Three reports from somebody driving the app through the MCP connector,
all the same shape: an argument accepted, an OK returned, and nothing
done.
**project_id on an update never moved the task.** Both the REST handler
and the MCP tool read it, both checked the caller's rights on the
destination — which reads as "understood" — and then neither passed it
on, because store.TaskUpdate has no such field: a project change is a
move, since sort_order is per project. So the task stayed where it was
and the answer said OK. Moving a task between projects was impossible
except by creating it again and closing the old one, which is what the
person ended up doing. Both paths now go through one helper, so they
cannot drift apart again.
**An unknown #project vanished without a word.** The parser reads
"#3Dekoration" correctly; the lookup then fails, the task goes to the
Inbox, and the name is gone from the title with nothing anywhere saying
why. Still not an error — a thought should not be thrown away over a typo
— but quick-add now returns unknown_project, and the MCP tool returns a
note written for a model to relay. The old behaviour was kind in the web
interface, where you see the row appear in the wrong place and fix it. In
a chat nobody sees the row.
**Sub-tasks were unreachable over MCP.** create_task had no parent_id,
though the REST API has had it all along, so importing a parent with
seven children produced eight unrelated tasks with the parent's name
prefixed onto each. A sub-task joins its parent's project — the two
cannot disagree — and a second level is refused rather than silently
flattened.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Projects have carried sort_order since the first migration and nothing
ever set it, so the sidebar was ordered by creation date for good.
POST /projects/reorder takes the whole list rather than one move. A
person has a handful of projects, so sending the order you want is
simpler to reason about than midpoints between neighbours — and it cannot
run out of precision, which is the failure the task ordering has to
respace a section to recover from. It also cannot land half-applied the
way a sequence of individual moves can.
Projects you do not own are skipped rather than refused. sort_order is a
column on the project, not a preference per viewer, so a shared project
sits where its owner put it and a member rearranging their own sidebar
does not rearrange somebody else's. The sidebar only makes your own
projects draggable, so that rule never shows up as a row that springs
back.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Adding the connector to Claude failed with "Couldn't register with
Verdande's sign-in service", on a server that has no sign-in service and
needs none.
Claude probes /.well-known/oauth-authorization-server before connecting to
an MCP server. The SPA fallback answered 200 with a page of HTML — which
reads as "yes, there is an authorization server here" — so it went on to
dynamic client registration and failed there instead of concluding the
obvious thing, which is that this server does not use OAuth and the key in
the URL is the whole credential.
/.well-known/ is reserved by RFC 8615 and is never a route in this app. A
404 is the true answer, and it is also the useful one: the client stops
asking and uses the URL as given.
This is the third time the catch-all has caused a confusing failure by
saying 200 to something that does not exist — /mcp itself was the second.
The fallback should hand back the shell for routes the app might have, not
for every path on the server.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A project could be created and shared and never renamed or deleted. Both
have worked through the API since the beginning; there was simply nothing
to click.
The heading is now the field — a heading that happens to be editable
rather than a form that looks like one — and saves on blur, with Escape
putting the old name back. Owner only, and the Inbox keeps its plain
heading because it cannot be either renamed away or deleted.
Deleting needed more than a button. It is soft, with thirty days to
change your mind, and restoring has been possible the whole time — from
an id, which is exactly what you no longer have once the project has gone
from the screen. A recovery window nobody can reach is not one. So
GET /trash/projects lists what you have deleted, with how many tasks went
with it and how long is left, and Settings → Data has the button.
Owned, not shared: a member of a project somebody else deleted has no
business bringing it back, and there is a test that says so.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
"I have updated and restarted, but I cannot tell whether it worked" is
not a question the panel could answer. The version was on the startup
line and in the settings page, and nowhere you can reach from outside.
/healthz carries it now. That is the one endpoint answering without a
session, which makes it the only way to ask the question from a terminal
or from the panel itself. The "listening" line repeats it too, because
that is the line the Rune's done_regex matches and therefore the one
sitting at the bottom of the console after a deploy.
The update notice could also never have worked. It reads GitHub's
releases/latest, and this repository had only tags — a tag is not a
release, and that endpoint answers 404 for a repository without any. The
release workflow creates one now, after the image is pushed so a release
never announces a version that failed to build.
