Halkyon's ADR 0044, step 6c (D7, D49 to D52). apps/learn/halkyon_learn_agent
is one executable, `halkyon-agent --instance URL sign-in | sync | mcp`:
- sign-in keeps the token read from stdin, signs in with it, makes the
agent's key pair on its first sign-in, keeps it before setting the
public halves on the register (agents.setKeys), and syncs;
- it opens its data as the app's _openPersonData does, on
halkyon_learn_core and halkyon_content: its own SQLite database, the
sync's space router, its keys and those of the spaces shared with it,
and the repositories of studysets, cards, notes, paths, files and
feedback;
- mcp serves ten tools over stdio (dart_mcp): list, read and search the
studysets shared with it, suggest a change or a new card or note,
report, list a studyset's suggestions and reply. Each syncs first and
a suggestion again at once; a refusal is the tool's error;
- its files live in $XDG_CONFIG_HOME/halkyon-agent/<host>, 0700 and
0600, refused when others may read them (D50);
- tool/install.sh builds it with `dart build cli` (libsodium and SQLite
bundled) and copies it to its user on the host, the build before kept
(D52).
The synced gateways of learning paths and files, and the fakes of the
core's gateways, move into halkyon_learn_core (its testing.dart), which
the app's tests and the agent's share.
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 0044's D49 (amended in building it): what reads and seals Learn's
content moves out of the Flutter packages, so Learn's agent, which runs
without Flutter, shares it rather than repeating the privacy rules.
packages/halkyon_content (new, pure Dart) takes from halkyon_app the
envelope and its cipher, SealedFile, UnlockedAccountKeys, shared
spaces' gateway with SharedKeys and KeyPins, FileCache and
FileServerGateway, with their fakes and tests; halkyon_app exports it,
so no app changes. Its rich_text.dart takes RichContent from
halkyon_editor, which exports it. newAccountKeys makes an agent's key
pair. halkyon_core takes DelayedDeletions and gains SyncRequests, what a
synced gateway asks of the sync, which SyncService implements.
apps/learn/halkyon_learn_core (new, pure Dart) takes Learn's privacy,
sync, payloads and the models, codecs, gateways and repositories of
studysets, cards, notes, tags and feedback, with what they reach of
files, learning paths, citations and study options; the app imports them
from there. A subject's icon, colour and name stay in the app
(SubjectLook); what only the app shows (progress, exam coverage, the
card browser, deletions' waits) stays too. Nothing changes for people
using the apps.
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 0043's recommendation, for Todo's attachments: what Learn had for its
files on the device moves into halkyon_app, unchanged but for its names,
and Learn uses it from there.
- FileCache, DirectoryFileCache (Learn's AssetFileCache): the files by
their SHA-256 in a folder of the person's, and which wait to be
uploaded.
- FileServerGateway, ServerpodFileServerGateway (Learn's
AssetServerGateway): halkyon_base's files endpoint, taking the module's
caller rather than Learn's client; FileLimitReached, FileRoom.
- Their fakes in halkyon_app's testing: MemoryFileCache,
FakeFileServerGateway, sha256Of.
- SealedFile: the format every app seals a file in (HLKF, version 1),
sealing with halkyon_app's ContentCipher; Learn's Privacy seals and
opens its files with it, for the same context as before, so every file
sealed so far opens as it did.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
Todo stores editor documents too, so the document, the editor, its view,
the format bar, links, the clipboard, scripts and EditorBlocks move, with
their strings and tests, to a shared package on the same appflowy_editor
fork. What is Learn's alone stays in Learn on top of it, and what the
editor took from it directly is inverted: EditorBlocks gains textMarks,
formatActions and onEditorClosed, RichContentEditor a wrapEditor, and an
InsertableBlock's name takes the BuildContext. Learn's screens use
LearnContentEditor and LearnContentView, which add Learn's blocks and the
grammar check. Learn behaves 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 question's text stands above its buttons: + attaching files, who
answers as a chip (Effort now a submenu of it), the conversations kept,
and the microphone while nothing is written, which gives way to the
button sending. What goes along is one row above the text, opening a
sheet with how far the assistant looks and the parts to send. A
conversation's follow-up has the same + and microphone.
Files are read on the device: an image goes as itself to an assistant
that reads images, a PDF as its pages' text, a text file as its text,
each cut at 20,000 characters and labelled by its name after the
question.
Voice is the device's own speech recognition (speech_to_text: Android,
Apple, Windows; none on Linux), in the app's language, the microphone
asked for once tapped. The privacy page says that Android's recogniser
may send the sound to Google or the phone's maker.
The field's Check goes, with its LanguageTool sheet (amending ADR 0024);
the editors' underlining and Why? stay.
