 §)
that a client not reading the onboarding skill would never see. The
create preview now carries accounting_method_note with the condition
whenever cash was defaulted, and the v1 pitfall states it for API
integrators. The registry cannot verify turnover, so the confirm-time
human check is the gate; the default itself stays (a brand-new EF has
zero turnover by definition).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
|
||
|
Jakob Wennberg
|
4ac7b45c8a |
perf(layout): dashboard layout in two waves, nav flags as one RPC, local JWT verification (#1946)
The dashboard layout runs on every hard load, hard refresh, company switch and the 16 router.refresh() sites, and loading.tsx cannot paint until it resolves. It cost ~20 network calls in 4 sequential waves: a third getUser() round trip to Supabase Auth (after the proxy's and the route guard's), the company resolution, then 16 reads including four limit-1 probes whose only job is to decide whether to render the Webshop and Körjournal nav rows, and an entitlements read that itself ran two waves. - lib/auth/claims.ts: claimsPinned/userFromClaims extracted from require-auth.ts (unchanged) so the dashboard request context shares the exact pinning + mapping. getDashboardAuthContext verifies the JWT locally and falls back to getUser() only when claims are missing, unpinned or unverifiable: the proxy already performed the per-request revocation check before the layout runs (same semantics approved for routes on 2026-07-23). - Wave 1 (user-keyed, parallel with the company resolution): team membership, profile, user preferences and the memberships join, which now also supplies the active company's row and role, so the separate companies and company_members reads are gone. - Wave 2 (company-keyed): settings, agent profile, the switcher's settings names, entitlements in ONE wave (getCompanyEntitlements takes the team_id the join already carries and runs the grants read alongside config + subscription), and get_dashboard_nav_flags(). - supabase/migrations/20260826120000_get_dashboard_nav_flags.sql: SECURITY INVOKER, STABLE, EXECUTE for authenticated only; RLS applies inside. lib/dashboard/nav-flags.ts wraps it with the pre-RPC four-probe fallback on PGRST202/42883/42501 (self-hosted not yet migrated, deploy ordering) and degrades to hidden rows on any other error. - tests/pg/dashboard-nav-flags-rpc.pg.test.ts (6): fresh company, active vs pending WooCommerce, active Shopify, mileage trips, RLS for a member of another company, EXECUTE grants. Unit tests for the wrapper (RPC row, single-object payload, each fallback code, other errors) and for the entitlements teamId option. ~20 calls / 4 waves -> ~12 calls / 2 waves, 0 auth network calls. Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Jakob Wennberg
|
1a41119682 |
perf(bundle): drop the BAS chart and the Node crypto polyfill from the shared client baseline (#1942)
* perf(bundle): drop the BAS chart and the Node crypto polyfill from the shared client baseline
Two chunks rode along in the first-load JS of almost every dashboard route:
the full BAS 2026 chart (315 KB uncompressed, in 81 route manifests) and
the browser polyfill for Node's crypto/vm/Buffer (327 KB, in 26 routes
incl. login and register). Neither was needed on first paint; both got
there through static imports of helpers that happen to live next to code
that needs the data or the builtin.
Node polyfill (4 pure splits, behaviour unchanged, re-exported from the
original modules for server callers):
- lib/auth/bankid-flags.ts: isBankIdEnabled (login, register, security
settings imported it from bankid.ts, which imports crypto).
- lib/import/bank-file/formats.ts: the format registry + detection (the
import history imported getFormat from parser.ts, which hashes).
- lib/salary/personnummer-format.ts: parsing/validation/formatting (the
employee forms reached the encrypting personnummer.ts via tax-column).
- lib/auth/api-key-scopes.ts: scope catalogue, groups, tool map, helpers
(the API key panel imported STAGING_SCOPES from the key generator).
BAS chart:
- lib/bookkeeping/bas-lazy.ts + use-bas-reference.ts: the chart becomes a
dynamic import, fetched once per session after first paint; components
that show BAS names/descriptions call useBasReference() and re-render
when it lands. Until then (and on the server) only the hardcoded
account-descriptions answer, so SSR and hydration agree.
- lib/bookkeeping/bas-labels.ts: class/group labels out of bas-reference.ts
(account-descriptions needed a label and paid for the whole chart).
- lib/bookkeeping/bas-account-numbers.ts (generated, ~11 KB) +
scripts/generate-bas-account-numbers.ts (--check) + parity test:
isStandardBASAccountNumber for AddAccountDialog/ChartOfAccountsManager.
- lib/bookkeeping/account-classifier-{heuristic,client}.ts: the BAS-aligned
heuristic shared by the server classifier and a client variant that uses
the lazy chart.
- lib/bookkeeping/invoice-accounts.ts: INVOICE_FX_RATE_MISSING,
InvoiceFxRateMissingError, getRevenueAccount, getOutputVatAccount out of
invoice-entries.ts, whose engine import pulled account-backfill and the
chart into SendInvoiceDialog/PaymentBookingDialog.
