compose.languagetool.yaml runs LanguageTool's open-source server
(meyay/languagetool 6.8-13, pinned by digest) on `edge` with no site and
no published port, as user 61081 on a read-only filesystem, without a
log. The server alone reaches it (LANGUAGETOOL_URL): grammar.check is
for those signed in, takes 20,000 characters, 60 checks a minute each
(counted in memory), maps a studyset's language to LanguageTool's code
and keeps nothing. instance.info() names the languages it checks; the
privacy page says what goes there. sidecars/README.md has the steps.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
- deploy/assistants.env.example: every variable commented, with a
Claude example and a second assistant on the Responses API. The file
itself is git-ignored and reaches Learn's server alone
(compose.yaml's env_file, required: false, so an instance without it
offers no assistant; that syntax needs Docker Compose v2.24 or later).
- deploy/.env.example points to it.
- The Caddyfile forwards /privacy to the web server, as it does
/simplelogin-callback.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
SIGN_UP takes open (unset), approval or closed; anything else stops the
server as it starts. Closed refuses a SimpleLogin sign-up before its
handle is counted, an address sign-up only once the code has proved the
address, so nothing tells whether an address has an account. Approval
lets a new account wait, at most ten at once: no session is issued for a
waiting or suspended account, only a pass (32 random bytes, kept as its
SHA-256, 30 days, five per account) and four BIP-39 words the operator
reads out to match it. The pass looks, forgets, and deletes an account
that was never approved; nothing else.
Operators hold the scope operator, given first by SQL (D32), then named
in the app: InstanceOperatorEndpoint lists, approves and refuses waiting
accounts, finds accounts, suspends them (ending their sessions), and
names and takes back operators, never the last. account.operating()
counts the waiting for the badge. Members ask the register (five new
operations, a refusal kind signUp), and say closed while it is out of
reach. Never-approved accounts go after 14 days in the daily pruning.
The migration only adds the tables waiting_account and waiting_pass, so
the server before it keeps working on the new schema.
BREAKING CHANGE: a register and its members must run this version
together: a member asks the register operations an older one doesn't
know.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
Every account signs in with SimpleLogin or a password now, so GitHub
goes from the module, the shared app package, Learn and Todo, ahead of
ADR 0017's GITHUB_SIGN_IN_ENDS.
- halkyon_base: no `gitHubSignIn` endpoint, GitHub identity provider or
GitHub refusals; no `githubClientId`/`githubClientSecret`, no
`GITHUB_SIGN_IN_ENDS` and no `forgetGitHubAccounts`; the register
loses `account-with-github`. `SignInWays.gitHub`, the basics'
`gitHubClientId` and `gitHubSignInEnds`, and `PasswordRefusal`'s
`signInAgain` and `noAddress` go. A first password takes a code
mailed to the person's own address; an account without one sets none
(`wrongCredentials`).
- Learn's server: `instance.info()` without `gitHubClientId`, the
pruning without GitHub, and `minimumAppBuild` 24, as builds up to 23
(v0.20.0) require `SignInWays.gitHub`: the server deploys with the
release of build 24.
- halkyon_app: no GitHub button or problems on the start screen, no
password offered after a GitHub sign-in, and `ServerpodAuthGateway`
without the browser half only GitHub used (`(client, base:)`). The
account section offers a first password only to an account with an
address. `LoopbackWebAuth` and the apps' schemes stay, SimpleLogin's
ways back.
- Learn's Profile lists no GitHub; the texts go (English, German).
- Serverpod's `serverpod_auth_idp_github_account` stays, as its
migrations make it; a row goes with its account, and ADR 0020's D36
gives operators the SQL to delete those left.
- deploy/: compose.yaml and .env.example without GitHub. The READMEs,
AGENTS.md files and ADR 0017 amended.
BREAKING CHANGE: builds up to 23 can't read the instance's answers
about an account's ways in; minimumAppBuild is 24.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
The server's half of D25: every account has a way in, and the anonymous
accounts left from before go with the daily pruning.
- The module signs nobody in anonymously: `anonymousIdp` and
`gitHubUpgrade` are gone, and so is linking GitHub at the register.
- A password is set only by someone whose address is theirs (a
`firstPassword` code) or who signs in with GitHub alone (a fresh
sign-in and a `signUp` code); an anonymous account is refused
`wrongCredentials` before any code is used.
- SimpleLogin refuses to link onto a caller without `identified`
(`wrongCaller`), the member's endpoint too, so nothing turns an
anonymous account into one with a way in.
- `pruneAnonymousPeople` deletes every anonymous account left, however
new, never one identified. Learn and Todo schedule it as
`prune-accounts` and cancel the old `prune-abandoned-anonymous-people`
entry once.
- Learn: someone without `identified` keeps no files;
`FILE_LIMIT_ANONYMOUS_MB` is gone. `minimumAppBuild` is 23, as builds
up to 22 start with an anonymous account.
- Learn's sync tests make their person and session by SQL in the
development database, as Serverpod stores sessions.
- Docs, store listings and ADRs 0010, 0014, 0016, 0017 and 0019 amended.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
simpleLogin.begin takes the way back (SimpleLoginReturn: loopback,
scheme) rather than the platform, so macOS can come back through its
scheme as iOS and Android do; the markers become learn-loopback,
learn-scheme, todo-loopback and todo-scheme. Not breaking: the
endpoint has never been released.
