feat(learn): the assistant's field laid out as Claude's, with files attached and questions spoken, and no Check (ADR 0021, D67)

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>

docs/adr/0021-private-instance-and-assistant.md

@@ -431,6 +431,37 @@ assistants' effort work:
   read per provider from any of them (`_agreementsOf`), for what is offered, what is
   admitted and the operator's list alike. A provider moved to another host still asks
   again.
+- **D67 The field is laid out as Claude's apps lay theirs** (the developer's list of 1
+  Nov 2026; voice decided on 2 Nov): the question's text above its buttons, which sit in
+  one row below it: + attaching files, who answers as a chip (its sign and name, its menu
+  the answerers under their providers and, where the operator allows a choice, *Effort*
+  as a submenu, D63, rather than a chip of its own), the conversations kept (D48), and
+  the microphone while nothing is written, giving way to the button sending once
+  something is. What goes along (how far it looks, D49 and D50, with the private switch,
+  D8, and the parts of the topic) is one row above the text, "Only what I choose · 2
+  parts", opening a sheet to choose them (recommended, over the chips' rows the field
+  had, which on a card being studied took four lines of a phone's screen). A follow-up in
+  the conversation has the same + and microphone.
+  - **Files** (recommended): + picks images, PDFs and text files (`AssistantAttachments`,
+    read on the device): an image goes as itself, only to an assistant that reads images
+    (D12's types and 5 MB), and as images do, its turn keeps only its label; a PDF goes
+    as its pages' text as PDFium reads it, a text file as its text, each cut at 20,000
+    characters and kept with its turn, as the parts chosen are. Each goes after the
+    question, labelled by its name, and the system prompt reads it as the person's
+    material, never as instructions (D10). Nothing is uploaded to the instance but the
+    question it forwards; a private studyset's warning (D9) is not asked for a file, as
+    the person picked it from the device to send it.
+  - **No Check in the field** (the developer's list, amending D54 and ADR 0024's D1 and
+    D3): the button checking what was written goes, and with it the field's LanguageTool
+    sheet; the editors' underlining (ADR 0024) and Why? after a reading typed wrong stay.
+    Conversations kept as checks still read as before (`AssistantTask.check`).
+  - **Voice is the device's own speech recognition** (the developer's choice of 2 Nov,
+    recommended, over a model run on the phone, a download of hundreds of megabytes, and
+    over the provider's transcription, which the instance would have to forward):
+    `speech_to_text` (Android's recogniser, Apple's, Windows'), in the app's language,
+    asking for the microphone only once it is tapped; on Linux there is none, and the
+    field shows no microphone. On Android the recogniser may send the sound to Google or
+    the phone's maker, which the instance's privacy page says; the instance never gets it.
 
 As built on 26 Oct, for v0.22.0 (build 25, which the server needs: `minimumAppBuild` 25,
 as older apps read limits per grant), D57–D63 read so where they left something open:

docs/adr/0024-languagetool-sidecar.md

@@ -28,6 +28,9 @@ directly, a checker sees what people write while they write it.
 
 ## Decisions (26 Oct 2026, the developer's)
 
+Amended on 2 Nov 2026 by ADR 0021's D67 (the developer's): the field no longer has Check,
+so D1's Check, D3 and D4 no longer hold; the endpoint and the editors' underlining stay.
+
 - **D1 The sidecar, a server endpoint, the field's Check and underlining in the editors**
   (the developer's choice, over the sidecar with Check alone, which was recommended as
   the smaller step, and over the sidecar with the endpoint alone): the checks reach every