Add Coach: an in-editor writing coach for drafts

[leomattic]

Closes #189

What this adds

Coach is a new right-rail mode in the draft editor that reviews a draft and helps improve it, aimed at people writing in a non-native language and anyone who wants coaching beyond a spell check.

Open it from the rail. Three sections:

Review (whole document)

Scans the draft into three lenses, each with its own editor underline and a teaching popover on click:

• Grammar (red): real errors (agreement, tense, articles, plurals, spelling).
• Style (amber): wordiness, filler, weak word choice, unnatural collocations ("make a photo" → "take a photo"), repetition, weak verbs, unnecessary passive.
• AI tells (violet): machine-sounding constructions (false contrasts, stock phrases, empty tricolons) so the writer learns to spot them.

Each finding explains why, not just what to change. Pills toggle each lens. The scan auto-runs on entering the tab when the draft changed since the last scan; "Rescan" forces it. A one-line summary shows register (from the scan) plus reading level (Flesch grade, computed locally and live).

Rewrite (selection)

Selection driven, falling back to the current sentence. Six actions, two per line: Beautify, Natural, Fix, Simpler, Rephrase, Humanize. Each returns one suggestion as a before/after diff before you apply. My voice is a placeholder for a later samples-based feature.

Structure (whole document, opt-in)

A separate heavier pass (Sonnet) on organization and flow: buried ledes, paragraphs doing two jobs, weak openings/endings, abrupt transitions. Notes link to the paragraph. Opt-in so the fast loop stays cheap.

Hover popover

Hovering a word shows a definition, the meaning in context, tappable synonyms, and a one-line sentence rewrite. Synonyms and the rewrite apply in place.

Notes

• Coach flags em dashes and other AI tells, so its own output is de-dashed; it never emits what it warns about.
• Reuses the checks pipeline (offset anchoring, decorations, bulk apply) and the language-aid hover from an earlier iteration.
• New helpers (deDash, wordDiff, readability) and the shared model-JSON parser are unit tested.

Testing

• tsc, ESLint, Stylelint clean; 482 unit tests pass.
• Verified end to end via Playwright: each lens flags and applies, Beautify vs Simpler differ clearly, the rewrite diff/apply work, auto-rescan fires only when stale, structure notes jump to the right paragraph, hover popover is rich. Synced onto latest trunk (History tab + voice work); both coexist, no console errors.

Known rough edges

• Register can vary between scans on short drafts (Haiku judgment).
• Category bucketing is occasionally loose (a passive finding once landed under Grammar). Detection is solid; the bucket is tunable with prompt work.

🤖 Generated with Claude Code