VERDANDE_UPDATE_CHECK is a Rune variable, so it can be switched on
without editing the container. Still off by default: it is an outbound
request, and that is a choice to make rather than a default to inherit.
Rune version 1 -> 2, so the panel offers the update.
Note that the check still cannot see anything while the repository is
private — releases/latest is 404 anonymously, exactly as the GHCR package
was. Making the repo public, or teaching the check to carry a token, is a
separate decision.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two bugs with one shape: the interface confidently told you something
that was not so.
Connecting Gmail without a registered OAuth client said "Det er der
allerede." Thirteen different situations answered a plain `conflict`, and
the frontend renders that as "that already exists" — which for four of
them is not vague but wrong. An unconfigured Gmail is not a Gmail that is
already connected; a second factor that is off is not one that is on; the
Inbox is not a project that already exists somewhere else. Those four
have their own codes now, and their own sentences.
The mail-to-task address came out as todo+…@localhost. VERDANDE_SMTP_FROM
defaults to verdande@localhost and the domain was taken from it
unconditionally, so an instance with no mail server handed out an address
that looks real, can never receive anything, and gives no hint which of
the two it is. The default no longer wins over the public hostname, and
the response now carries whether a mail server exists at all — because
even with the right domain, nothing routes to it until one does.
errors.go has always claimed the code list is kept in one place "so the
frontend's mapping can be checked against it". Nothing checked it, and
the mapping was missing bad_request. Now a test reads both files and
fails on a code with no message.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
verdande could not be added to Claude at all. The documented setup —
POST /api/v1/mcp with an Authorization header — cannot be configured from
Claude's custom-connector dialog, which asks for a name, a URL and OAuth
client credentials and has no field for a bearer token. The instructions
described something the interface does not offer.
Worse, the obvious guess failed silently in the most misleading way
available: /mcp was not a route, so the SPA fallback answered 200 with a
page of HTML. The connector reported a successful connection and then
failed to parse the response, which reads as a broken client rather than
a missing endpoint.
So /mcp exists now and takes its credential in the query string, exactly
as the calendar feed does and for the same reason: a client that cannot
be told to send a header means the address is the credential. The
request logger records paths without their query strings, so the key does
not reach the log.
It is deliberately outside /api/v1 and accepts the key and nothing else.
Reading a session cookie here would make it a state-changing POST outside
the CSRF check — a cross-site request that acts as whoever is signed in —
and there is a test that fails if that ever becomes true.
The API path is unchanged for anything that can send a header. The tokens
page now shows the connector address next to the token it belongs to,
since that is the one moment the token exists.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The brief asked for one. docs/api.md is complete and hand-written, and
there was nothing a tool could read.
Written by hand rather than generated from annotations, which would have
meant every handler carrying a second copy of itself in a comment. The
risk with a hand-written spec is that it goes stale, so it does not
depend on anybody remembering: openapi_test.go walks the chi router and
fails both on a route nothing describes and on a description with no
route behind it. CI separately checks it is a valid OpenAPI document.
The test reads the file with a forty-line scanner rather than pulling in
a YAML library. It needs two things — which paths are listed and which
methods each has — and both sit at fixed indentation in a document the
test also owns, which is worth more than a dependency in a project whose
whole shape is one static binary.
The CalDAV routes are excluded deliberately: PROPFIND and REPORT are
outside what OpenAPI can describe, and docs/caldav.md covers them.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
docs/api.md has told people to make a token under Settings since it was
written. The store has held CreateAPIToken, ListAPITokens and
DeleteAPIToken since the first migration. There were no routes: nothing
between the two, and no way to make one.
GET, POST and DELETE under /tokens, all behind a new requireSession
middleware. A bearer token is refused there with 403 even though it is
accepted everywhere else — otherwise a leaked token is permanent, because
whoever has it issues a second one and revoking the first changes
nothing. The plaintext comes back from POST and never again; only the
hash is stored, so the interface has to say so at that moment rather than
offering a reveal that cannot work.
