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>
ADR 0019 gets "The final reading (22 Oct 2026)": what decides a key is
bound into what is sealed or signed, the walk up the tree, a device
without the keys writing nothing, settings' newest versions and Count me
folding off-wins, publishing and Finish as shown, the rewrites; the
hostile-instance test, its invariant and what it found, on the final code
and on the code before; the residual risks as they stand. Learn's
AGENTS.md gets the rule that every column deciding a key is bound into
what is sealed or signed, and the test.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
Publishing is refused on the device where anything in the tree is sealed,
before any tag's name goes out: the tree and the tags' links are the
instance's to change. A suggestion leaves sealed files out, as the
instance could name another private tree's file through a planted row.
Docs: ADR 0019's "The short reading (22 Oct 2026)", and the general
residual risk that any device can be held at an earlier genuine record;
Learn's AGENTS.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>
Learn's AGENTS.md says how private studysets work (keys, the envelope,
locked rows, rewriting a tree) and what changed for the connector,
publishing, uploads and the learner count; ADR 0019 records what was built
otherwise than decided or not yet (a private studyset's files); the
halkyon_app README describes the content cipher and device keys.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
ADR 0017's rules as they stand, halkyon_app's README and Learn's
AGENTS.md: where an instance offers passwords, the apps offer no Add GitHub
and no hint on adding it, an anonymous person's Sign in with GitHub reaches
only the account someone has, and every refusal says why.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
ADR 0017 gets "The short reading": its findings, the developer's four
decisions of 21 Oct, and what changed. "The rules as they stand" says a
reset takes effect 14 days after it was asked at the latest, that a member
keeps the session a reset signed in and deletes nobody while the register
is out of reach, and no longer claims every secret is kept out of every
capture: Android holds them back, Windows from 10 2004, macOS from
screenshots but not ScreenCaptureKit on macOS 15 and later, iOS covers
the app switcher's picture alone, Linux nothing. The module's and
halkyon_app's READMEs and Learn's AGENTS.md 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>
ADR 0017 gets "Reading the fixes (21 Oct 2026)": each finding and what
was done, and "The rules as they stand" brought up to date (resets take
effect once 72 hours outside hold-outs passed and lapse 30 days after
they were asked; GitHub-only accounts go 30 days after the end; a member
takes the register's moment by its own clock; nothing secret copied on
the device). The module README and Learn's AGENTS.md say the same.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
ADR 0017's full-reading part says what the device's fixes did: the
clipboard, plain HTTP refused but on this machine, sheets never left busy,
what Save leaves, secrets out of debug logs, keyboards, screenshots and
backups, memory, and the real cryptography against the register's checks.
halkyon_app's README says what an app must add (its activity's screen
guard), Learn's AGENTS.md the HTTPS rule for instances.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
ADR 0017's "Full reading (21 Oct 2026)": each server finding and what was
done (H1, M1, M2, M3, L1, L4, L5, L6, L7, L10), those accepted with their
reasons (L2, L3, L8, L9), and the device's M1, the owner's proof bound to
its operation and the whole base URL, still to do. "The rules as they
stand" brought up to date; the module README and Learn's AGENTS.md 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>
ADR 0017 records what the fourth reading found and what was done, and
states who changes what of an account with which proofs, as a table the
full reading can check against; the module README and Learn's AGENTS.md
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>
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>
Anyone who knew an address could spend its whole hour: 10 sign-ins, 5
phrase proofs and 3 codes, each new code voiding the last (the second
reading's M3). Now:
- sign-ins begun count per address and network (10 an hour), phrase
proofs likewise (5), and networks as before (30);
- failed proofs count per address from every network: past 30 an hour
(wrong passwords, phrases' signatures by another key, an address nobody
has alike) no sign-in or proof begins for it anywhere, so guessing from
many networks makes at most 30 guesses an hour, while holding an owner
out takes failing without end from at least three; longer windows
would outlive the hourly pruning of the limits' notes;
- a signed-in change of password begins 10 times an hour per account;
- codes by mail count per address and network (3 an hour), per address
from every network (10, no mailbox flooded) and per network (10), and
up to 3 are valid at once, a fourth voiding the oldest, so codes others
ask for don't void the one its owner is typing.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
ADR 0017's step-3 section: codes by mail sign nobody in, the ways in
for accounts without a working password, changes while signed in,
sessions ended in every service, the apps' screens (the QR package and
the alternatives considered), the keys in the keyring, anonymous
people's keys waiting for step 4, and libsodium on devices (results
copied before they leave an isolate; a C compiler and make for every
build). The module's and halkyon_app's READMEs and the AGENTS files say
the same; the snap's build installs make for sodium's build hook.
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 hackathon's judging is over, so learn.halkyon.eu follows main once
more. Pushed together with build 21's release, as released apps up to
build 20 can't sign in against the new server.