D9 corrected by the developer after research: Android, iOS and macOS
come back through the app's scheme, Linux and Windows through the
loopback. Verified links protected nothing here: a code caught on its
way is worthless without the attempt's ID, and a hostile app beginning
an attempt of its own can ask for a way back that other platforms need
open. The ADR notes the residual risk and the check of prompt=consent
still to do. assetlinks.json stays served.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
Someone whose SimpleLogin identity is new to the register but who has an
account ("I already have one") names their handle and proves the
account's recovery phrase; the new identity takes the old one's place.
- `simpleLogin.beginMend(attempt, handle)`: the keys wrapped under the
phrase's key, the public keys and a nonce, for the account with that
handle (whatever its case). Only an account with keys and a SimpleLogin
identity is mended, a link it had, as nobody adds SimpleLogin signed
out (D10); a handle nobody has, or any other account's, gets a decoy
from `signInDecoyKey`, the same each time. At most 10 an hour per
handle and 30 per network.
- `simpleLogin.mend(attempt, proof)`: the phrase's proof for `mendLink`,
binding the attempt (`mendLinkDigest`, on both sides), so a proof
relayed from another sign-in links nobody's identity but the device's
own; checked for the waiting identity first, which leaves the proof
unspent. Every session of the account ends, in every service, and the
sign-in stands ten minutes for a change, such as a lost password.
- SimpleLogin needs `SIGN_IN_DECOY_KEY` alone of ADR 0017's three keys
(`signInDecoyKeyOf`); `addInstanceRoutes` stops a server with a
SimpleLogin app and without it (`checkSimpleLoginMending`).
- Members ask the register for both; `RemoteRegister` and the instance
API carry them.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
The `simpleLogin` endpoint on halkyon_oidc's SimpleLogin preset: `begin`
for the app and platform (the state's marker for the forwarding page), at
most 30 an hour per network; `complete` with SimpleLogin's code, signed out
a session for the account with that SimpleLogin ID, signed in SimpleLogin
added to the caller's account, who becomes identified and anonymous no
longer. Someone new gets no account yet: their identity waits 30 minutes
under the attempt's ID (D16), with a handle suggested from their name,
`anna_mueller` (D17), for the sign-up of step 3.
The register alone holds the instance's SimpleLogin app
(`simpleLoginClientId`, `simpleLoginClientSecret`) and exchanges codes; a
member asks it for both steps and issues its own session. `SIGN_IN_WAYS`
says which ways the register offers, the instance's basics and the ways
in carry `simpleLogin`, GitHub links nobody more where SimpleLogin is
offered, and Learn's and Todo's pruning forgets the unfinished sign-ins
hourly. No migration: the tables are halkyon_oidc's, already in place.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
SimpleLogin takes only https redirect addresses, each registered exactly,
so the instance's SimpleLogin app registers one page on the register's
web server, /simplelogin-callback, which sends the browser on to the app
that began the sign-in: a 303 to the desktop's loopback (127.0.0.1, a port
per app) or the phone's scheme, chosen by the marker the state begins with
(`<app>-<platform>.<rest>`, the rest 1 to 512 characters of base64url),
from a fixed table, so it sends nobody anywhere else. It forwards the code
or SimpleLogin's error (`deny`) with the state and nothing else, refuses
anything malformed, repeated or not a GET with a page that repeats nothing
it was given, and keeps and logs nothing (no-store, no referrer). It is
built ahead of the sign-in itself so development can sign in through
learn.halkyon.eu's page, which the development app registers too.
Android's assetlinks.json for Halkyon Learn signed with the release key
is answered by the Caddyfile on learn.halkyon.eu alone: the one domain the
project's builds name, and the server knows no app's key. A root test
holds its fingerprint to tool/release/android-certificate.sha256. The
app's intent filter comes with the sign-in; iOS's association waits for
an Apple team ID.
The profile page's language choice and look move to common/web_page.dart,
shared by both pages.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
The developer decided on 23 Oct to drop the hand-off to an AI assistant
outside the app, and Claude's connector with it, in v0.19.0, the app and
the server together: build 22 raises minimumAppBuild anyway, so no app
calling the connector's endpoints uses the instance after it. An instance
is to have one way to an assistant, its operator's (ADR 0021, proposed).
The app: the field at the foot of Home, and the button in the Library, on
a studyset and while studying, search the web alone, through the engine
chosen in Settings or Android's own search, with the parts of a card the
person picks; a studyset's name and description are never offered. Gone:
the assistants a question opened in (AiAssistant, AssistantHandOff), the
switch between them and the web, Settings' choice of assistant, Connect
Claude with its route, cubit, repository and gateway, and their texts. A
settings file from before keeps the old keys until the next change of a
setting writes it anew; nothing reads them.