PATCH /auth/me alongside it. Name, timezone and language could not be
changed anywhere — not through the API and not through the interface —
and an account created from an invite inherits a default timezone. That
is not cosmetic: every date in the app is resolved in it, so somebody in
another country had every due date land an hour out with no way to say
so. The zone is validated against the running system's database rather
than a list kept here, which is the same database the dates are resolved
with. The email address stays read-only: it identifies the account and is
what invites were sent to.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
authenticate() cannot tell "no such session" from "the session lookup
failed" on its own — both arrive as an error. Answering 401 for both
meant a disk or file-permission problem presented as every session
ending at once: the frontend clears its state and shows the sign-in
screen to everybody, and the logs record a wave of failed logins rather
than the fault.
Now only a missing or unknown credential is 401. Anything else is a 500
and is logged as what it is.
Found while writing the smoke tests, where a deleted database file
produced SQLITE_CANTOPEN on the connections that had not been opened
yet — so most requests succeeded and a few reported the user as signed
out, which reads exactly like a flaky auth bug.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two people saving at the same moment got a 500 after two milliseconds.
The DSN has set busy_timeout since the first commit, and it covered none
of the contention it looks like it covers. With a plain BEGIN a write
transaction starts as a reader and asks for the write lock at its first
UPDATE — and SQLite refuses that upgrade with SQLITE_BUSY immediately,
without applying busy_timeout to it. Backing off cannot help: the
transaction is already holding a read snapshot the other writer is
invalidating.
_txlock=immediate takes the lock at BEGIN instead, where the timeout does
apply and the second writer waits its turn. The driver leaves read-only
transactions deferred, so readers still run concurrently, and every one
of the twenty callers of db.Tx is a write.
Found by opening a task and saving a description, a sub-task and a
comment at once — three writes, one of which came back a 500. The test
fails without the flag.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Gmail now connects. Authorisation-code flow with PKCE, the verifier kept
server-side rather than in a cookie — the callback is a top-level
navigation from Google, and a SameSite=Lax cookie is the kind of thing
that works until a browser tightens its defaults. State is checked, the
attempt expires after ten minutes, and the spent verifier is deleted
rather than left at rest.
Scope is gmail.readonly and the sync is one-way on purpose. Unstarring a
message does not delete the task; completing the task does not touch the
mailbox. Two-way would mean deciding what happens when both sides
changed, and the only honest answer is a sync model far larger than the
feature. The query carries newer_than:30d so connecting an account with
a decade of starred mail does not create a decade of tasks.
The update check is off unless the operator turns it on, asks GitHub at
most every six hours, and sends nothing about the instance — no
identifier, no version, no counts. A development build never reports an
update: somebody running their own tree is ahead of every release, and
telling them otherwise trains them to ignore the notice. Only
administrators see it, because only they can act on it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The MCP server speaks JSON-RPC over Streamable HTTP with the six tools
the brief names, authenticated by a personal API token — so a connector
reaches exactly what its owner reaches and there is no second identity
to reason about. A failed tool call returns a tool error, not a protocol
error: a model can recover from the first and can do nothing with the
second.
CalDAV serves VTODO, two-way, so Apple Reminders and Thunderbird can
read, create and complete tasks. Clients authenticate with Basic using
an API token as the password, never the account password — a client
stores what it is given, and what it is given should not be able to sign
in to the app. Completed tasks stay in the collection: a client that
stops seeing one treats it as deleted and drops its own record of having
done it.
The AI layer is one interface with adapters for Anthropic, OpenAI,
Google and any OpenAI-compatible endpoint. That last one is why the
abstraction exists: an integration that only spoke to a hosted API would
be useless to somebody running their own models. Every feature degrades
to "off" when nothing is configured, and the stored API key is never
returned to the settings page.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Comments and attachments close out phase 2. Files are stored
content-addressed by hash, so the path never contains anything the
uploader chose — a filename is attacker-controlled text, and building a
path from it is how "../" becomes a write outside the data directory.
Everything downloads as application/octet-stream: an uploaded SVG served
inline would run its own script on this origin with the session cookie
attached.
Web Push is written directly against RFC 8030/8291/8292 rather than
pulled in — it is three specs' worth of well-defined crypto and about
two hundred lines, and the alternatives carry a JWT and crypto stack
each larger than the file.
Phase 4 imports Todoist's CSV and exports it back. Todoist numbers
priorities the opposite way to how it displays them — its 4 is P1 — so
the two are converted rather than copied; getting it backwards silently
inverts the urgency of every imported task. Nesting is positional, so
the reader keeps an indent stack and the writer never sorts.
