Commit graph halkyon/packages/halkyon_llm/README.md
Author SHA1 Message Date
b4290f71c8
refactor!: packages, paths and app IDs under Halkyon Days' names (ADR 0045, step 3)
Learn's folder is apps/halkyon, Todo's apps/myrmex; every package of
ours takes the prefix halkyone_ (halkyone_halkyon_*, halkyone_myrmex_*,
halkyone_train_*, halkyone_meditate_*, halkyone_app, halkyone_base_* and
the other shared ones; halkyon_oidc_* keeps its name, D11). Halkyon and
Myrmex take the app IDs one.halky.halkyon and one.halky.myrmex (D10);
Train and Meditate keep theirs, while every app's Android namespace,
Kotlin package and method channels move to one.halky. The snaps are
halkyon and myrmex (D17), each built in a clone with its recipe in
place of snap/, as Train's and Todo's were.

Serverpod finds a database's migrations by its module's name, which
changed with the packages':
- each server's new, empty migration ("rename") records itself under
  the new names and deletes the rows under the old ones;
- deploy/redeploy.sh copies each database's record to the new name
  before the migrations (carry_module), so the renamed servers find
  their databases migrated; a copy, never a move, so a deploy failing
  before the migration commits leaves the old rows in place;
- Train and Meditate keep their devices' databases: halkyone_app's
  carryMigrationRecord copies their record before the sync engine
  opens them (openPersonDatabase);
- the client migrations' generated moduleName getters, one line each,
  say the new names: Serverpod's client asserts in debug builds that
  every migration names the module, and create-migration never writes
  a migration's files again.

Train's WorkManager work is enqueued under new names, the former ones
cancelled: the work kept names the workers' former classes.

The release tool: an app's tags of its earlier names count as its
history (legacyTagPrefixes), the shared packages' former paths count
for its version, and every snap is built in a clone (atRoot gone).

The databases, Docker images and Compose project keep their names
until step 5.

BREAKING CHANGE: Halkyon (Learn) and Myrmex (Todo) are new apps to
Android and the Snap Store: installs of app.halkyon.learn and
app.halkyon.todo don't update to 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>
2026-11-12 23:57:02 +01:00
0590a4b43d
feat(llm): a request carries how much the model thinks, in each provider's field
LlmRequest.effort (LlmEffort: low, medium, high, xhigh, max, Anthropic's
names) goes as Anthropic's output_config.effort, as reasoning.effort to
the Responses API and OpenRouter, and as reasoning_effort to OpenAI's chat
completions, Gemini and any other host. endpoint.efforts says which
levels a host takes (Gemini low to high, Mistral none); send refuses
another before anything goes, the fake client too. LlmEffort.outputCeiling
gives the room for output each level wants, as thinking counts towards
the ceiling (Halkyon Learn's ADR 0021, D63).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
2026-10-26 16:21:53 +01:00
6e63312bfa
docs: the assistant answering in Learn, as step 3 built it
Learn's AGENTS.md describes the field with the assistants answering in it
(ADR 0021, step 3): who answers, the scope, the conversation sheet and its
lookups, images, private studysets' warnings, kept conversations and the
tasks; and the testing trap of awaiting a finished subscription's cancel.
The shared packages' READMEs name providerTermsOf and onThisMachine.
Learn's README no longer says there is no end-to-end encryption, which
private studysets changed (ADR 0019).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
2026-10-26 07:17:09 +01:00
038aca5439
feat(llm): halkyon_llm, language models' APIs streamed over plain HTTP (ADR 0021, step 2)
A pure Dart package, no SDKs, one interface (LlmClient) over Anthropic's
Messages API, OpenAI's chat completions, in each known host's dialect and
the local servers' (D45), and OpenAI's Responses API, stateless (D46). A
model's turn keeps what its provider wants back with the tools' results:
Anthropic's thinking blocks, the Responses API's reasoning items and
message phase, Gemini's thought signatures. Failures come by kind, the
provider's words cut and without keys.

Tested on replies written from the providers' documentation, replayed in
pieces of 1, 7 and 4,096 bytes; tool/record_llm.dart records real ones
with a key of one's own, which the tests then run on too (D36).
FakeLlmClient for the code that uses it. BSD-3-Clause, as halkyon_oidc.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Tizian 「ティツィアーン」 <raisondetredev@pm.me>
2026-10-26 07:17:06 +01:00