- CorrectOpeningBalanceDialog re-seeds names when the chart lands;
OpeningBalanceRowEditor builds its Fuse indexes lazily; the
ChartOfAccountsManager BAS-katalog tab awaits the chunk.
Tooling:
- scripts/perf/client-import-closure.mjs: static import closure of every
'use client' module with the shortest chain to a target (file or bare
specifier); found every path above without a build.
- scripts/checks/client-node-builtin.mjs wired into check:guards: a client
module reaching a Node builtin is a hard failure (0 today).
Left as is: invoices/[id], its credit page and SendInvoiceDialog still
reach the chart through lib/invoices/issue-credit-note -> invoice-entries
-> engine -> account-backfill; splitting the engine is out of scope here.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(perf): unambiguous import-edge regex in the closure walker (CodeQL js/redos)
One quantifier per span: a greedy [^'"]* up to the specifier quote, which it
cannot cross, so a run of whitespace has a single parse. Same edges as
before (multi-line named imports, re-exports, side-effect imports; type-only
imports still skipped).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
|
||
|
Jakob Wennberg
|
e7e4efbfbc |
feat(oauth): one-click consent: all scopes pre-selected, list collapsed, Allow above the fold (#1953)
The read-only default forced every agent-first user to scroll a scope
list and hand-tick write rows before the flow could work. Founder call
2026-08-26: pre-check ALL scopes when the client requests none (Claude's
connector case), collapse the scope list into an expandable details fold
('Alla förvalda, visa och justera'), and keep the Allow button visible
without scrolling.
Why this is defensible: every write is STAGED for explicit approval
before anything touches the ledger, each scope row stays individually
untickable inside the fold, the warn line states the staging rule right
above the button, and the grant is revocable under Inställningar >
API-nycklar. A client that requests explicit scopes still gets exactly
that set (RFC 6749 3.3 least-privilege unchanged), and the tampered/empty
POST fallback stays read-only.
CONNECTORS.md gains the share link (connectorName/connectorUrl params)
plus the starter prompt to pair with it.
Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
|
||
|
Jakob Wennberg
|
41424a1650 |
feat(ux): one silhouette from route fallback to detail content (#1944)
Opening a row on the customers, invoices, verifikat, supplier-invoice, supplier, article and salary-run lists flashed three unrelated layouts: the segment's list-shaped loading.tsx (or, for suppliers and articles, the Hem-shaped dashboard fallback) during the RSC round trip, then the client page's bare centred spinner in an h-64 box while it fetched, then the content with a full layout change. Two flashes per click on the most travelled drill-down path. - components/common/DetailPageSkeleton.tsx: back link + title row + card grid + line block, the silhouette of a document/register detail page; InvoiceEditorSkeleton for the editor routes (same shape the Ny faktura dialog shows while its chunk loads). - loading.tsx for every [id] segment (customers, invoices, invoices/edit, invoices/credit, bookkeeping, supplier-invoices, suppliers, articles, salary/runs) and for the two list segments that had none (suppliers, articles, cloned from customers/loading.tsx). - The client pages render the same skeleton while they fetch instead of the centred Loader2, so the RSC fallback to client fallback handoff is invisible. The reference-data gates are already gone (A1 to A6); this only covers the primary-entity fetch. - app/(dashboard)/__tests__/detail-loading-states.test.ts pins both. Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Jakob Wennberg
|
3ee3565d6d |
perf(reference-data): sweep the remaining raw reads onto the session cache, ratchet to 0 (#1941)
Final consumer migration of the responsiveness plan: the 35 files still fetching fiscal periods, settings, accounts, cash accounts, dimensions or templates on their own now read lib/reference-data, and every client write site invalidates the shared cache instead of refetching locally. Settings and registries: FiscalYearsManager, FiscalPeriodEditor (period snapshotted once per company so a revalidation cannot reset dates being edited), BookingTemplatesPanel, ChartOfAccounts, ChartOfAccountsManager, EditAccountDialog, CorrectionEntryDialog, StrikeLinesDialog, InvoicePaymentAccountsSettings; the dimensions registry (DimensionsManager, DimensionCombobox, LineDimensionFields, DimensionFilter, bookkeeping/[id]) reads useDimensions and the ad-hoc fetchDimensions/fetchDimensionsCached helpers are deleted. Pages and pickers: CashAccountSelector (FyPicker-shaped restore, once per company load), use-account-names, FiscalYearGapNotice, OpeningBalancePeriodStep, BankFileConfirmStep, ImportReviewStep, the import page (invalidates accounts + periods after a SIE execute), customers list, invoices list + detail, pending, salary employee, asset dispose, year-end and periodisering pages (invalidate periods after closing), reports DimensionPnlView (its pivot picker read the wrong payload key and was always empty; it now populates), SkatteverketPanel, TemplatePicker, ArticleForm (vat_registered). Invoice dialogs and extensions: SendInvoiceDialog, PaymentBookingDialog (init reduced to the credit-note lookup + catalogue, proposal and voucher preview fire on open when cached; a local getSession replaces the network getUser for the fallback CC), InvoiceInboxWorkspace, TicWorkspace, ArcimMigrationWorkspace (invalidates after each SIE import step), enable-banking AccountPickerDialog. raw-reference-fetch ratchet: 35 -> 0 files. Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Jakob Wennberg
|
4560ccbfc9 |
perf(forms): supplier-invoice form, register forms and review dialogs read the session cache (#1938)