The round trip is tested as the brief asks: a CSV in, out, and back
produces an identical project, and two exports of one project are
byte-identical.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The board is one column per section with native drag-and-drop, which
works with a trackpad, a mouse and a screen reader without any of them
being reimplemented. Dropping is optimistic: a card that snaps back for
200ms reads as a failed drop.
Weeks in the calendar start on Monday. That is ISO 8601 and what every
Danish calendar does; a Sunday-first grid makes a Dane misread the month
at a glance. On a phone the chips give way to coloured bars — at that
width the month's shape is the useful part, not the titles.
Label pages run the filter language rather than a bespoke endpoint, so a
label page and a saved filter can never disagree about what a label
selects.
fix(recurrence): "hver mandag", not "hver uge mandag"
A weekly rule that names its days does not also need the word for the
interval. That is how the phrase is said, and it is what the parser
accepts back — the round trip through Describe now holds for weekday
rules too.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Reminders are absolute or relative ("10 minutes before"). A relative one
is resolved against the task in SQL, so moving a task moves its reminders
too. They are marked sent before delivery, not after: a reminder that
goes out unrecorded repeats every minute forever, which is worse than
one recorded that fails to send.
The calendar feed is a per-user token in the path, because Apple Calendar
and Google fetch with no credentials and no way to prompt for any. Tasks
go out as VTODO — the same component CalDAV will serve — and timed ones
also as VEVENT, since Google ignores VTODO entirely.
Templates store relative day offsets rather than dates, which is what
makes them reusable: an onboarding checklist is "day 1, day 3, day 14",
not three dates in March. Building a project from one is a single
transaction, so a failure cannot leave half a checklist.
Backups use SQLite's VACUUM INTO, not a file copy — a copy of a live
WAL database can miss recent writes or catch a page mid-update, and the
result looks like a database right up until it is needed. Rotation keeps
fourteen by count, not by age: a container that was off for a month must
not come back and delete every backup it has.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
"today & p1", "#Firma & @regnskab", "overdue | (today & assigned to: me)".
The grammar is Todoist's, so the import path from it does not require
relearning anything, and both languages' keywords are accepted.
It compiles to parameterised SQL rather than filtering in Go: one
indexed query instead of ten thousand rows crossing the boundary. No
typed text is ever concatenated into the statement — this is a language
people write by hand, which makes it exactly where string-building would
become an injection.
An expression is compiled before it is saved. A filter that cannot run
is worse than none: it is discovered later, as a list that is empty for
a reason nobody can see.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
"vand planterne hver mandag" becomes FREQ=WEEKLY;BYDAY=MO. The stored
form is RFC 5545, not a private format, so the same field will drive the
CalDAV server and the ICS feed without a translation layer.
Ticking off a repeating task advances it instead of closing it, keeping
the same row — so its sub-tasks and comments stay attached. A weekly
review is one thing that recurs, not fifty-two things. The clock time
survives the move, and the completion is still written to the activity
log, because "what did I get done this week" has to include the chores
that repeat.
Recurrence is parsed before dates, and that ordering is load-bearing:
"hver mandag" contains a weekday, so a date parser reading it first
would claim Monday as a one-off and strand "hver" in the title. It is
also resolved before the time-only case, or "hver mandag kl 9" typed
after nine would land tomorrow morning instead of on Monday.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
CI runs in UTC and the API resolves due dates in the user's timezone,
which defaults to Europe/Copenhagen. For two hours out of every
twenty-four those disagree, so tests written against time.Now() passed
all day locally and failed on a late-evening push.
The application was right; the tests were asking it the wrong question.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The interface for phase 1: sign-in and first-run setup, Today, Upcoming,
project views, sharing, Cmd+K search and the keyboard shortcuts.
Quick add highlights what the parser understood underneath the text as
you type — a mirrored layer behind a transparent input, sliced by byte
offset because the server counts bytes and JavaScript counts UTF-16
units, which disagree on exactly the Danish letters this app is full of.
Every local action is optimistic and reconciled afterwards: ticking a
task off strikes it through and animates it out before the request
leaves, and rolls back with a message if the server disagrees.
