Repository navigation
Added
- A corpus of bank statement shapes, contributed rather than guessed.
fixtures/statements/<country>/<bank>/
holds an invented statement and a note of what the file says;fixtures/validate.pychecks one before it is
sent, andengine/tests/test_statement_corpus.pyreads every one of them on every build, so a shape that is
read today keeps being read while the reader is changed for somebody else's bank. Two worked examples are
included: money-in/money-out columns under a letterhead, and a single signed amount column with the minus
written after the figure. Contributing one needs no code — see fixtures/README.md. - Two issue forms for the two reports that matter most, each asking for what is actually needed to fix the
problem: A bank statement is not read correctly (the top of the file with invented figures, and what the
file says) and A figure is wrong (the account, the period, both figures, and where yours comes from). - A ten-minute way in at the top of Run it: install, import the Harbor Line demo company's 14 months of
invented books, and follow a figure down to one journal entry. - A corpus of charts of accounts, contributed by the accountants who use them.
fixtures/charts-of-accounts/<country>/<slug>/holds a chart with the account class in the words that
country's systems write and the role each account should get as an accountant judges it;
engine/tests/test_chart_corpus.pyholds the classification to it on every build, and
fixtures/charts-of-accounts/validate.pyshows a contributor what the product makes of their chart before
they send it. Contributing one needs no code — see
fixtures/charts-of-accounts/README.md.
A fixture can record a gap — a class word the product does not know, or a role it suggests wrongly — with
a reason, instead of bending the expectation to match the code. A recorded gap is asserted to still
reproduce, so when somebody extends a word list the build fails and says to delete it. The two charts
included carry one gap each, both found by writing them: a UK loan written (over one year) is suggested as
short-term debt, because the long-term words are «long-term», «long term» and «non-current»; and an
account class written in Arabic is not recognised at all (the ledger import then refuses the row with
"unknown account class"), becauseCLASS_ALIASESis English only while the role words beside it are in four
languages — the same is true of French and Spanish.
Changed
-
The account class column is read in four languages, not one. A ledger arrives with its class column
written in the words the customer's own system uses — Debtors, Creditors, «الإيرادات», Proveedores,
Les charges — andCLASS_ALIASESheld twenty English words, while the role words beside it
(analysis/roles.py) have always been in English, Arabic, Spanish and French. An unrecognised class word is
not a degraded suggestion: the import refuses that row with "unknown account class", so the account never
reaches the statements or the ratios. In a product whose interface and documentation are Arabic first, an
Arabic class column was refused row by row.The table is now
CLASS_WORDS, keyed by the canonical class, with each of the eleven classes written the way
all four languages write it — 181 words, normalised on the way in the same way the role words are, so
«الإيرادات» and «الايرادات» are one word and Amortización needs no second spelling. The definite article is
read as a second reading: Arabic joins it to every word it governs, so it comes off each one and
«الالتزامات المتداولة» reaches «التزامات متداولة»; Les charges and La caisse lose theirs from the front.
A word that two classes could claim goes to the narrower reading or stays out — bare dettes and deudas
are a liability while dettes fournisseurs and proveedores are the payables control account — and a word
that names an account rather than a class stays out altogether, so «مخزون» and existencias are still
inventory accounts of class Asset rather than a class of their own.engine/tests/test_cost_centres.pystates the four words for each of the eleven classes in a list written
independently of the table, fails if a class is added without them, fails if any word is claimed by two
classes, and reads an Arabic ledger end to end. A word nobody knows is still unknown: the row is named with
its number so a person can correct the file, because guessing is worse than refusing. The class-word gap
recorded infixtures/charts-of-accounts/sa/trading-companyis deleted, which is what the corpus test asked
for in so many words.The ledger template's own Guide sheet now says so, where somebody filling the file in will read it, and the
monospace block on that sheet follows the list of field lines instead of a hard-coded row count — which is
what hid the three new lines in the first draft of this change. -
The product's owner is named: OFIN INC, 8 The Green, Suite B, Dover, DE 19901, United States. The
copyright holder inNOTICEwas «The OfinTech Desk authors», which names nobody and leaves the marks
belonging to no one in particular. The entity now appears inNOTICEwith its registered address, in the
installer metadata the operating system shows before anyone runs the file, in the published package
metadata (@ofintech/embedded-agent, the engine, the desktop app and the UI), and in the website footer
beside the Riyadh operation it runs from. "OfinTech", "OfinTech Desk" and the OFINTECH logo are stated as
marks of OFIN INC, which is what section 6 of the licence leaves out of the grant.LICENSEis deliberately untouched: the copyright notice at the end of the Apache-2.0 text is the
licence's own template, not a field to fill in, and the real notice belongs inNOTICE.