The supplier-invoice editor issued four requests on every mount (suppliers, accounts, settings, fiscal periods) and defaulted vatRegistered=true, entity type and rounding until /api/settings landed, so the moms controls visibly flipped. The register forms fetched the whole chart of accounts to fill one konto combobox, and each transaction review dialog refetched accounts, cash accounts or settings per open. - use-supplier-invoice-data: thin composition of useSuppliers, useAccounts, useCompanySettings and useFiscalPeriods; the settings-driven gates come from a pure deriveSupplierInvoiceDefaults() (tested) instead of state that flips when the fetch returns; the per-invoice öresavrundning toggle is the one local override. Inline supplier create invalidates the shared list instead of patching local state. - SupplierForm, ArticleForm (posting accounts), QuickReviewDialog, InvoiceMatchDialog, supplier-invoices/[id] (payment dialog chart): useAccounts; ArticleForm's inline account create invalidates the chart. - BulkBookDialog, MatchVoucherDialog, DuplicateBookingDialog: cash accounts from useCashAccounts (resolveAccount over the cached list; an empty list still resolves to 1930 with the fallback note). - QuickReviewDialog, BulkBookDialog, NewEmployeeDialog, customers list (default payment terms), salary run page (payment format, bank, IBAN, dimensions): derived from useCompanySettings; the salary page's post-settings-modal refetch becomes a cache invalidation. raw-reference-fetch ratchet: 45 -> 35 files. Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Jakob Wennberg
|
40e773548c |
perf(invoices): the invoice editor renders on the first paint from the session cache (#1937)
"Ny faktura" was the slowest form to fill in: the list page lazy-loaded NewInvoiceDialog, which lazy-loaded InvoiceEditor (ssr:false), which then issued four requests on mount (customers, articles, chart of accounts, company settings) and hid the ENTIRE form behind a spinner until the customers query alone resolved, even though the other three had landed. Reopening the dialog paid all of it again. - InvoiceEditor reads customers, articles, posting accounts and settings from lib/reference-data (seeded by the dashboard layout). The whole-form spinner gate is gone; the customer picker shows "Hämtar kunder ..." only while the list is genuinely uncached. Company settings are applied once per editor instance through a guarded effect, so a background revalidation can never re-run the create-mode prefills over notes or a reference the user has typed. Inline customer/article creation invalidates the shared cache (awaited, so the new option resolves before the line points at it). Customers now come through /api/customers, which masks the personnummer column; nothing in the editor rendered it. - NewInvoiceDialog imports the editor statically: the dialog is itself a next/dynamic chunk on the list page, so this is one deferred chunk download when the dialog opens instead of two sequential ones. - New strings: invoice_editor.loading_customers (sv + en). Per "Ny faktura": 2 sequential chunk loads -> 1; blocking mount requests 4 -> 0 (cached) with every field populated on the first render. raw-reference-fetch ratchet: 46 -> 45 files. SendInvoiceDialog and PaymentBookingDialog (init() flows) stay in the baseline for a later PR. Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Jakob Wennberg
|
567fae654c |
perf(bookkeeping): booking dialogs render populated on open from the session cache (#1935)
The bookkeeping dialogs were the customer's "fields load late" in its purest form: Bokför (TransactionBookingDialog + the embedded JournalEntryForm) issued five requests on every open (fiscal periods, accounts, settings, cash accounts, then the voucher preview once the first two had landed), Nytt verifikat the same minus one, BookDirectlyDialog four, and the template dialogs two. Each Radix dialog unmounts on close, so every reopen paid the full price again, and several fields visibly flipped: the bank line seeded '1930' then rewrote itself, the series defaulted to 'A' until settings arrived, the period select was empty. All of them now read lib/reference-data (seeded by the dashboard layout): - JournalEntryForm: periods, accounts and settings from the hooks; dimensionsEnabled derived, not fetched; the voucher-number preview is keyed on the entry date (the route resolves the period from it) so it fires as soon as the series is known instead of after the period fetch; after activating accounts it invalidates the shared accounts cache; the create-period dialog callback invalidates the periods cache. - TransactionBookingDialog: settlement account and its name derived with useMemo from the cached cash accounts; the form mounts on the first paint. - BookDirectlyDialog: cash accounts, periods and accounts from the hooks; the '1930'-then-rewrite disappears because the resolved account is known on the first render. - TemplateBookDialog, BookingTemplatePicker, TemplatePicker: templates (and periods) from the hooks. - BookingTemplatesPanel (delete, import) and CreatePeriodDialog (create) invalidate the corresponding cache entries so every picker sees the change at once. - fetchers.ts: booking templates are booking_templates rows (BookingTemplateLibrary), not the static BookingTemplate shape. Per open: Bokför 5 requests -> 0 blocking (voucher preview is a non-blocking hint), Nytt verifikat 5 -> 1 non-blocking, BookDirectly 4 -> 0, Mall 2 -> 0, template pickers 1 -> 0. raw-reference-fetch ratchet: 51 -> 46 files. Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Jakob Wennberg
|
ec9cab24cc |
feat(mcp): ask the bank first: connect link deep-starts the named bank's consent (#1951)
* feat(mcp): ask the bank first: connect link deep-starts the named bank's consent The connect card used to open the generic picker page; the user then chose the bank there. The agent now asks 'vilken bank har företaget?' among the opening questions and passes it to gnubok_connect_bank, whose connect_url becomes /import?mode=psd2&bank=<name>. BankSelector resolves the name (exact, then unique prefix, then unique substring: ambiguous names fall back to the prefilled picker rather than guessing an institution) and auto-starts that bank's consent through the same onConnect handler, so the duplicate-pending and renew-instead guards stay fully interactive. The param is stripped via history.replaceState after the one-shot so an aborted bank flow plus back-navigation does not silently relaunch. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(mcp): first-year suggestion keeps the AB vs enskild firma distinction Compliance-review finding on #1949: the suggestion text collapsed both forms onto a 31 December end. Only an enskild firma's first year MUST end 31 December; an AB may pick any end within BFL 3 kap 3 §'s 18-month cap, with 31 December as the common default. The lookup tool's still_to_ask line and the skill now say so explicitly, and first-year-defaults documents that fiscalYear-null is a strong-not-perfect filed-report signal that must only ever feed confirm-question suggestions. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Jakob Wennberg
|
9a56b7aff9 |
perf(bookkeeping): fiscal-year pickers and cash accounts read the session cache (#1934)
First consumer migration onto lib/reference-data. FyPicker and
FiscalYearSelector (14 consumer surfaces, 47 fiscal-period fetch sites
before this series) now read useFiscalPeriods(); with the layout seed the
restore of the persisted scope runs in the first effect tick and onReady
fires on mount instead of after a round trip. Their restore rules are
extracted into a pure resolveInitialFiscalScope() (lib/reference-data/
fiscal-scope.ts) so the two pickers cannot drift apart again, and the
restore runs once per company load, not on every background revalidation.
- /reports: the static catalog renders immediately; only the "no fiscal
year" empty state waits for the picker (previously six skeleton bars
until /api/bookkeeping/fiscal-periods resolved).
- JournalEntryList (/bookkeeping): resolves its initial scope from the
cached list instead of its own fetch; the saved-scope shortcut still
unblocks the entries fetch first when nothing is cached, and resolution
is guarded to once per company so a revalidation can never snap a
deep-link "all years" visit back to the stored year.
- /transactions: the account chooser reads useCashAccounts({ enabledOnly })
(seeded) instead of fetching /api/cash-accounts on every visit; the bank
sync button invalidates that entry after a sync.
- STORAGE_KEY_PREFIX / ALL_YEARS_VALUE move to a dependency-free
fiscal-year-storage.ts (re-exported from FiscalYearSelector) so lib/ code
can import them without a React component.
raw-reference-fetch ratchet: 55 -> 51 files.
Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
|
||
|
Jakob Wennberg
|
b436d47b64 |
feat(mcp): onboarding flow v2: migration-first SIE import in chat, one-confirm momentum, extended first-year heuristic (#1949)
Three changes from the first real E2E run (Arcim, 2026-08-26): 1. gnubok_sie_preflight: read-only scan of a SIE file shared in chat BEFORE anything is staged: parse, validate (per-verifikat balance, IB, closed-year P&L residual), CP1252-mojibake tripwire, duplicate file/period check, org-number match against the company (the wrong-company import is the worst silent failure this flow can have), and suggested account mappings shaped for direct passthrough to gnubok_import_sie. Both tools now also accept file_content_base64, decoded with the same encoding detection as the HTTP upload route so CP437 exports keep their åäö. 2. Onboarding skill v2: opens with TWO questions (orgnr + 'vilket system hade du innan?'), imports history before the bank (PSD2 rarely reaches far enough back), and a momentum rule: the create preview is the ONLY stop; connect tools are called without asking, and categorization starts as soon as the bank is active. connect_bank instructions now describe the account-selection dialog that actually gates the first sync, and the bank history cap. 3. deriveFirstYearDefaults: no closed fiscal period in the registry now extends the first-year window from 12 to 18 months (BFL 3 kap 3 §): a 13-month-old company with no annual report is still in its first, extended räkenskapsår (the Arcim case the 12-month rule missed). Applied in the web journey and the lookup tool. tools/list ceiling 61.5K to 62K, documented in the bench. Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Jakob Wennberg
|
52bfd7a399 |
perf(auth): skip the MFA factor lookup once the session is at AAL2 (#1933)
The enforced-MFA branch of the proxy called supabase.auth.mfa.listFactors() on every page, RSC and prefetch request for every MFA-verified user with a company. auth-js implements listFactors() as a getUser() network round trip, so hosted page requests paid two Supabase Auth calls in sequence. At AAL2 a verified factor exists by construction (the session got there by verifying a challenge on one), so the lookup only runs on the aal1/aal1 path (users mid-enrolment), where it still gates exactly as before. The one case deferred is a user who unenrols their last factor mid-session: the JWT keeps aal2 until the next refresh, so the enrolment bounce lands on the refresh instead of the next click. Tests: aal1/aal1 still asks for the factor list and bounces to /mfa/enroll; at aal2 listFactors is never called (even with a factor list that would read as empty) on page, RSC and prefetch requests; the step-up bounce and the no-company skip are unchanged. Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Jakob Wennberg
|
47fe193c48 |
feat(perf): session-cached reference data layer, server-seeded, with a raw-fetch ratchet (#1932)
* feat(perf): session-cached reference data layer, server-seeded, with a raw-fetch ratchet
Customer report (2026-08-26): "it takes time before all fields load when
clicking around". The cause is on the client: fiscal periods, settings,
accounts, cash accounts, dimensions and templates are fetched raw from 47 /
27 / 14 / 8 / 12 / 5 independent call sites, uncached, on every mount and
every dialog open, each request paying the auth proxy and route wrapper
before its own query. SWR was adopted for exactly this on 2026-07-13 but
reached only three files.