The server: lib/src/assistant/ (MCP, its OAuth, the five tools, connect
codes and tokens), AssistantEndpoint, the daily forgetting of the
connector's secrets (its scheduled call cancelled at start), its lines in
the deletion hook, instance.info().assistantConnectorUrl (nullable, so an
older app reads it as none), and Caddy's forwarding of /mcp, /oauth/* and
/.well-known/oauth-*. The migration drops the connector's five tables,
children first:
DROP TABLE "assistant_token" CASCADE; -- tokens of a connection
DROP TABLE "assistant_grant" CASCADE; -- OAuth codes of a connection
DROP TABLE "assistant_connection" CASCADE; -- a person's connected client
DROP TABLE "assistant_connect_code" CASCADE;
DROP TABLE "assistant_client" CASCADE; -- OAuth clients registered
No other table references them, so CASCADE drops only their own foreign
keys. For the minute between the maintenance run and the switch, the old
server answers the connector with errors and fails to delete an account
(its hook still deletes from these tables); ADR 0022 says so.
BREAKING CHANGE: a connected Claude gets no answer from the deploy on, and
the app no longer opens a question in an AI assistant.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
Caddy's network had IPv4 alone, so Docker's userland proxy forwarded
every IPv6 visitor to Caddy from the network's gateway, and the server
counted all of them as one network (the second reading's M5, reasoned).
The edge network gets a unique local /64 (fd4a:6c6b:7e80::/64), Caddy a
fixed address in it beside its IPv4 one, both trusted as the proxy; the
CI runner's unit denies fd00::/8 and its sandbox check tries Caddy's IPv6
address too. A commit of its own: it needs Docker 27 or later with
ip6tables on the server and the network made anew (compose down, never
-v, then up), checked by hand first (NEXT.md).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
With Todo and the shared packages in the workspace, a cold `dart analyze`
peaks at about 5.4 GB (3.7 GB warm). Above the 4 GB soft limit the
kernel slowed it to a crawl, and at 5 GB it killed the job. Analysing
parts from the root drops the lint plugin's rules, so the limit rises
instead, and the kernel kills CI first should the server run out.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
Every per-network limit counted by the connection's address, which behind
Caddy is Caddy's for everyone: new anonymous accounts, codes by mail and
password sign-ins were each limited for the whole instance.
`networkOf` (now `common/network.dart`) takes the last address of
X-Forwarded-For, the one the proxy saw itself, for a connection from a
proxy the operator trusts (`TRUSTED_PROXIES`, which deploy/compose.yaml
sets to Caddy's fixed address); otherwise, or where that is no address,
the connection's. Never the Forwarded header, which Caddy passes on as
sent, so never Serverpod's `remoteInfo`. A value that isn't addresses
stops the server as it starts.
Serverpod's own limit on anonymous accounts counts by the connection's
address too and can't be told otherwise, so it is off; the anonymous
endpoint counts them itself, 30 an hour per network as before.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
TOTP_KEY, RECOVERY_CODE_KEY and SIGN_IN_DECOY_KEY switch password
sign-in on (each at least 32 characters, or the server stops), and
GITHUB_SIGN_IN_ENDS names the day GitHub stops signing anyone in. CI
gives the test runs throwaway keys as it does the other passwords.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
halkyon_base runs as the instance's register or as one of its members.
The register keeps the ways in, handles and profiles; members ask it
over HTTP with the instance's shared secret (instanceSecret), and open
sessions of their own for its accounts, under the same account ID.
- AccountsMode from the environment: INSTANCE_MEMBERS on the register,
ACCOUNT_REGISTER_URL on a member. Learn always holds the register
(Explore joins the profiles); Todo is a member where it is set.
- AccountRegister: LocalRegister, RemoteRegister, an in-memory one in
tests. Calls between services on the web server, refused without the
secret (constant-time compare); a server needing it doesn't start
without it.
- Deleting an account deletes it everywhere: notices per member kept in
the deletion's transaction (account_deletion_notice), told at once and
hourly until confirmed; members prune daily whom the register no longer
knows, never while it knows none of them or can't be reached.
- Deploy: INSTANCE_SECRET and INSTANCE_MEMBERS in compose and .env.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
- Melos 8.9 as the root's dev dependency, with scripts across the
workspace (analyze, test:dart, test:flutter, generate, check) and its
filters (--diff, --include-dependents); cli_util overridden to 0.5.2
for it, since Learn's flutter_launcher_icons asks for 0.4 and uses only
its logger, unchanged in 0.5.
- tool/check.dart in place of tool/check.sh: packages, formatting of our
own files, analysis, the tools' tests, then every package's tests
through Melos (--changed-since=<commit> for only what changed and what
depends on it). tool/pre-push is the hook: the signature check, then
the checks. CI generates every server's code and runs the same check.
- The images build from the root with a workspace narrowed to what they
need (tool/workspace_only.sh); compose, the deploy's migration test,
the snap, the release script, the editor's launch files and Learn's
asset tools (run from apps/learn) follow the new layout.
- The project's source is named as project-halkyon/halkyon (source link,
snap, F-Droid listing, release uploads); the AGPL packages carry the
licence's own text.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
halkyon_shared's 96f2d14: an operator's DATA_LOCATION and
BACKUP_LOCATION may name a subdivision (ISO 3166-2, US-CA, DE-BY)
besides the EU or a country, and the app adds "(EU)" to a member
country and the country to a subdivision ("Germany (EU)", "California,
United States") under the most specific flag there is. deploy/.env and
the backups' guide say so; a Storage Box is named by its country now.