Added
-
/readysays whether the engine can actually work./healthhas always answeredokwhenever the
process was running, which is the right answer to the question it asks and the wrong one to rely on: a
container on a full disk, a deploy whose migrations did not run, a bucket whose policy changed after
deployment and a background loop stuck on anawaitall answer it perfectly, and none of them fails a
request until somebody's work is already lost. The new route runs five checks — the schema version against
the build in both directions, a write transaction taken and rolled back, free space on the data volume,
the object store, and how long since the background scheduler last came round — and answers503when the
engine should be taken out of rotation,200withdegradedfor a warning./healthis unchanged and
stays what a restart policy points at: nothing/readyreports is fixed by restarting, and a schema
mismatch or a refusing bucket would turn a restart policy into a restart loop.The route is unauthenticated, because a container probe has no credentials to offer, so what it says is
public and is treated that way: a status and a reason code from a closed set, with no path, version or
free-byte count anywhere in the body. The numbers go to the log, once each time the overall state changes.
One evaluation serves two seconds, so holding down a reload cannot make the engine hammer its own object
store. It does not replace an outside monitor — a process that is down answers nothing at all, and no
internal check can report its own absence. -
Every document search records what it cost. One line per search: how many vectors were compared, whether
the matrix was already in memory, the keyword and semantic phases separately, the total, the memory the cache
holds and the resident memory of the process. A search slower than 500 ms is logged as a warning rather than as
routine information. This is what makes the next decision about retrieval a measured one: the thresholds worth
watching are chunks per workspace, process memory and the slow-search warnings, and nothing in the product
could report any of them. The query, the matched text and the document names are deliberately not recorded — a
search line is written on every lookup, including the ones an agent makes over a customer's ledger, and a log
is the easiest place to spill the contents of the thing being searched. -
The app can be installed from the browser. Chrome and Edge now offer to install Desk from the page, Safari
can add it to the Dock, and the result is an icon and a window of its own with no browser chrome. It matters for
who can actually try the product: a finance lead on a managed laptop often cannot run a downloaded installer,
and an unsigned one asks them to step past Gatekeeper or SmartScreen before they have seen anything. This is the
same page over the same HTTPS, so there is nothing to sign, notarise or warn about.e2e/installable.pychecks
the parts that each fail quietly — the manifest linked and served as a manifest, icons at 192 and 512, a
maskable icon whose corners are opaque so a cropping platform does not cut through the mark, and a
Content-Security-Policy that still permits fetching it. -
What the model calls cost (Settings → Usage). The product has recorded every token it spends since the first
release and has never been able to say what they came to. Rates are entered per workspace, per million tokens, as
the provider quotes them, and each applies from the day it takes effect — so entering next month's price does not
re-cost last month, while correcting a rate typed wrongly does, with the audit log keeping what it was. Cached
tokens are billed at their own rates rather than as input, which the token column alone would get wrong by more
than five times on a workspace that caches well. Usage with no rate behind it is reported and left out of the
total instead of being counted as free, and currencies are kept apart and never converted. Nothing is costed
until someone enters a rate: no prices ship with the release, because a stale one produces a confident wrong
figure, and bundling a vendor's list would assume which vendor you use. -
A 90-second brand film (
docs/brand/film/): the figure, the question behind it, the answer with its source,
and the three limits that make the answer worth having — type and motion on the brand field, no screen
recordings, so it does not go stale when the app changes. 1920×1080 at 30 fps with a quiet bed at about
−21 LUFS, and a voiceover script timed to the picture for whoever records one.film.htmlis the whole film as
one timeline (seek(t)paints the frame attand nothing else depends on time, so renders are identical);
frames.pypipes frames into ffmpeg a chunk at a time;music.pywrites the bed. -
The public website (
site/): four static pages written for a finance lead rather than an engineer — what the
assistant answers, the four steps to using it, what it will not do (no payments, nothing outbound without a
named approval, sealed mode, an audit log the application cannot rewrite), and how to get it. Everything
technical moved to a singledocs.html. Sign-in and sign-up pages post to the engine's existing
/auth/form/sign-in,/auth/form/sign-upand/auth/form/google, and the script turns the error codes the
engine redirects with (#signup-exists,#signin-rate, …) into a message on the page that error belongs to.