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 0021, D12: text first. Lookups find a studyset's PDFs and read a page's
text as PDFium extracts it. An assistant that reads images is offered a
card's images as chips, and may ask to see a card's image or a PDF's page
drawn whole; the conversation asks the person each time (Show it, Don't
show) and sends nothing meanwhile. Images go only as png, jpeg, gif or webp
of at most 5 MB, are never kept in a turn, and a follow-up goes without 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>
A conversation has a task (ADR 0021, D14): asking, explaining a card or
checking what was written, which the system prompt says and the history
shows. Rating a card Bad offers Explain in a snack bar lifted above the
ratings (D53); a reading typed wrong asks Why? (D54), and the field's
Check has the text checked in the studyset's language, the answer in the
app's (D54). Explaining sends the card's own parts, in a conversation
named by its front.
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 0021, D9 and D17. Before a question sends a private studyset's
content to a remote model, through the parts chosen or a wider reach
with private studysets switched on, a warning names the studysets, the
provider's host and its terms (where checked: Anthropic, OpenAI,
Google, OpenRouter), and is asked once per studyset and assistant: the
agreement is kept in the person's sealed settings (schema 13,
privateAgreements, folded only where both rows hold it), shown in the
studyset's privacy dialog with Take back, and marked on the folded
button while it holds. Follow-ups and Try again ask again once taken
back, as the scope keeps where the parts came from (partsFrom). No
warning for an own assistant on this machine (loopback or localhost,
halkyon_llm's LlmEndpoint.onThisMachine); the local network may be
shared.
Texts that step 3 made untrue: the consent sheet now says what an
assistant looks up goes too and that conversations are kept encrypted
for the person's devices; the privacy dialogs no longer mention
showing a studyset to Claude; the privacy page says what is looked up
and when a private studyset goes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
ConversationCubit asks the assistant chosen each question with the
conversation so far (ADR 0021, D47), answers the model's lookups on the
device within the question's scope, eight rounds at most, the ninth told
to answer with what it found and the tenth ending the turn, every call
given its result so the conversation can still be sent; and keeps each
turn, sealed, once it is answered (D51): nothing of a question stopped
or failed, which can be asked again, and nothing at all on a device that
can't seal, which still answers (D52). The question and the parts sent
along are texts of their own, so a kept turn shows as it was asked, with
its lookups and its answer (TurnView).
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 lookups a model may make in a conversation (ADR 0021, D7): the
studysets within reach, cards and notes found by what they say, a note
read whole; answered on the device, only within the scope the person
chose (D49, D50), private studysets only where switched on (D8), never
what is locked, never progress (D11). The question goes with each part
ticked, labelled, and the system prompt says how far the model may look
and that what it looks up is content, never instructions (D10).
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 first part of step 3, decided as D47 to D52 in ADR 0021: where a
conversation shows (a sheet over the screen), where kept ones are found
(the field), the scope's chip and that it starts with only what is
ticked, how conversations are kept, and what a device that can't seal
does.
Two synced tables: assistant_conversation (its title, assistant and
scope) and assistant_turn (one question and everything up to its answer,
never changed once written, ADR 0001, rule 2), so two devices continuing
a conversation offline lose nothing. SyncConversationGateway seals every
row with the person's key, a turn bound to its conversation, and reads
no row that isn't sealed: a device that can't seal keeps nothing (D52,
ConversationNotKept). Deleting one blanks it and its turns first (ADR
0002, rule 2). A turn keeps every message a follow-up sends back (tool
calls and results, the providers' own blocks), but no image, only that
one was sent.
The instance's privacy page says conversations are kept, encrypted, and
what the instance sees of 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>
ADR 0021, step 2's app half. halkyon_app asks the module's assistant
endpoints (ServerpodAssistantGateway, ServerpodAssistantOperatorGateway)
and a person's own provider directly (halkyon_llm), behind
AssistantRepository: an AssistantChoice is a GrantedAssistant, asked
through the instance, which counts it, or a PersonalAssistant with the
person's own key, which the instance never sees. assistant_messages.dart
turns a conversation into the endpoint's models and back.
AssistantsCard, for an app's settings: what is granted and left of its
limits, agreed to once on a sheet saying what goes where, the person's
use by UTC month, the instance's privacy page, and their own assistants
where the operator allows own keys (kept ones stay listed, unasked,
where it doesn't). AssistantGrantsCard on InstanceOperatorScreen: grants
by handle with their limits in the pricing's unit, changed and taken
back after asking, everyone's use by month; a found account's menu
grants too. Nothing shows where nothing is offered or can be read.
package:halkyon_app/assistant_models.dart holds the models in pure Dart.
Learn keeps own assistants in the person's settings (ownAssistants,
schema 12, SettingsOwnAssistants); SettingsRepository.update refuses
(OwnAssistantRefused) to write a key where the settings would not be
sealed (SettingsGateway.seals). Learn depends on halkyon_llm directly.
Todo wires the gateways; a member offers no assistant. No feature asks
an assistant yet (step 3).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