The package registers its notices (the flags, CLDR's subdivision
names); Learn's own CLDR notice for the world map stays, its
registration renamed to say so.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
Its operator sets DATA_LOCATION and BACKUP_LOCATION in deploy/.env (a
country's code or EU, the backups also none), and instance.info() hands
them on. The first start's tile and Switch account show the flag of
where the data is and the words, as labels ("Data: European Union",
"Backups: Switzerland"), the countries by the world map's CLDR names:
where the disks are, never which law governs. The flags come from the
flag package, which draws flag-icons; its MIT notice joins the licence
list.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
Studying:
- A round is one list in the order its cards come: those done, the one
shown, those to come. A card due again within the day goes back in
after 4 more cards when rated Bad, at the end otherwise, so the pills
above the card follow it; the buttons say "in 4 cards", "at the end",
"now" (round_order.dart).
- The studied studyset's daily limits hold for everything inside it
together, as Anki's do for subdecks: twenty topics start 20 new cards
a day in all, not 20 each.
- The study header's studyset name shows whole on hover or long press.
- The assistant's button sits at the lower left: in the Library across
from Create studyset, in a studyset above its bar, and higher while
studying, above the buttons that answer the card.
A studyset:
- Progress also counts the learning paths' steps done and lists the
PDFs not read to their end, begun first, each opening the reader,
across its topics (MaterialProgressCubit).
- Next up starts with studying for the exam while one is ahead: its
reviews due and the new cards its pace leaves for the day.
- Learning paths tagged with a level count towards its automatic exam
plan and offer it, as cards do.
- A step's first heading equal to its title is dropped on import and
hidden in paths imported before.
Signing out, or a session the server ended, leads to a start screen
(GitHub, email, without an account, another instance) rather than a new
anonymous account; only the very first start makes one by itself.
Settings names the kind of account and its instance.
Also: the Watch screen reads the cards once for held words and their
progress; its ±10 s buttons tick as the navigation bar does; a map card
keeps its view as its caption changes and brings a picked region
towards the middle; an empty or failed library can be pulled to sync.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
Once the person agrees, the device holds MeCab's IPADIC and the
instance's JMdict and cuts lines as Jisho.org does, so their text never
leaves it (ADR 0012). MeCab comes through mecab_for_dart, whose build
hook compiles it; its tokens are grouped by Ve's rules, written anew in
Dart, with に kept a word of its own. Each word is found as it stands or
as a form of a word, else by its dictionary form, else as the longest
words it holds from token to token; longer words over token boundaries
are offered as before.
The curator packs Ubuntu's compiled IPADIC beside the dictionary files
(curate ipadic, and every update run). The server exports its JMdict
from its tables, so the device finds the same keys and keeps the same
versions, and serves both at /device-dictionaries/, named in
dictionary.deviceFiles('ja') with their hashes; Caddy forwards the path.
The app checks each file against its hash, unpacks IPADIC, builds a
SQLite database from the export on an isolate, and asks the instance
for newer files every 30 days, downloading only what changed. The Watch
screen offers the download once per device with its size (34 MB, some
150 MB once set up); Settings downloads and removes them. Without them,
or should cutting on the device fail, lines go to the instance as
before. MeCab's, IPADIC's and Ve's licences join the licence list.
Every build now needs a C++ compiler; mecab_for_dart holds code_assets
at 1.x.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
A profile now has a page on the web, at the root of the account domain
(https://halkyon.eu/@raisondetredev where the operator set ACCOUNT_DOMAIN,
else on the instance's web server), so the address belongs to the
identity rather than to Learn and stays when other services share it
(ADR 0011). The page shows what the app's profile sheet shows, in English
or German, without JavaScript or cookies, asks search engines not to list
it, marks every link rel="me" so other sites can verify links to it, and
sends a visitor who asked on another host to its own address. Caddy
forwards /@… to the web server, and serves nothing else on the account
domain; compose now passes ACCOUNT_DOMAIN on, which it never did.
A link is verified as Mastodon verifies one: its page links back to the
profile's with rel="me". profile.checkLinks fetches each link's page, at
most ten times an hour, over HTTPS on port 443 and to public addresses
only, connecting to the address it checked and verifying the certificate
for the host, with three redirects, 1 MiB and ten seconds at most, and
keeps what links back on the profile (verifiedLinks, a column of a table
no device syncs). A changed link loses its check, a new handle all of
them. The app shows the page's address to copy, checks a link saved by
itself and on Check links, and marks verified links in the form and in
others' profiles; a check answered after a newer save only adds its
checks.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
Email sign-in's codes are mailed over SMTP (pub.dev's mailer), through
whichever service the passwords name: emailSmtpHost, emailSmtpPort,
emailSmtpUser, emailSmtpPassword and emailFrom (EMAIL_SMTP_* and
EMAIL_FROM in deploy/.env, with EMAIL_SIGN_IN_PEPPER). TLS only:
STARTTLS, or TLS directly on 465. learn.halkyon.eu sends through Proton's
SMTP submission. Without a service, codes still go to the log in
development, and elsewhere there is no email sign-in.
The mail is plain text, in English or German: the code first and in the
subject, so it shows in the notification and the system can offer it,
the instance it signs in to, its 15 minutes, and that it can be ignored;
no links, no tracking, no HTML. A mail that couldn't be sent is a refusal
of its own (mailFailed): nothing of the request is kept, and the log
names the error's type only, as the service's answer may name the
address. A fake SMTP server on loopback drives the real mailer in the
tests: it delivers, and nothing goes to a server without TLS.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
What applies of Japanese's features to Chinese, Korean, Greek, German and
English, with the developer's choices and each decision's reason: Unihan's
characters, KRDict's words and levels, Make Me a Hanzi's and the curator's
own strokes, typed pinyin, radicals and readings by language; what doesn't
apply or has no open data. AGENTS.md, the server's README and the
instance's example environment follow.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
For parity with Japanese's KANJIDIC2 and kana:
- Unihan (Unicode License v3) makes characters' dictionaries: the 8,105
standard simplified characters (unihan-hans) and the traditional ones in
use (unihan-hant), told apart and ranked by CC-CEDICT's words (Unihan
calls 这 its own traditional variant; 後 ranks as the 后 of 以后), with
pinyin, meanings, strokes, Kangxi radical, HSK 2025 character levels,
the 3,500 common ones and Hong Kong's grades; and Korean Hanja
(unihan-ko) with their sounds, the 1,800 of education, and their native
readings from Wiktionary (물 수), with the 40 Hangul letters as kana are.
- The Greek alphabet, the curator's own (CC0): letters with names,
transliteration and sound.
- KRDict, the Basic Korean Dictionary for learners (CC BY-SA 2.0 KR,
downloaded without an account), replaces Wiktionary for Korean's words:
learners' levels, pronunciation where it differs from the spelling
(같이 [가치]), Hanja, English senses; Wiktionary's update keeps the
words' forms (가다's 99) and the Hanja's readings beside it for both.
- The update converts KRDict when its page offers another export and the
characters weekly, and removes the files of languages no longer kept
or of sources replaced.
Measured on the real data: Unihan 3 s; KRDict 54,870 entries in 24 s
and under 600 MB.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
A studyset in German, Greek, English, Chinese or Korean finds its
dictionary as one in Japanese does (ADR 0008), and its subtitles are cut
into words by its language (`WordCutting`): Chinese by its longest
words as they stand, the longer words over each cut offered as for
Japanese; Korean at its spaces, a word the dictionary lacks as it stands
taken without the particle it ends with (집에서 is 집, 아이 stays a
child); every other language at its spaces, a word written exactly so
first (the noun Essen before the verb essen). A studyset finds the
dictionary of its language tag, else of its language alone (zh for
zh-Hans).
A Wiktionary word shows its genders, principal forms, IPA, romanisation
and Hanja; a German or Greek noun of one gender goes onto its card with
its article (das Haus, το σπίτι); its tag is its word class, named in
full (Noun, Adjective, Preposition), as Wiktionary's adj and adv would
read as JMdict's codes. Articles, pronouns, prepositions and
conjunctions are left out of a video's words. Chinese words show their
pinyin over each character on the Watch screen, carry their HSK level
as a tag (HSK-1 … HSK-7-9), and the dictionary adds a level whole; the
Watch screen filters by HSK levels for Chinese, JLPT's for Japanese, and
by none for a language without. A reading is typed in a session only
when it is kana, so pinyin isn't. Words are looked up on the web on
Jisho in Japanese, on Wiktionary at the language's section otherwise,
and a search's hint speaks of kana and romaji only in Japanese.
`DICTIONARY_LANGUAGES` in deploy/.env names the languages an instance
keeps (unset: de,el,en,ko,zh; empty: none).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
With Jellyfin's default of three failed tries, anyone could switch the
demo off with three wrong passwords; an account without a password has
nothing for the lockout to guard. The runbook also says how people sign
in to it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
rsync has to be installed on both ends, and learn.halkyon.eu's server
has none; tar through SSH needs nothing the server lacks.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
The sidecar's steps made Jellyfin's user with its own group, to which
useradd gave a system group ID (987 on learn.halkyon.eu's server), while
compose.jellyfin.yaml runs it as 61096:61096. The group is now made
first with the same ID, and the steps say why useradd's warning about an
ID above the system range is meant.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
A failed deploy put the checkout back at the deployed commit but left
the containers as the new one had made them: when Caddy couldn't start
on the new network (2026-10-14), learn.halkyon.eu stayed down until they
were brought back by hand. Failing after the switch, the deploy now also
starts the containers as the deployed commit has them, on the new
images, which fit the database the migrations have changed. Before the
switch, the running containers are never touched, as before.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
The deploy that brought the edge network failed and left learn.halkyon.eu
down: the server, started before Caddy, got the first free address on
the network, 172.30.80.2, which is Caddy's fixed one, so Caddy couldn't
start ("Address already in use"). Containers without an address of their
own now get one from 172.30.80.128/25, below which the fixed ones (Caddy
.2, sidecars from .10) are never taken.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
The servers' names are no secret, so the settings that carry them stay
as they were, and nothing on a server or a laptop needs changing: CI
runs on `sahaquiel` again, the release uploads to `ramiel-fdroid`, and
deploys by hand go through `sahaquiel-deploy`. Prose still names a
server by the domain it serves, which says what a reader needs.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
After v0.13.0's deploy every request to learn.halkyon.eu took a minute or
two and syncs failed, until the server was restarted by hand. The
migration changed a synced table, so the new server, as it first used
the database, found the sync engine's registry of synced tables changed,
and then rebuilt every space's projections in every session until it
stopped (serverpod_offline_sync issue #150). The maintenance run that
applies the migrations doesn't spare it, as AGENTS.md wrongly said.