The Content-Security-Policy is computed at startup by hashing the inline
scripts in the built index.html. Both of them change on every frontend
build, so a hard-coded hash would go stale silently and show up only as
a blank page in production. The alternative, 'unsafe-inline', would drop
the one protection that makes an injected script in a task title inert.
Today is derived from the store by date rather than from the ids that
came back with the first request — filtering against a captured list
meant a task added through quick add never appeared until a reload.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Completes the MVP surface: project and section CRUD, task CRUD with
sub-trees, drag-and-drop ordering, the quick-add endpoint over the
natural-language parser, Today and Upcoming, search, sharing with roles,
an activity log and a WebSocket for live updates.
Permission checks are route middleware rather than a first line in each
handler, so an endpoint is guarded by where it sits instead of by the
author remembering. Anything the caller may not see answers 404, never
403 — a 403 confirms the thing exists. Visibility is enforced inside the
list query, so another user's project cannot be reached by filtering or
searching at it.
Ordering is a float, so a drop writes one row instead of renumbering the
list. When repeated drops into one gap exhaust the precision, the
section is spread out again and the move retried — in a separate
transaction, since the attempt that discovered the problem rolls back.
The connection pool is not 1: a single connection makes the pool a lock,
and a query run while an earlier result set is still open then waits for
a connection it is itself holding. The store now closes its rows before
querying again, and the pool leaves that mistake slow rather than fatal.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Sessions are stored as hashes, so a database dump contains nothing that
can be pasted into a browser. A login that has passed the password but
not the second factor is modelled as a pending session: it expires, is
revocable, and reaches exactly one endpoint.
Signup is invite-only and takes the email address from the invite rather
than the form, so a link cannot be redirected to somebody else. The
first admin is created by an endpoint that refuses once any account
exists, checked and inserted in one transaction because it is
necessarily unauthenticated.
Login failures and forgotten-password requests answer identically
whether or not the address has an account — otherwise the form is an
account-existence oracle. Password rules are a length floor and nothing
else: composition rules push people towards worse passwords.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The HTTP server serves /healthz, the API namespace and the PWA, with the
frontend embedded behind a build tag so a clean checkout compiles without
a Node toolchain. /healthz reads the database rather than just answering:
a process that cannot reach its own data is not healthy, and a 200 there
would keep a broken container in rotation.
The image is distroless and static — the SQLite driver is pure Go, so
CGO stays off. The Rune manifest uses keep_entrypoint because there is
no shell in the image to run a startup command with, and its log watchers
match the JSON log format the container actually emits.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Password hashes are self-describing PHC strings, so the cost parameters
can be raised later without locking anyone out: NeedsRehash reports which
hashes to upgrade at the next login, when the plaintext is briefly at hand.
Session, invite, reset and API tokens are stored only as SHA-256 hashes;
the plaintext exists once, in the link or the reply. Unsalted is correct
here and only here — the input is 256 bits of randomness that has to be
found by value in a WHERE clause.
feat(store): project roles with viewer/editor/owner and explicit denial
Roles are strictly ordered so every check is "at least X". Access to a
task is access to its project: being the assignee or the author confers
nothing, or removing someone from a project would leave them holding
tasks inside it. Checks take a Querier so they can run in the same
transaction as the write they guard.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Parses date, time, priority, project and labels out of one typed line,
e.g. "betal moms i morgen kl 10 p1 #Firma @regnskab".
Both languages are read at once rather than behind a locale switch,
because people mix them mid-sentence. Locale settles only one genuine
ambiguity: a bare hour after English "at".
Each stage reads a copy of the line with earlier matches blanked out,
so no two readings can claim the same characters — without it "kl 10.12"
is a time and, to the date parser, also the 10th of December. Matches
are returned as byte spans so the input box can highlight what it
understood as you type.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Initial data model for verdande: users, sessions, invites, projects,
sections, tasks, labels, reminders, comments, attachments, activity
and filters, plus personal API tokens.
Search indexes a generated `fold` column alongside the original text.
FTS5's remove_diacritics does not touch ø and æ — Unicode treats them
as letters rather than accented forms — so without it a Danish user
cannot find "grøn" by typing "gron". MatchExpr quotes every term so
free text containing FTS5 operators cannot reach the query planner as
syntax.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