Arabic is the source of truth with English indata-enattributes, one file per page,html[lang]and
html[dir]switched together. No build step, no framework, no third-party request; the brand's own fonts and
palette.e2e/site_check.pydrives four pages × two languages × five widths in CI and fails on any sideways
overflow, an unswitched string, a form that posts anywhere but the engine, or a console error.
Removed
- PipesHub. The optional external search engine is gone: keyword search (SQLite FTS5, BM25), semantic search
through the workspace's own embedding model, reciprocal-rank fusion and OCR with Tesseract are all built in, so it
added an outbound path, a stored credential and a live integration to prove, for a capability the product already
had. Migration 22 dropsdocuments.pipeshub_status,documents.pipeshub_record_idand
doc_collections.sync_pipeshub, and removes any PipesHub connection and its secret. Documents left in the
remotestatus (readable only there) becomefailedwith a message asking for a searchable copy. The
Connections → Knowledge section and the PipesHub card are gone from the app.
Fixed
- The Team Server image no longer ships a vulnerable
apt, and the build that produces it is no longer the
place to attack. CVE-2026-109518: apt's handling of a resumed HTTP download could be made to merge two
downloads when a redirect gave them the same address, which let an on-path attacker substitute a package
after it had passed verification and run code as root. Debian published the fix as apt 3.0.4 for trixie on
4 November; the image was built on 3.0.3, so the Trivy gate in CI — which fails a build on any HIGH finding
that has a fix available — stopped the first build that ran afterwards. The image now upgrades apt, and the
rest of the base image with it, before installing anything, so it carries the operating-system fixes Debian
has already published instead of whichever ones its base layer happened to be built with. Retries are
switched off first, because the apt that fetches the fix is still the vulnerable one and with no retry there
is no resumed download to merge — the mitigation the advisory gives for a build whose package sources are not
yet HTTPS. - A workspace no longer reports nothing as embedded after its embedding settings are saved again. Saving the
settings re-queues the workspace, and queueing cleared the embedded flag on every document before walking each
one to embed it. A document whose chunks already carried a vector from that same model had nothing to embed, so
the walk returned without putting the flag back: semantic search went on working while Knowledge → settings
reportedembedded: 0, and on a large workspace that wrong answer cost a pass over every document. Re-queueing
now selects only the documents whose chunks are not already embedded with the current model, so saving an
unchanged setting queues nothing, and embedding a document that needs no work restores its flag. - The cache of embedding matrices has a ceiling. Semantic search holds a workspace's vectors in memory as one
float32 matrix, which is what makes the search a single multiplication rather than thousands of reads. There
was one such matrix per workspace and model, kept for the life of the process, with nothing to evict it: a
server with many workspaces could exhaust its memory without any single workspace looking large, and a deleted
workspace went on holding its matrix until a restart. The cache is now least-recently-used and bounded by
DESK_KNOWLEDGE_CACHE_MB(512 MB by default,0for no limit), and deleting a workspace releases its matrix.
A workspace whose own matrix exceeds the budget is still kept — rebuilding it per query would cost far more
than the memory — and logged, because that is the point at which it has outgrown an exhaustive scan. - The macOS installers build again.
@electron/rebuildran on every desktop package, searched for node
modules, found none — the shell has no runtime dependencies andfilesships withoutnode_modules— and on
macOS stopped the build with<projectDir> not a file. Windows was unaffected, which is why it went unnoticed:
thev1.1.0tag produced no release because the release job waits for installers from all three systems, and
desk-v1.0.0failed the same way on macOS while succeeding on Windows.npmRebuild: falsein both builder
configurations removes a step that had nothing to do and could only ever fail. - The website at
ofintech.aiis now the one insite/. The four-page site — the product and who it is for,
a separate documentation page, sign-in and create-account, Arabic and English on every page — existed in the
repository but was wired to nothing: the domain still servedlanding/, a single English page with no language
switch.site/now carries its ownDockerfile,nginx.confand compose files, andlanding/is gone.
Addresses written down when the site was one page still arrive:/featuresredirects to how it works, and
#signinand#signupare sent to their pages bysite.js, because a fragment never reaches the server.
/docsand/sign-inwork as well as/docs.html. There is a 404 page in the site's own chrome.
The Content-Security-Policy had to change with it: the inherited one saidscript-src 'none', which was right
for a site with no script and would have switched off the language switch and pointed the sign-in form at
itself.e2e/site_serving.pyserves the pages under the directives read out ofnginx.confand fails if the
policy blocks the site or an old address stops arriving; it fails on the inherited policy. - A UI change now reaches a browser that already loaded the app. The build wrote the bundle at a fixed name,
assets/app.cssandassets/app.js, while the engine served everything under/assets/with
cache-control: public, max-age=31536000, immutable. A browser holding an immutable response does not revalidate
it, so every browser that had ever opened Desk kept the stylesheet and script it first downloaded — for a year,
past anygit pull, image rebuild or restart. The interface fixes in this release were invisible to returning
users for that reason. The build now writes the content hash into the bundle name and rewritesindex.htmlto
point at it, and the engine sendsimmutableonly for a name that carries a hash;index.htmlitself stays
no-cache, so it is re-fetched and hands the browser the new URLs, which also clears the stale copies already
cached. Builds made with--watchkeep plain names and are servedno-cache.e2e/bundle_cache.pyfails the
build if the bundleindex.htmlpoints at could be pinned at an unchanging URL, and it fails on the old build. - Money figures no longer leave the box that holds them. A figure in the millions or billions painted over the
KPI cell border into the card beside it, and was clipped inside dashboard widgets and the ledger totals. Each of
those cells is now its own query container and the figure's type size follows the cell's width, with a floor; the
dashboard grid, which was three fixed columns at every screen width, and the ledger totals now reflow instead of
squeezing.e2e/figures_fit.pyrenders the three patterns at eight widths in English and Arabic (RTL) and fails
if any figure is wider than its box; it runs in CI.
Added
-
Embedded Agent Gateway + AG-UI adapter (docs/embedded/): partner applications can run Desk's financial agent
inside their own UI over the AG-UI protocol. A workspace admin registers an integration (allowed origins,
capabilities, limits); Desk creates a dedicated viewer service user for it and an integration key for the
partner's backend, which mints 15-minute session tokens per end user. Workspace, acting user and role always come
from the integration, never from the request; client-supplied tenant/user/role fields are ignored and audited.
Runs are limited to a read-only tool scope (no write tools, nothing that leaves the network), stream AG-UI
events over SSE (POST /api/embedded/agents/{agent_id}/run), and map Desk approvals to AG-UI interrupts that can
only be decided inside Desk by an editor. Generative UI through a validated, versioned component registry (ten
financial components, ar/en, RTL) — the model never produces UI. Usage metering per integration and an
entitlement hook (no billing). Migration 21. React SDK@ofintech/embedded-agent(sdk/embedded-agent) and a
CopilotKit example (examples/copilotkit-embedded). 53 engine tests (tests/test_embedded_*.py), 12 SDK tests,
browser verificatione2e/embedded_browser.py. Threat model with 20 threats in
docs/embedded/SECURITY-THREAT-MODEL.md. Not a production-readiness claim. -
Agent.start_turn(..., state_extra=)andstate["tool_scope"]: a run can carry a tool scope that narrows the
tools it is offered; absent, nothing changes for the Desk UI, workflows, channels or MCP. -
The members API refuses to promote an embedded integration's service user above viewer (
service_user_role). -
Executive overview (docs/OVERVIEW.md): the landing page for every workspace. A financial health score with its
reasons (judged against the finance lead's targets, or reference levels until targets are set), revenue, expenses,
net profit and cash with a plain-language note on each, the twelve-month trend, cash and liquidity, risks that need
attention, what the numbers say (facts from the ledger, plus observations the workspace's model writes on request,
always labelled AI estimate), and what is waiting for a decision (approvals, forecasts to sign off, open closes,
budgets running out early). Every figure carries its provenance: actual, unverified (account roles not yet
confirmed), budget, forecast or estimate. -
Drill-down from any executive figure: profit and loss → the accounts behind a line → the ledger lines of one
account in the month → one journal entry with its source dataset and audit trail. The reporting month travels in
the address, breadcrumbs lead back, and the viewer role is checked at every level (/api/w/{ws}/overview…). -
Statements, for the CFO (docs/OVERVIEW.md): monthly profit and loss, balance sheet and cash-flow statement from
the ledger (the plan in grey when one is approved); budget against actual by line with favourable or adverse
variances and the accounts that drove the overruns; receivables and payables by customer or supplier, aged 0–30,
31–60, 61–90 and 90+ days with settlements applied oldest first and a check against the ledger balance; anomalies
(unusual amounts, missing regular postings, first postings, possible duplicates) above a materiality threshold.