This PR adds the layer; consumers migrate in the follow-ups.
- lib/reference-data/keys.ts: one key builder per data set, company id in
position 1, null without a company; company_settings keeps the shape
useCompanySettings already uses so that hook is seeded without a change.
- lib/reference-data/fetchers.ts: browser Supabase for fiscal periods and
cash accounts (mirroring period.list and listForCompany ordering, pinned
by tests), /api for the lists whose routes do real work (accounts RPC,
dimensions ensure, template scoping, customer masking).
- lib/reference-data/hooks.ts: useFiscalPeriods, useCashAccounts,
useAccounts, useDimensions, useBookingTemplates, useCustomers,
useSuppliers, useArticles (+ re-exported useCompanySettings); one-minute
dedupe, keepPreviousData, background revalidation kept on so writes from
MCP/agents/other tabs surface.
- lib/reference-data/invalidate.ts: invalidateReferenceData(kind) for the
success path of every client write.
- lib/reference-data/seed.ts + components/providers/ReferenceDataSeed.tsx:
the dashboard layout fetches fiscal periods and cash accounts in its
existing batch and hands them, with the settings row it already had, to
SWR as fallback, so the first form of a session renders its period, bank
account and settings-driven fields on first paint. getDashboardSettings
now selects the full row for that (its other consumers read a subset).
The chart of accounts is not seeded (hundreds of KB for large charts).
- scripts/checks/raw-reference-fetch.mjs, wired into check:guards as a
per-file ratchet: GET-shaped fetch('/api/<reference path>') anywhere in
client-facing code and .from('<reference table>').select( in 'use client'
files. Baselined at 55 files; new sites fail CI; at 0 the entry goes.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(checks): anchor every optional whitespace run in the raw-reference-fetch regex
CodeQL js/redos flagged the `\s*,?\s*\)` tail: two adjacent optional
whitespace runs around an optional comma backtrack polynomially on a long
near-miss. The URL and init-object pieces are now named fragments and
every whitespace run is followed by a literal, so there is one way to
match. Behaviour unchanged (same 7 fixtures + baseline count of 55);
a worst-case timing test pins the linear scan.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(checks): make the use-client directive regex unambiguous (CodeQL js/redos)
An unclosed /* let the lazy comment body be re-split at every later /*.
The body is now (?:[^*]|\*(?!\/))* which cannot cross a */, so the outer
repetition has one parse. Pinned with a 3000-comment worst-case test.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* ci: re-trigger checks for the rebased head
No workflow ran for dd560a7af (nor after close/reopen); an empty commit
gives the pull_request event a fresh head. No code change.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(checks): single-character whitespace alternative in the use-client detector (CodeQL js/redos)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
|
||
|
Jakob Wennberg
|
a8980a3a41 |
perf(hem): stream the home page in three sections and trim its query plan (#1945)
Hem was one ~33-query render behind a single fallback: the greeting waited for the slowest worklist scan, and the request also paid a sequential bank_file_imports read after the batch, a second scan of the suggested matches (getWorklistCounts counted the same 200 rows the pane listed) and an awaited stale-dismissal delete on the read path. - page.tsx awaits only what the greeting shell and the redirects need (settings, profile, agent profile, the Skatteverket flag); the notice line, the setup checklist and the Att göra + Fortsätt panes are async server components behind their own Suspense (hem-sections.tsx). RSC streaming applies to client navigations too, so the greeting paints first on every visit and each block fills in as its queries land. - DashboardContent becomes the shell with three slots; HemNotices keeps the one client-side action (the wrong-account sign-out); HemSkeletons are the two fallbacks. - getWorklistCounts accepts the suggested matches the caller is already fetching (a promise, so it stays parallel); listSuggestedMatches runs once at the scan cap and the pane shows the first five. - countInboxDocuments runs its id chunks in one wave instead of N sequential round trips. - getCompanyNotices takes deferReap; Hem passes Next's after() so the stale-dismissal delete runs after the response. - bank_file_imports joins the checklist section's batch. Tests: worklist aggregate (precomputed matches skip the rescan), notices aggregate (deferReap receives the reap; the delete does not run inline and runs when the task does). Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Jakob Wennberg
|
a3326a0296 |
perf(nav): hover-intent prefetch for the dashboard nav + 30 s client router cache (#1943)
* perf(nav): prefetch dashboard routes on hover intent, not on viewport
DashboardNav renders ~45 links, all dynamic routes with a loading
boundary, so Next prefetched every one of them as soon as the nav mounted.