redeploy.sh now restarts the server once it answers whenever the deploy
brought a migration, and waits for it to answer again: started a second
time, it finds the registry as the first start left it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
An instance may now run services beside itself, each a compose file of
its own switched on in deploy/.env (COMPOSE_FILE=compose.yaml:compose.
jellyfin.yaml), so the deploy runs them with the instance. The first is
Jellyfin, which people watch and listen to learn from: the instance then
suggests it (JELLYFIN_URL from JELLYFIN_DOMAIN), and Caddy serves its
site, refusing every change to a user from outside, so a passwordless
demo account stays everyone's; its admin signs in through an SSH tunnel.
The instance's compose gets a second network, `edge`, for Caddy (at a
fixed address, which sidecars trust as their proxy) and the server;
Postgres stays on the internal one. Caddy loses its needless way to
Postgres, and a sidecar never gains one. A sidecar publishes no port,
as the CI runner may reach 127.0.0.1, and runs as a user ID of its own.
Caddy imports the sidecars' sites from sidecars/, mounted by their
compose files (none without), and redeploy.sh recreates it when one
changes. sidecars/README.md has the steps.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
An instance now names a Jellyfin server for its people to learn from
(instance.info().jellyfinUrl): its operator's JELLYFIN_URL (`none`: none),
else https://jellyfin. on the domain registered for its public host, as
learn.halkyon.eu suggests jellyfin.halkyon.eu, once that answers as a
Jellyfin server. The registered domain comes from Mozilla's Public Suffix
List, kept in the server's data folder, so an instance at example.co.uk
never names someone else's jellyfin.co.uk. The instance looks as it
starts and again hourly when asked; the app never reaches for a guessed
address.
In the app, Add material offers From Jellyfin as soon as the instance
suggests a server, before anyone signed in to one. The sign-in fills the
server in and finds it by itself, and lists the people the server shows
on its login screen: one without a password, as a demo's, signs in with
a tap; one with a password is named in the form, which asks for it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
Files called the servers by their own host names, which are the
operator's and say nothing a reader needs. They now say which domain a
server serves: learn.halkyon.eu's server (the instance and its CI
runner), the halkyon.app server (Forgejo and the F-Droid repository).
Two names are settings rather than words. CI's runner label becomes
`instance` (`runs-on: instance`, `labels: ["instance:host"]`): the
runner reads its installed copy of deploy/runner/config.yml, so jobs
wait in the queue until that copy is updated and the runner restarted.
The SSH aliases the scripts and docs use become `learn-deploy` and
`fdroid-upload`. AGENTS.md says to keep it so.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
The deploy reloaded Caddy when deploy/Caddyfile changed, but a reload
reads the file the container sees, and that is the old one: Docker
mounts a single file by its inode, and git writes a changed file anew.
The first deploy of the assistant connector showed it (the container
still had the Caddyfile without the connector's routes). A changed
Caddyfile now recreates the Caddy container, which mounts the new file;
its certificates are kept in their volume, and the instance is away for
a few seconds.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MMm7usAo6q9ZZRoPkWjoxi
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
A connect code, an authorization code and a refresh token are each
consumed by deleting their row and going on only if that deleted it, so
of two requests racing with the same one only one succeeds (as OAuth
requires); tests race them. Registration, open to anyone, takes at most
200 clients an hour (429 after), so a flood can't fill the table before
pruning. The deploy reloads Caddy in place when deploy/Caddyfile changed,
as `docker compose up` doesn't restart it for a file it mounts: without
it the connector's paths wouldn't reach the web server.
Also formats the app's fake assistant gateway, new and so unchecked by
the last commit's check (tool/check.sh formats tracked files only).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MMm7usAo6q9ZZRoPkWjoxi
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
A Claude plan can't drive a chat inside another app, but Claude's apps
(every plan, the phone apps too) can add a remote MCP server as a custom
connector. The instance now is one (ADR 0006): Serverpod's web server
serves the MCP endpoint (/mcp, Streamable HTTP, JSON answers, no
sessions) with five read-only tools over the person's own synced rows
(their studysets, a studyset's cards and notes a page at a time, a
search across cards, a note, their progress), and the OAuth the
specification asks for: protected resource and authorization server
metadata, dynamic client registration, authorization code with PKCE
(S256 only), tokens bound to the resource, iss in the response, refresh
tokens that rotate. Caddy forwards these paths alone to the web server,
which production now runs on 8082.
People connect with a code from the app (Settings › Assistant › Connect
Claude): eight Crockford characters, valid for ten minutes and once,
typed on the authorization page, which names where it returns to, loads
nothing from elsewhere and may not be framed. Anonymous people can
connect too. Codes and tokens are kept only as SHA-256 hashes; an access
token lives an hour, a refresh token 30 days. The app lists what is
connected and disconnects it, which ends its tokens at once; deleting
the account deletes it; a daily job forgets what expired and clients
nobody connected within a week. InstanceInfo says where the connector is
(additive); the migration only adds server-only tables, so devices and
older apps are untouched.
The handlers take plain requests, so the integration tests drive the
whole flow (registration, the code, tokens, refresh, the tools, another
person's content, disconnecting, expiry, pruning, account deletion)
without a web server.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MMm7usAo6q9ZZRoPkWjoxi
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
A person with an identity and a public profile publishes a studyset with
everything inside it as its next release (ADR 0005, rule 3): the server
reads their synced rows and writes them through the git command as a
commit on main and the annotated tag vN, in a bare repository of the
studyset's own, named by a random public id made at the first publish.
A release is a directory per studyset with a file per card, note, path,
step and PDF (its id, position, tag names and payload as synced, keys
sorted and indented) and the files the content names, without what is
only the author's (a card's report, a PDF's deadline). Publishing is
refused without a licence allowing passing it on in every studyset ("no
derivatives" allowed), with revisions to settle, with files not uploaded
yet, when nothing changed, or for a name that won't do; the name is
fixed from the first release, as is the person's handle.
In the app, Share in a studyset's action row opens the Publish sheet;
publishing syncs first, and settles by itself the revisions that merge
cleanly, listing only real conflicts. The server's image gets git, the
repositories a volume of their own, and the hourly backup a git bundle of
each.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MMm7usAo6q9ZZRoPkWjoxi
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
Every dictionary entry now has a version: 16 hex digits of the SHA-256 of
what cards take of it (headword, reading, entry), made by the loader and
carried on every hit. A card keeps the version it took (card schema 8).
DictionaryEndpoint.changed returns, of the entries named with the
versions cards keep (1,000 at most a call; the app asks 500 at a time),
those whose version is another now.
The card browser says how many cards are behind their dictionary and takes
the corrections at a tap (decided 2026-10-09: by offer, never by
themselves). Taking one writes the entry as it is now over the revision
in which the card last took it, beside whatever came after, so the
revisions merge per side as two devices' edits do (ADR 0003): a side only
the dictionary changed takes the correction, a side the person changed
too is a conflict settled in the composer, whose words now name that
case. Cards made before versions are never offered one.
The server rereads its dictionary folder hourly now, not daily: the
dictionaries service fills its volume after an instance's first start,
and an hour with nothing changed costs a hash of each file.
Migration 20261009145434785-dictionary-entry-version: adds
dictionary_entry.version, text NOT NULL DEFAULT '' (rows loaded before
have none and are never reported as corrected), and by hand empties
dictionary_source."fileHash", so every source loads once more and its
entries get their versions; the schema stays as definition.sql says. The
old server goes on working until the switch.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MMm7usAo6q9ZZRoPkWjoxi
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
EDRDG's licence asks of a server showing its dictionaries that it update
them at least monthly. The instance now does it alone: a `dictionaries`
service in compose, built from the curator (halkyon_learn_curator/
Dockerfile, its own unprivileged user 10001, 1 GB at most), runs `curate
update --every 7`, downloading EDRDG's files of the day once its files are
a week old and replacing them in a volume the server mounts read-only.
The curator never enters the server's image or process (decided
2026-10-09: a container of its own, weekly). The server rereads the
folder daily at 04:07 (DictionaryReloadFutureCall), loading what changed,
and loads now run one at a time. redeploy.sh builds both images.
An operator who'd rather supply the files removes the service and mounts
deploy/dictionaries/ into the server, as compose.yaml says.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MMm7usAo6q9ZZRoPkWjoxi
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
The server now holds dictionaries for everyone: `dictionary_entry` and
`dictionary_source`, reference data owned by no one, never synced. It
loads the dictionary files in its `dictionaries` folder at every start,
those whose SHA-256 changed: in the background, each source replaced in
one transaction, so searches meanwhile find the old entries; a file it
can't read is logged and left out. All of JMdict loads in under half a
minute. On an instance the folder is deploy/dictionaries, mounted
read-only (compose.yaml); the server's README says how.
DictionaryEndpoint (signed in): the dictionaries with their credits; a
search (exact key, exact loose key, key starting with the query, loose
key starting with it, each in the source's order; katakana as hiragana,
macrons written out), some 150-450 ms over all of JMdict; and the
entries with a label, counted and a page at a time (all JLPT N5 words).
The curator now keeps spellings that aren't quite right apart, as
`looseKeys` (romaji with long vowels typed once): "sushi" finds 寿司
before 数詞 (すうし).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MMm7usAo6q9ZZRoPkWjoxi
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
A Storage Box answers only inside Hetzner's network unless External
Reachability is on, for the box and for the sub-account. FTP and SFTP on
port 22 can't be turned off, so the sub-account turns off Samba and
WebDAV only. A host key of another type is checked against Hetzner's
list too.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
learn.halkyon.eu dumps its database every hour (people's files are in it
too) and restic stores the dump encrypted in a Hetzner Storage Box,
keeping 24 hourly, 7 daily, 4 weekly and 3 monthly snapshots. The box's
own snapshots, removable only in Hetzner's Console, protect the backups
from a taken-over server. A systemd timer runs it as deploy; once a day
it also prunes and checks the repository. Failures go to the journal.
