Settings: a landing page that answers three questions, and a blurb for every screen (v0.21.266) - #69
Merged
Merged
Conversation
Per-key help answered "what is this knob"; nothing answered "what is this screen, and why would I come here". Every settings section hand-wrote its own header description, the Overview showed bare link lists, and the generated reference had section headings with nothing under them — three surfaces, three copies, none of them shared. `settings_help.yaml` gains a `sections` store keyed by screen id, three fields (summary / what / why), all 21 screens authored. `ALL_SECTIONS` in the registry is now the full screen directory: the 14 that own settings plus the 7 that configure their own subsystem (prompt layers, templates, themes) and previously had no server-side identity at all. Manifest v3 carries them as a `sections` block with each screen's label, blurb, and writable-key count — counts derived from the entries in the same response, so "22 settings" cannot drift from what the screen holds. The generated reference is restructured to read like the app: one heading per screen in nav order, blurb, then that screen's written-up settings, then a table of what is left. Previously it was one flat list of documented keys followed by a wall of tables, which does not scale as the cadence fills in. Fixes a real defect while there: every table on that page rendered as literal pipe text. Blocks are joined with a blank line between them, and the generator appended table rows one block at a time, so Markdown saw each row as its own paragraph. Tables are now emitted as a single block. Section blurbs gate fully from day one — there are only ~20 screens and the Overview lists all of them, so unlike per-key help there is no cadence to run. Assisted-by: Opus 5
The Overview knew what you'd changed. On a fresh install that is nothing, so it said "everything is on its shipped default" and stopped — a landing page with nothing to land on. The tiles beneath it were bare link lists. **Search moved to where you land.** State lifts out of SettingsNav into a shared context so the Overview hosts its own box driving the *same* query, with results inline. This is a mobile fix as much as anything: there the nav is a full-screen takeover, so a landing page whose only search lives in the nav is not a landing page. **Getting set up.** Three rows for the things nothing else works without — connect a provider, pick a model per role, add a search key — each read from the manifest and each a link to the screen that fixes it. A key counts as configured because the server reports a non-empty value, never because the value was inspected: secrets arrive redacted and stay that way. The list disappears once every step is done rather than becoming a row of ticks. **Tiles gained substance** — each screen's authored blurb and how many of its settings you've moved, counted from the manifest so they cannot drift. Still no reset from the digest, deliberately: rows land on the control, which has its own reset beside its help and its bounds. A second write path from a list with no context and no undo would be worse, not more convenient. Two fixes found on the way: `ModelPickerField` now takes the field kit's chrome. It sat outside the kit, so every model-valued setting — the most common kind left in the settings area, ~25 of them — rendered with no default chip, no reset, no help, and no `data-setting` anchor. Settings search could find those keys and then fail to land on them. Unbound callers (the profile editors, which have no manifest) render exactly as before, asserted by test. One humanizer, not two. Search capitalised the first letter of the whole path and the digest capitalised each segment, so the same setting read as "Search · max results" in one list and "Search · Max results" in the other — on the same screen. Assisted-by: Opus 5
The Recall panel had the same hole the rest of the area did: its HyDE and Self-Query model pickers shipped unbound, because the picker sat outside the field kit. Both keys are fully written up, so search found them and then dropped the user at the top of the section with no help beside the control. Bound now — the exemplar has no unanchored keys left. Its test stub forwards the anchor so a regression fails rather than passing silently. ADR-17 gains the section-blurb property: the authored-once rule holds one level up, and unlike per-key help it gates from day one — ~20 screens, and `ALL_SECTIONS` is asserted against the client's SECTION_HIERARCHY. The manifest endpoint test now pins to `MANIFEST_VERSION` rather than a literal 2; the shape assertions are the point, and a version bump is a deliberate act elsewhere. Release notes consolidated per the file's own rule (two web bullets into one, two settings-fix bullets into one) to stay inside the size budget, and CLAUDE.md trimmed back under its ceiling. Assisted-by: Opus 5
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Wave 3.2 of the settings campaign. No per-key write-ups in this one — it lands the two things every remaining slice depends on, plus the comprehensive Overview.
Section blurbs — authored once, three surfaces
Per-key help answered "what is this knob". Nothing answered "what is this screen, and why would I come here". Every section hand-wrote its own header description (37 of them), the Overview showed bare link lists, and the generated reference had section headings with nothing under them.
settings_help.yamlgains asectionsstore keyed by screen id — all 21 screens authored, three fields each.ALL_SECTIONSin the registry is now the full screen directory: the 14 that own settings plus the 7 that configure their own subsystem (prompt layers, templates, themes) and previously had no server-side identity at all.Manifest v3 carries them as a
sectionsblock — label, blurb, and writable-key count, the count derived from the entries in the same response so "22 settings" cannot drift from what the screen holds.Unlike per-key help, this gates from day one: there are only ~20 screens, the Overview lists all of them, and a missing blurb is a bare link.
The chrome gap that was blocking the cadence
ModelPickerFieldlived outside the field kit and rendered its own label, so it could not take a binding. ~25 of the remaining keys are model-valued. Every one rendered with no default chip, no reset, no help popover, and nodata-settinganchor — so Wave 2 search could find those keys and then fail to land on them.It now renders through
FieldShellwhen bound. Unbound callers (the profile editors, which have no manifest behind them) render exactly as before — asserted by test. Recall's two pickers and Conversation Context's three are backfilled; the golden section has no unanchored keys left.Overview
SettingsNavinto a shared context, so the Overview hosts its own box driving the same query, results inline. Mobile fix as much as anything: there the nav is a full-screen takeover.Two defects found on the way
The reference is also restructured to read like the app: one heading per screen in nav order, blurb, that screen's written-up settings, then a table of what is left. The old flat "documented keys, then a wall of tables" does not scale as the cadence fills in.
Verification
task test:sterile1243 OKtsccleanruff check api/clean; pyright baseline 0task docs:checkgreen, zero warnings; production build OKmemory:recall_hyde_modelwith anchor, flash, dot, help and reset — a control that had none of it before. 375px pass, no horizontal overflow.Assisted-by: Opus 5