Each prefetch is a full request through the auth proxy (Supabase Auth
round trip, active-company RPC, MFA check) whose only payload is the
shared loading skeleton; prod logs showed 1,000 to 1,300 such hits per nav
route per day.
NavLink wraps next/link with prefetch={false} and an explicit
router.prefetch on mouseenter/focus/touchstart (Link's own hover prefetch
is disabled together with viewport prefetch, so the warm-up must be
explicit). The link to the current route and non-routes are skipped
(shouldWarmNavRoute, tested). A source-shape test pins that DashboardNav
has no bare next/link left.
Cost: an un-hovered click shows the route's loading skeleton ~50-100 ms
later than before; the skeleton is all a dynamic prefetch ever carried.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* perf(router): keep dynamic routes in the client router cache for 30 s
experimental.staleTimes.dynamic was 0: every back/forward or repeated nav
click re-requested the RSC payload through the auth proxy. 30 s covers the
click-around pattern the customer described while the 16 router.refresh()
sites after mutations keep the pages that must not go stale fresh.
Separate commit so it can be dropped on its own if stale numbers are
reported.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
|
||
|
Jakob Wennberg
|
c31933b15b |
perf(api): write routes stop re-resolving the active company (#1928)
* perf(api): write routes stop re-resolving the active company
withRouteContext resolves the active company (one resolve_active_company
RPC, ~40 ms p50 on prod) and then, for the 256 routes that pass
requireWrite: true, called requireWritePermission(), which resolved it a
second time before its role select. Two sequential round trips repeating
work the wrapper had just done, on every mutating request.
requireWritePermission() and getCompanyRole() now accept an optional
`known` context; the wrapper passes { companyId }, so the helper goes
straight to the membership select. Callers that pass nothing behave
exactly as before, and the shared selectRole() keeps both helpers on the
same query. The role is still looked up, never trusted from the caller.
Tests: known companyId skips resolution, known role skips the select, a
known viewer is still 403, a known company without a membership row is
still 403, legacy calls unchanged; new lib/api/__tests__/with-route-
context.test.ts pins that the wrapper resolves the company exactly once,
hands it to the guard, never calls the guard on read routes, passes the
guard's 403 through with a request id, and emits Server-Timing.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* test(customers): viewer gate expects the wrapper to hand over the resolved company
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
|
||
|
Jakob Wennberg
|
b2e15bbd2a |
feat(perf): measure the auth proxy per request (Server-Timing + proxy completed log) (#1922)
The proxy in front of every page, RSC, prefetch and /api request makes several sequential network calls (getUser, session state, the resolve_active_company RPC, MFA factor lookups) and nothing measured them, while the route wrapper has logged authMs/companyMs/handlerMs per API call for months. This is the first PR of the responsiveness plan (customer report: "it takes time before all fields load when clicking around"): the baseline every later change is measured against. - lib/supabase/proxy-timing.ts: pure helpers (request classification from the app-router headers, route template that collapses ids and tokens, Server-Timing formatting, a timed() accumulator). - lib/supabase/middleware.ts: updateSession wraps updateSessionInner, times each phase, sets Server-Timing on page/RSC/prefetch responses and X-Proxy-Timing on /api responses (withRouteContext owns Server-Timing there), and emits one "proxy completed" log line per request. - scripts/perf/log-percentiles.ts: p50/p90/p99 per group over `vercel logs --json` output, for both "op completed" and "proxy completed"; scripts/perf/README.md documents the protocol and targets. Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Jakob Wennberg
|
6dd0e951e6 |
ci: publish accounted-mcp and gnubok-mcp to npm when their version changes (#1920)
* ci: publish accounted-mcp and gnubok-mcp to npm when their version changes accounted-mcp has never been published (npm view is E404) although every "connect Claude" doc says `npx -y accounted-mcp`, and gnubok-mcp is at 1.0.1 on the registry while the repo has carried 1.1.0 since #706. No workflow published to npm; this adds one. .github/workflows/npm-publish.yml runs on a push to main that touches a packages/*/package.json, and on workflow_dispatch (package: all or one, plus a dry_run that packs and validates without touching the registry). One matrix job per package: it fails first with a message naming the NPM_TOKEN secret if it is absent, then compares the package.json version with `npm view <name> versions` (E404 counts as "never published", any other failure is an error), skips when the version is already on the registry, and otherwise runs `npm publish --provenance --access public`. Permissions are contents: read plus id-token: write for the provenance attestation. Actions are pinned to the same SHAs as the sibling workflows. npm rejects a provenance attestation whose package.json repository.url does not match the source repository, and gnubok-mcp still pointed at erp-mafia/gnubok, so both repository fields now name erp-mafia/accounted in npm's canonical form with the monorepo directory. `npm pkg fix` normalised the bin paths, and accounted-mcp's index.mjs gets the executable bit gnubok-mcp's already had. Versions are not bumped. Both READMEs get a Releasing section: bump version, merge to main, the workflow publishes; the NPM_TOKEN repository secret must exist. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Signed-off-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com> * fix(packages): keep the ./index.mjs bin form the package tests pin Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * ci(npm-publish): scope NPM_TOKEN to the publish step and keep the matrix static The token was job-level env, visible to checkout, setup-node and the version gate; it now reaches only npm publish. The matrix no longer interpolates the workflow_dispatch input into an expression: both packages always get a job and a Select step skips the one not requested. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Signed-off-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com> Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Jakob Wennberg
|
9396e54965 |
docs: correct stale product facts (arkivplan, architecture, agents, self-hosting, extensions, database map) (#1931)
Every statement was verified against the code on main before editing; the docs had drifted from the product in ways a customer or agent would act on. - public/docs/arkivplan-mall.md: product named erp-base; magic-link login; BAS 2025/2026; eu-central/eu-west region; US subprocessors for AI. Now Accounted, e-mail + password + TOTP (BankID optional), BAS 2026, eu-north-1 Stockholm, Bedrock in EU with Resend as the only US subprocessor; adds rättelselogg, Peppol inbound, skattekonto imports and the säkerhetsbackup ZIP to the räkenskapsinformation tables. - ARCHITECTURE.md: adds the inline-rättelse correction path, OAuth 2.1 and lazy MCP auth, accounted-mcp and claude-plugin, 150+ tools. - AGENTS.md: defers to CLAUDE.md instead of a drifted copy; keeps the Codex-only constraints with the Supabase project name fixed (erp-base). - README.md: drops LangChain/OpenAI (not dependencies), Node 20/22 facts, 150+ tools, adds betalfil, Peppol, skattekonto and the Claude plugin. - docs/PEPPOL_FOUNDATION.md: the two sentences denying network delivery and inbound support now describe the live Qvalia path. - docs/SELF-HOSTING.md, docs/DOCKER.md: clone URLs and directory names, Sentry DSNs are not read by the app, image pinning uses the 7-char SHA tags the workflow actually publishes (no semver tag has been cut). - docs/EXTENSIONS.md: replaces the fictional sector tree with the 19 real extensions/general directories; lib/reports/sru-encoding.ts. - .claude/rules/database.md: 680+ migrations, ~170 live tables, adds the tables and RPCs that matter since July, drops sandbox_users. Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Jakob Wennberg
|
188816652d |
docs(api,mcp): tool counts, changelog backfill, version-header honesty, lazy auth, endpoint map (#1929)
Brings the developer-facing API and MCP docs back in line with origin/main (audit 2026-08-26). Docs only; no runtime behaviour changes. - Tool counts: the server registers 153 tools; docs said 90+/100+/120. All now say "150+" (connect-claude, gnubok-mcp README, plugin README, mcp-server rules, CLAUDE.md, registry entry with refreshed updatedAt). Not derived from the tools array: lib/ must not import @/extensions/. - REST changelog: backfilled the additive 2026-08 changes (#1909 report date ranges + PDFs, #1864 POST /companies, #1773 vat-declarations, #1405 PATCH settings, #1724/#1788 customer personal_number, #1809 cash_account_id filter). API version date unchanged. - Version headers: Gnubok-Deprecation is planned, not emitted; the Gnubok-Version request header is not read today (version.ts comment, versioning page, conventions overlay, regenerated skills/accounted-api). - connect-claude Path A documents lazy auth (connector works before an account exists; sign-in on the first company-scoped call). - MCP server README: real Anthropic SDK call sites, real resource URIs, pending-operations widget, public-tools/tasks/origin-guard/pii-guard. Rules file gains Lazy auth + feedback/tasks paragraphs. - api-routes endpoint map regenerated from the filesystem (560 routes, 55 families incl. v1, agent, reconciliation account-keyed, dimensions, peppol, rot-rut, webshop-orders, mileage, billing, skatteverket, receipt-hunt). - gnubok-mcp/accounted-mcp: /settings?tab=api is the pre-redesign URL; now /settings/api (README + help hints, no version bump). Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Jakob Wennberg
|
b8605aabfc |
fix(settings): derive API-key scope groups and tool counts from the scope catalogue (#1924)
The API-key settings panel carried a hand-copied list of scope groups that
had drifted to 24 of the 30 scopes in API_KEY_SCOPES: articles:read/write,
companies:write and the three reconciliation scopes were missing, so a key
minted in the dashboard could not call gnubok_create_company, the article
tools, the seven reconciliation tools or the matching v1 endpoints. The
per-scope "N verktyg" counts in the panel and in the API_KEY_SCOPES
descriptions were hand-maintained and wrong (reports:read said 18, actual
30; bookkeeping:write said 11, actual 22).