Every table opens into the same drill-down as the overview (/api/w/{ws}/overview/statements|budget|parties|anomalies). -
General ledger, for the accountant (docs/OVERVIEW.md): journal entries searched and filtered by month, text,
type, account, party, amount or balance, each opening with its lines, roles, source and audit trail; the chart of
accounts with roles, activity and month-end balances, suspense accounts flagged; and validation the way a
reviewer would do it (entries that do not balance, single-line entries, missing accounts or dates, future dates,
both sides on one line, zero lines, suspense balances, unconfirmed roles, balances on the wrong side, postings on
the unusual side), for a month or the whole ledger (/api/w/{ws}/ledger/entries|accounts|validation). -
Oversight, for the accounting manager: close and reconciliation status by month, unmatched bank items,
pending approvals, exceptions and anomalies, team activity from the audit log, documents, operational measures
and reporting progress, every item opening the page where the work is done (/api/w/{ws}/oversight). -
Three views of one application: Executive, Finance and Accounting change what the navigation rail shows; the
default follows the member's role (viewer, admin, editor) and the person can switch. Permissions are enforced by
the server whatever the view.
Changed
- OfinTech identity (docs/DESIGN.md). The app and the website now follow the OfinTech design system: Deep Navy
(#14243A) as the brand foundation — the navigation rail and dark surfaces — Primary Blue (#2563EB) as the one
interactive colour, white and cool neutral surfaces, and the four semantic colours for success, warning, error and
information. Colours are defined once as brand primitives, mapped to semantic tokens (--color-*) and from there
to the names the components use, so the palette is maintained in one place. Latin text and figures are set in
Inter (bundled, SIL Open Font License; no font download), Arabic stays in IBM Plex Sans Arabic; financial
figures use tabular numerals everywhere. Tables lost the green-bar banding for a quiet zebra on the muted surface.
Dark mode keeps the same palette on navy surfaces. Value provenance badges (actual, forecast, budget, estimate,
unverified) are part of the design system. The website's product screenshots show the new identity, with the
executive overview of the demo company as the opening image.
Security (enterprise audit, docs/audit/)
- The Zoho sign-in callback accepts only Zoho's own accounts servers and API hosts, matched exactly.
- The CSRF guard compares the whole request origin with the server's, never a prefix.
- Every CSV and workbook the engine writes keeps imported text as text (
= + - @, tab and carriage-return
prefixes), including the month-end close journal and pack. - Request body limits hold whatever
Content-Lengthsays, on every public endpoint and the JSON API. - An open event stream closes within one keep-alive when the session ends or the member is removed.
- Team Server: outbound requests are judged against the private network at every connection
(egress.PrivateNetworkGuard); shared address space (100.64/10) is refused; the S3 endpoint and
DESK_EGRESS_ALLOWhosts are exempt. - Behaviour change: a stored file without the encryption header is refused unless
DESK_STORAGE_ALLOW_PLAINTEXT=true;python -m desk_engine storage-encryptrewrites such files once. - Office files that would inflate beyond 400 MB or 200:1, images outside raster formats or above 60 MP, and OCR
runs past 15 minutes are refused or cut short before a parser spends the memory; Caddy caps bodies at 260 MB. - The bank-identifier masker no longer masks the engine's own ids (
doc_…,ds_…). - A custom skill written by an editor stays off until an admin turns it on, and cannot take a built-in's name.