Hourly, because a restore costs people who started after the backup
what their device held: the app no longer finds its session and starts
anew. deploy/backups.md describes the setup, the restore and what a
restore loses; both were rehearsed on the local Compose stack.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
An anonymous person may keep 100 MB of images and sounds, an identified one
500 MB, and one person more by a file_allowance row the operator adds; all
of them together 15 GB, since every file is in the database and in each dump
the server keeps (decided 2026-10-02). Each comes from the environment
(FILE_LIMIT_*_MB), with these defaults, documented in deploy/.env.example.
The server checks when an upload starts, counting unfinished uploads so they
can't fill the disk unnoticed, and again when it finishes, counting what is
kept, so uploads started side by side can't overshoot; a refusal is a
FileLimitExceededException, and an upload refused as it finishes is deleted.
deletePerson deletes the allowance with the person. Migration: the new
table file_allowance, only added.
The app keeps a refused file waiting on the device and stops that round of
uploads; the Anki import warns when a deck's files may not fit in the room
left (asset.usage()).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
They are the way back from a deploy, not backups: those go off-site (3c).
Every file people keep is in each dump, so fewer dumps leave room for more
files (decided 2026-10-02).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
Decided 2026-10-01. CI run #32 died in its analysis: a cold `dart analyze`
of the workspace with the lint plugin now peaks at about 3.7 GB (2.1 GB
warm), measured with an empty analysis cache, while the runner's unit
allowed 3 GB. The kernel killed it, systemd's default OOMPolicy (stop) then
stopped the whole unit, and the job's remaining processes got SIGTERM: the
"signal: terminated" in the log. The packages this branch added tipped it
over.
The runner may now use 5 GB of sahaquiel's 8, and above 4 GB the kernel
reclaims from CI first (MemoryHigh), so the instance keeps at least 3 GB;
it uses under 1 GB today. Installed by hand on sahaquiel, like the rest of
the unit.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
Two of Serverpod's own conventions, as its v4 project template and roles have them.
CI now runs `serverpod generate` (CLI pinned to the server's Serverpod version, like
the template's workflow) and fails if the committed generated code differs from its
output, so a model or endpoint changed without regenerating can't reach main. Checked in
a clean clone: today's code passes, a committed stale generated file fails.
redeploy.sh now applies the migrations in a run of their own, with Serverpod's
maintenance role (--role=maintenance --apply-migrations), after the dump and before
switching to the new server. That run exits with 0 only if the migrations worked, and
each migration is one transaction, so a failing one stops the deploy while the old
server keeps serving; before, the new server applied them at start and would have
crash-looped. Tested on the local stack: a deploy through the new step, and a failing
run (wrong database password) that exits with 1 after six attempts. AGENTS.md says
migrations must keep the old server working until the switch.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
Installing the runner on sahaquiel, `useradd --system` gave `runner` user ID 999: the ID
under which PostgreSQL and the server run inside their containers, which Docker doesn't
remap. To the kernel, those processes were the runner's own, so a job could have read the
server's environment (database password, session pepper, GitHub secret) and, through
/proc, PostgreSQL's files and its local socket, which the official image trusts. The
sandbox's other walls don't help there: ProtectProc hides only other users' processes,
IPAddressDeny doesn't cover Unix sockets. Only our own check ran as 999; `runner` now has
user and group ID 61000.
The check job now starts with a Sandbox step that fails unless only the runner's own
processes are visible, all in one process namespace, /home/deploy is invisible, the
Docker socket is unusable and Docker's bridge network unreachable. It would have caught
this; outside a sandbox it fails as it should. The unit and AGENTS.md say how to create
the user.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
Pushes to main were deployed by hand. Now Forgejo Actions runs tool/check.sh on every
push, and a green main deploys learn.halkyon.eu by itself.
The runner (forgejo-runner v13.1.0) runs on sahaquiel as the unprivileged user `runner`,
with jobs directly on the host: containers would take Docker access, which is root on the
production server, and Forgejo warns that host jobs have no isolation from the runner's
user. So its systemd unit (deploy/runner/forgejo-runner.service) provides it: /home and
the Docker socket invisible, Docker's container networks unreachable, no capabilities,
two cores and 3 GB at most, lowest priority. It is registered to this repository only,
with the UUID and token of Forgejo's "Create new runner" (the old `register` is
deprecated); the token reaches it as a systemd credential. The configuration
(deploy/runner/config.yml) was dry-run against v13.1.0, the unit rates 1.8 in
`systemd-analyze security`. CI runs on pushes only: a pull request from a fork would run
its code on the production server.
The deploy job calls deploy/redeploy.sh as `deploy`, through a key that authorized_keys
pins to that script and to connections from sahaquiel itself. The script deploys the
commit CI checked while it is still main's tip, as a fast-forward, only if every new
commit is signed by a key in deploy/allowed-signers (as the deployed checkout lists
them), after dumping the database (the last ten are kept), and fails unless the server
answers afterwards; a failure puts the checkout back. Tested against the local stack:
refusals of a malformed argument, a stale commit, a forged unsigned commit and a
non-fast-forward, a real deploy with its dump, a repeat, and a failed build.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