- Move the pure scope catalogue (API_KEY_SCOPES, scope lists, SCOPE_GROUPS,
TOOL_SCOPE_MAP) into lib/auth/scope-catalog.ts with no server imports, so
the client-side panel can bundle it. api-keys.ts re-exports everything,
so existing imports are unchanged.
- SCOPE_GROUPS becomes a list of { domain, label, scopes } covering every
scope (reconciliation has three), shared by the panel and the OAuth
consent page. scopeKind() replaces the ad hoc suffix checks.
- TOOL_COUNT_BY_SCOPE is derived from TOOL_SCOPE_MAP at module load; the
hand-written counts are removed from the catalogue descriptions.
- The panel renders groups and cards from the catalogue; i18n keys are
derived from domain and scope id. The "(REST API)" heading suffix is
computed from the counts instead of baked into the labels.
- New unit test asserts every scope belongs to exactly one group and that
counts equal TOOL_SCOPE_MAP occurrences.
- New sv/en strings for the articles, companies:write and reconciliation
scopes.
Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
|
||
|
Jakob Wennberg
|
f338850bd0 |
fix: hide API-archived customers and suppliers from lists and pickers (#1927)
* fix: hide API-archived customers and suppliers from lists and pickers The v1 API soft-archives customers and suppliers (archived_at, plus is_active=false on suppliers) and its own list routes hide those rows behind ?include_archived=true. No other surface filtered archived_at, so an archived counterparty stayed a normal row in the dashboard rosters, the internal /api/customers and /api/suppliers list routes, the MCP list tools and every customer/supplier picker. Apply the same canonical `archived_at IS NULL` filter on every non-v1 list and picker path: - /api/customers GET, /api/suppliers GET (feeds the customers page and the supplier-invoice form) - suppliers dashboard page (reads suppliers via browser Supabase) - InvoiceEditor and NewRecurringScheduleDialog customer pickers; an invoice or schedule being edited keeps its current customer visible (archiving does not refuse on drafts, so a draft can point at one) - deadlines page and CalendarWorkspace customer pickers - InvoicePreviewCard sample customer - gnubok_list_customers and gnubok_list_suppliers: hidden by default, optional include_archived boolean mirroring the v1 flag; rows now carry archived_at so an agent can tell them apart when opted in Detail routes and by-id lookups are untouched: an archived row still opens. The delete-vs-archive semantics are unchanged. The tools/list payload guard moves 60.7K to 60.8K: main had ~6 tokens of headroom, so even the bare boolean contract crossed. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * test(schema): raise the unresolvable-expression ceiling by 2 for the archived-customer picker filters The two .or('archived_at.is.null,id.eq.<uuid>') filters keep an edited draft's archived customer selectable. The uuid is a runtime value, so the scanner cannot resolve the expression; both columns exist and the filter is covered by the archived-counterparty tests. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Jakob Wennberg
|
1185ab4294 |
fix(mcp): honest tool text and build-derived server version (#1923)
Tool text that lied to agents: - gnubok_create_voucher pointed at gnubok_reverse_entry, which does not exist; the tool is gnubok_reverse_journal_entry. A scan of server.ts, skills/, prompts/ and structured-errors.ts found no other phantom names. - gnubok_reverse_journal_entry said reversal_date defaults to today; the executor passes undefined and reverseEntry() uses the original entry date (same as the dashboard). Description now states that. No behaviour change. - gnubok_get_vacation_balance promised an estimated semesterloneskuld in SEK but returned none. The tool now returns estimated_liability_sek using the same BFNAR 2016:10 day valuation as the year-close and the v1 vacation-balance route (dayValueSek exported from semesterberedning), floored at zero for overdrawn balances. Descriptions trimmed so the tools/list payload stays under the 60.7K-token ceiling (60,696 after). - gnubok_create_invoice said the invoice number is assigned at approval; it is assigned on send or mark-as-sent (ensureInvoiceNumber). - gnubok_convert_invoice: "har redan makuleras" -> "har redan makulerats". - lib/entitlements/keys.ts comment claimed bank_sync has no MCP tool while the map right below gates gnubok_connect_bank on it. Version: MCP serverInfo.version, the extension version and /api/health all hardcoded '1.0.0', so clients could not tell deploys apart. They now share currentAppVersion() (commit SHA prefix inlined at build), resolved once at module load so the definitions layer stays deterministic, with '1.0.0' as the self-hosted fallback so Docker healthchecks keep a value. serverInfo is not part of tools/list, so the catalog payload is unaffected by this part. Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |