Skip to content

Releases: johannes-kaindl/lingotuner

0.6.1

Choose a tag to compare

@github-actions github-actions released this 03 Oct 22:34

Changed

  • The changelog is now written entirely in English.

0.6.0

Choose a tag to compare

@github-actions github-actions released this 30 Sep 22:18

Added

  • The GitHub release now also carries a ready-to-unpack lingotuner.zip (the plugin folder with main.js, manifest.json and styles.css) and a checksums.sha256 file. For a manual install, download the zip and unpack it into .obsidian/plugins/ instead of creating the folder and saving three files by hand.
  • Apple Intelligence (on-device) as an endpoint. With the LLM Endpoint Manager 0.4.0 the endpoint dropdown also offers "Apple Intelligence (on-device)"; a tune then runs through an Apple Shortcut instead of HTTP. Limits: the answer arrives all at once (no stream), is capped at about 4096 tokens, Shortcuts briefly comes to the front, and only one request runs at a time. The Model field stays empty. Failures of the shortcut (timeout, busy, cancelled, error) show their own message instead of an HTTP error. See the new guide "Use the Apple Intelligence endpoint".

Changed

  • obsidian-kit 0.46.0 and code-kit 0.9.0 (was 0.43.0 and 0.7.0, chat client 0.44.0). Brings the shortcut bridge and transport choice; the endpoint dropdown, model picker and other kit parts are updated with it. A source that uses the Shortcuts transport no longer probes a backend.

0.5.0

Choose a tag to compare

@github-actions github-actions released this 26 Sep 22:24

Changed

  • Chat client from obsidian-kit 0.43.0 (createChatClient), replacing the local XHR client. Visible effects:
    • After the first token, a stall of 120 seconds without data now ends the run with "No answer within 120 seconds". Before, only the wait for the first token was limited (setting Timeout), and a stream that went silent later hung forever.
    • A server that refuses the XHR stream (origin / CORS check) is retried once without streaming through Obsidian's requestUrl; the answer then appears all at once. The "stream blocked (CORS)" message only shows when that request fails as well.
    • HTTP errors carry the server's message on one line, also for 200 answers that hold an error body, and a context overflow is reported as an HTTP error.
    • The Lab recording measures ttftMs from the first byte of the answer (also reasoning) as before, now taken from the client's timing.
  • Lab recording through logToLab (kit lab-client): a Lab that returns a promise instead of an id no longer slips through, and a Lab with another apiVersion is named once per session in the console instead of looking like "no Lab".
  • Kit vendoring raised to 0.43.0; the copied styles for the endpoint list (> child selectors, so a nested setting keeps its label) and the streaming area (empty status and reasoning slots take no space) follow the kit.

0.4.0

Choose a tag to compare

@github-actions github-actions released this 26 Sep 19:23

Added

  • Help row at the top of the settings with links to the documentation and the issue tracker (English and German).

0.3.3

Choose a tag to compare

@github-actions github-actions released this 26 Sep 16:27

Added

  • Nutzer-Dokumentation. docs/README.md als Index, dazu Getting started und Troubleshooting (die Meldungen wörtlich aus dem Plugin, je mit Ursache und Abhilfe). Die README verlinkt sie im Abschnitt „Documentation“.
  • README-Bilder (Hero, Runden, Einstellungen, Regler), reproduzierbar per npm run shots; Aufnahme-Vertrag in docs/images/README.md.

Fixed

  • Die Zeile der aktiven Runde wurde in der Standardbreite der Seitenleiste links abgeschnitten. Ein Button hält seinen Text auf einer Zeile und zentriert den Überlauf; die Zeile bricht jetzt um. Der GUI-Smoke misst das mit dem neuen Punkt R1 (bei ~300 px, Gegenprobe am alten Stand rot).

Changed

  • docs/LAB.md und docs/SMOKE.md liegen jetzt unter docs/internal/ (Wartungsmaterial, keine Nutzer-Doku).

0.3.2

Choose a tag to compare

@github-actions github-actions released this 25 Sep 13:07

Fixed

  • Die Aufzeichnung in llm-lab war still aus. Der Lab-Client prüfte apiVersion 3, llm-lab liefert seit dem 2026-09-03 die Fassung 4 — readLabApi gab deshalb null zurück, und kein Tunen wurde aufgezeichnet (ohne Fehlermeldung). Der Client spricht jetzt Fassung 4.

Added

  • Jeder Tunen-Lauf trägt eine turnId in der Lab-Aufzeichnung (ein Tunen = eine Nutzer-Handlung), damit das Lab zusammengehörige Aufrufe klammern kann.

0.3.1

Choose a tag to compare

@github-actions github-actions released this 24 Sep 13:11

Changed

  • authorUrl im Manifest zeigt wieder auf das GitHub-Profil (Rückkehr in den Community Store); keine Funktionsänderung.

0.3.0

Choose a tag to compare

@johannes-kaindl johannes-kaindl released this 24 Sep 13:01

Changed

  • Sampling-Werte, Denkstufe und Tokenbudget kommen jetzt aus dem Anfrage-Profil (code-kit
    sampling-profiles) statt aus einer fest eingetragenen Temperatur.
    LingoTuner sendet im
    Modus transform; die bisherige feste Temperatur 0.3 wird zur Modus-Vorgabe 0.2 —
    Verhaltenswechsel, sichtbar im goldenen Request. Der alte Schalter „Denkschritt
    überspringen" wird einmalig zu einer Denkstufe je Modus migriert (suppressThinking: true →
    aus, false → niedrig); das alte Feld wird danach nicht mehr gespeichert.
  • Neuer, aufklappbarer Abschnitt „Anfrage" in den Einstellungen (unter dem Endpunkt-Abschnitt):
    zeigt Modellfamilie und Backend, welche Felder gesendet werden und warum, erlaubt eigene
    Werte je Familie, zeigt die letzte gesendete Anfrage und Abweichungen der laufenden Sitzung.
  • Der Denk-Knopf im Panel erkennt die Modellfamilie jetzt über den Endpunkt (LLM Endpoint
    Manager oder Namensschätzung) statt über eine reine Namensheuristik, und bietet bei Bedarf
    vier Stufen (aus/niedrig/mittel/hoch) statt nur an/aus.
  • Kein chat_template_kwargs/reasoning_budget mehr im Anfrage-Body; gpt-oss bekommt nie
    reasoning_effort: "none" (das Modell denkt dabei nachweislich mehr als bei "low").

Fixed

  • Thinking toggle now shows its state without relying on color alone (UI-STANDARD §8,
    state-button contract) — aria-pressed was missing entirely. Fixed: aria-pressed follows
    the state on every render, icon switches brain ↔ brain-cog for on/off instead of always
    showing brain. Native disabled and the tooltip via think.hint were already correct.
  • Endpunkt-Liste: Stil für Zusatz- und Schlüsselfelder nachgezogen (Kit 0.37.0).
  • Ein API-Lauf (quiet: true) zeigt bei einem fehlgeschlagenen Logbuch-Eintrag keine Notice
    mehr, sondern schreibt wie der Override-Fehler daneben nur nach console.warn — ein
    Fremdaufruf soll dem Nutzer keine unangeforderte Meldung ins Fenster schieben.
  • Der Status-Indikator wird jetzt auch bei einem reinen Patch (patchPanel, ohne Voll-Draw)
    nachgezogen — bislang blieb er stehen, bis der nächste Voll-Draw kam (heute folgenlos, weil
    jeder Aufrufer ohnehin einen Voll-Draw auslöst, aber keine Eigenschaft der Funktion selbst).