- Behaviour change: importing an e-mail attachment or a Drive file into the workspace and editing a person's own
Excel or Word file now wait for approval like other write actions; refreshing accounting datasets and creating a
new workbook or report still run without asking. - Behaviour change: an MCP tool set to run without asking does so only in the form an admin approved; after a
refresh shows it changed, or after this upgrade, it asks until an admin saves its policy again. - Tool arguments are validated against each tool's schema before a handler runs; every tool call has a time limit.
DESK_SMTP_SECURITYacceptsstarttls,sslornoneonly, andnonecannot be combined with a login.- Audit entries for MCP servers keep the URL without its query string; Telegram transport errors never carry the
bot token; formula parsers refuse deep nesting and absurdROUNDdigits cleanly. python -m desk_engine backupwrites an AES-256-GCM.desk-backupunder a separate backup key
(DESK_BACKUP_KEY_FILE);restoreauthenticates and integrity-checks a backup before writing it;backup-check
is a restore rehearsal.backup.shuses them.backup --plainkeeps the old unencrypted copy for one-off use.- CI: read-only workflow token, a
securityjob (gitleaks, ruff security rules, pip-audit, npm audit) and a Trivy
scan of the Team Server image gate the image and the release; the engine container drops all capabilities;
npm dependencies install from committed lockfiles; the desktop's IPC sender and link rules are unit-tested. - Webhook payloads shed credentials under any key that looks like one (
client_secret,refresh_token,
Authorization,X-Api-Key, …) and the query string of any URL value, before storage and the model. - A workflow run acts with its creator's rights only while the creator is still an active member of the workspace.
- The oversight desk and the audit activity it shows need the admin role; the executive overview's forecast and
dataset lookups are scoped to the workspace. - Export cells that begin with an invisible character before a formula prefix are kept as text too.
- Behaviour change: sessions on a Team Server end after 30 minutes of inactivity (was 60; configurable 30–45)
and after 12 hours whatever happens — 8 hours for admins and owners (was never;DESK_SESSION_HOURS,
DESK_ADMIN_SESSION_HOURS). High-risk actions (roles, connections, MCP servers, model providers, closing a
month, deleting data, approving an assistant action that leaves the workspace, …) ask for the password again
unless it was entered in the last 10 minutes; people can end their other sessions, and the instance
administrator can end anyone's. The page warns two minutes before with Stay signed in /
Log out now, signs every tab out together, and re-checks the session after a sleep or a reconnection. The
desktop app locks its screen after the same window and when the computer locks or sleeps. - Approvals expire (4 hours for actions that leave the workspace, 24 hours for edits of the person's own files),
are bound to the exact arguments shown, are consumed exactly once, and die with the requester's membership or
editor role. A pending approval older than that asks again. - The audit log carries an event id, a correlation id, the decision, the approval and the outcome on every entry;
entries cannot be changed or deleted; every tool call, refusal, 403, cross-tenant attempt, failed sign-in,
logout and expired session is recorded. - Behaviour change: the assistant of a viewer no longer creates dashboards, workbooks, reports, editable file
copies or month-end close runs, and does not refresh accounting datasets: these tools now need the editor role,
like the pages that do the same thing. Editors and above are unaffected.
These installers are not code-signed
macOS will refuse the app outright — “OfinTech Desk” is damaged and can’t be opened. You should move it to
the Trash. It is not damaged: Apple silicon will not run an application that carries no signature at all. Move
it to Applications, then clear the quarantine flag macOS puts on anything downloaded:
xattr -dr com.apple.quarantine "/Applications/OfinTech Desk.app"
Windows SmartScreen warns instead of refusing: More info → Run anyway.
Only Apple silicon is built for macOS; this release has no Intel build. What removes all of this is a signed and
notarised build — see docs/RELEASING.md.