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>
ADR 0017 records the decisions taken in building (SRP's identity, the
decoys, the 72-hour account reset proposed for confirmation, GitHub's
deprecation) and what step 3 must do; the module's README and the
AGENTS files say how password sign-in works.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
ADR 0017's step 1. Generic over group and hash (SrpGroup); production
takes RFC 5054's 3072-bit group with SHA-256. k = H(N | PAD(g)),
u = H(PAD(A) | PAD(B)), K = H(PAD(S)),
M1 = H(H(N) xor H(g) | H(I) | s | PAD(A) | PAD(B) | K),
M2 = H(PAD(A) | M1 | K). A and B must lie in 1..N-1 and u must not be
0; a and b are 256 bits from Random.secure(); proofs are compared in
constant time. x comes from the caller (halkyon_keys makes it); RFC
5054's own x is there for the vectors only.
- The register's half in halkyon_base_server (base.dart: srpChallenge,
srpVerify), the device's written by hand in halkyon_base_client
(srp.dart: SrpClient, srpVerifier). Serverpod's generator cleans and
writes only a client's lib/src/protocol/ (serverpod_cli 4.0.4's
_getDirectoriesRequiringCleaning; a generate leaves srp/ untouched).
- Tests: RFC 5054 Appendix B on both halves; the client against the
register, succeeding and every refusal.
- The module's tests depend on its client, and pub resolves every
member's dev dependencies, so Learn's server image takes the module's
client into its workspace too.
- The format check covers the hand-written half.
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 Learn and Todo each held now comes from halkyon_app:
- AccountTiles: whose account on which instance, switching, signing
out (asking first for an anonymous account, as it deletes it) and
deleting it (asking first), the app's own words passed in.
- showAppAbout: the app's name, version, icon and legalese as it gives
them, the source of what the instance runs (AppDependencies gains
sourceCodeUrl), the app's own parts, the licence list.
- resolveAppLocale: the first preferred language the app speaks, never
by country alone, British English otherwise.
- AssetCountryNames, the default of InstanceGate and InstanceScope:
Unicode CLDR 48.2's territory names in English and German
(tool/country_names.py, 7.6 KB), Hong Kong, Macao, Myanmar and
Palestine by their short names as Learn's world map has them, so
Todo says "Germany (EU)" where it said "DE (EU)".
Learn's texts read as before in every language; its About bloc and
gateway, its country names from the world map and the keys only the
moved code used are gone, as are Todo's copies.
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>
People's public profiles are the instance's register's (ADR 0016), so
the module now holds them: the public_profile table (unchanged, so
Learn's migration records versions only; devices get the table empty,
as they hold every module's tables), the profile and profiles
endpoints under `base`, the handle rules, the link checks and the page
at /@handle, whose foot names the app as the host's configuration says
(profilePageNote). Deleting a person deletes their profile in the
module's core.
Hosts read profiles through the module (profileOf, profilesOf,
personWithHandle, fixHandle; saveProfile for their tests), behind an
AccountRegister that is this server's own database for now. Learn's
feedback, discussions and publishing ask it; publishing fixes the
handle just before it writes the release, outside its transaction.
Explore still joins the table in its search, so Learn holds the
register wherever it runs. Learn's app calls the module's endpoints;
a profile not found is ProfileNotFoundException. Todo's database gets
the table too (it was never deployed).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
tool/release.dart releases one app: `dart run tool/release.dart learn
[major | minor | patch] [--dry-run]`, each app configured in Dart beside
it (apps/learn/release.dart). Every step and check of tool/release.sh
is kept: the repository as pushed, resuming after a release commit and
tag, Flutter as CI pins it, the tools, the key and its certificate pin,
secrets asked for without echo and never written down, CI green on the
newest commit, the signed release commit and tag by a key deploys
accept, the APK built from a checkout of the tag and checked, the snap,
the signed checksums, pushing only after both builds passed, the
release on Forgejo, the F-Droid repository, the Snap Store once a
channel is set.
New: tags per app (learn-v0.19.0, Learn's v… counting as its
history); only commits touching the app decide its version (its folder,
the shared packages it uses by path, the root pubspec and lock, its
release files, where its packages were before the move); --dry-run
reads all of it, lists every step and the notes, and what a release
would refuse, changing nothing. The pure parts are tested in
test/release/. Learn releases build 21 with tool/release.sh, then
switches.
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 root AGENTS.md holds what applies to every app and package (how
work is done, Serverpod, checks with Melos and the tools, CI); Learn's
own is apps/learn/AGENTS.md, Todo's a note that it is still the
scaffold. The root README introduces the project, Learn's moves to its
folder; LICENSE.md says which licence covers which directory, Todo's
packages unlicensed until decided; one CONTRIBUTING.md. ADR 0015
records the decision; ADR 0014 and 0003 note what moved.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
