Skip to content

fix: make auto-learn reliable across editable fields - #90

Merged
softmaxe merged 1 commit into
mainfrom
fix/auto-learn-reliability
Nov 15, 2026
Merged

softmaxe merged 1 commit into
mainfrom
fix/auto-learn-reliability

Conversation

@softmaxe

Copy link
Copy Markdown
Owner

Summary

automatic paste
  bind the intended input before Command-V
  track delivered ranges per field for five minutes
  reconcile completed corrections per session
  apply clear corrections and judge ambiguous ones in cleanup

Improve recovery, long-field edits, repeated dictation, and correction ownership. Auto-learn and learned hints pause when cleanup is disabled; manual vocabulary stays active. Keep the capsule unchanged and add feedback using the existing toast format.

This PR contains only application code, localization, native resources, and tests. It excludes AGENTS.md, GLOSSARY.md, skill configuration, design/interview documents, prototypes, README changes, and testing-document changes.

Evidence

  • Before: regressions reproduced missed first-second edits, duplicate-passage attribution, incorrect middle-insertion offsets, unconditional Chinese homophone replacement, and relevant corrections crowded out of cleanup hints.
    After: targeted regressions pass. The complete formatting, Clippy, and test checks pass: 669 passed, 14 ignored. This filtered branch matches the tested/reviewed code snapshot exactly.
  • Real isolated AppKit AX smoke passed for preparation, immediate edits, temporary clear, accessible Send-button clearing, and an 18k-character editor tail. Native reads were also checked in a T3 Code composer and Brave ordinary/contenteditable inputs.
  • Native UI screenshots and a timed recording verified toast states and the cleanup dependency. Real IME behavior and live ASR/cleanup accuracy remain manual checks.

Merge Danger

Door: one-way

The additive schema upgrade advances the database from version 3 to 4. Existing records are retained, but reverting to a version-3 binary alone cannot reopen the upgraded database.

Blast Radius: Dictation

Touches native field observation, learned dictionary transactions, cleanup hints, and learning feedback.

@softmaxe
softmaxe merged commit 23bb013 into main Nov 15, 2026
7 checks passed
@softmaxe
softmaxe deleted the fix/auto-learn-reliability branch November 15, 2026 19:11
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant