2026-08-26 -- Phase 3 batch N: attention without interruption

The last four Phase 3 items (UXA-4/7, UXIA-2/5). Two are about WHERE the panel asks for the user’s attention when a tool needs permission (the always-visible composer; a floating window instead of a view yank), two about what the eye finds when scanning (fifteen equal-weight settings cards; history rows that never said which model ran). Same discipline as batches L/M: every decision is a pure helper with a table test.

UXA-4: deny-with-alternative moves to the composer

The inline permission card’s deny-reason field lived inside the card’s COLLAPSED details — invisible unless the user expanded the card, which the default tool variant does not. The ux-spec (section 3.7) had always drawn this affordance in the always-visible composer, and ChatView’s own wiring comment promised “deny + type an alternative” there. Now it does:

• ResolveDenyMessage(inCardText, externalText) — pure: the in-card field wins when non-blank (the Window host’s only entry), else the composer text, else the protocol’s default refusal line.
• The inline card builds NO in-card field any more; it consumes the composer text through ConsumeExternalDenyMessage (ChatView wires it to ComposerView.ConsumeTextForDeny). Consume-on-read: a non-blank read clears the field, so the same words cannot also be sent as a chat message by a later Enter. The read happens only when the in-card field cannot answer, so a Window-host deny never eats a chat draft.
• While a request is pending, the composer placeholder announces the role (“Type what to do instead, then press Deny…”), driven by ChatView’s refresh so it tracks the request’s lifetime.
• The floating PermissionWindow keeps the in-card field: it has no composer.

UXA-7: a permission request stops yanking the active view

RefreshIfDirty force-switched to Chat whenever a can_use_tool arrived with Settings/History active — mid-keystroke, mid-rename. The REASON was real: ChatView’s refresh loop (and its inline/window hosting decision) pauses while Chat is inactive, so the request would sit invisible behind the title badge. The remedy was disproportionate. The floating PermissionWindow is view-independent and self-driving (its own AgentHub.Changed hook + 80ms refresh loop), so PermissionArrivalPolicy.Decide(activeView, windowAlreadyOpen) now returns OpenFloatingWindow for the non-Chat case — opened with autoOpened:true, which keeps keyboard Y/N unarmed per PermissionWindow’s existing guard — and BadgeOnly when Chat is active (ChatView’s hybrid hosting decides) or a window is already up. Forcing a view switch is no longer a possible outcome; the policy enum does not even offer it.

UXIA-2: settings sections gain progressive disclosure

Fifteen always-expanded cards gave “which model runs” and “read-only stderr tail” identical visual weight. AddCollapsibleSection — the Foldout-headed generalization of the danger-zone/agent-overrides pattern that AddFoldoutHeaderIcon already serves — now renders five advanced/plumbing sections collapsed by default: UapOps modules, Extension Profiles, uLoop, CLI, Diagnostics. Everyday sections (Conversation, Model, Custom Instructions, Display, Quick Actions, Notifications, Console Errors, Appearance, Account, About) stay plain expanded cards, and Conversation/Model keep the lead slots the 08-14 polish pass gave them (SettingsViewSectionIconTests’ index assumptions hold). Open/closed state persists per editor session via SessionState under SectionDisclosureKey(sectionId) — the HistoryView view-preference precedent, and the right lifetime for a disclosure preference. Body toggles bubble ChangeEvent<bool> through the foldout’s handler, so the persistence write is target-guarded.

UXIA-5: history rows say which model ran

The row meta line was relative-time + file size — and size is a weak proxy for “which conversation was this”. The SessionIndex shared scan (one bounded pass already extracting preview/ai-title/cwd) gained a fourth latch: the first assistant line’s message.model, skipping TranscriptUsage.SyntheticModelName exactly as TranscriptUsage itself does. A mid-session model switch keeps the FIRST real model — “the model it started on” is the honest one-line label. The meta line becomes time + size + shortened model via FormatRowMeta (reusing SettingsView.ShortenResolvedModel, the Settings dropdown’s own rule), with no dangling bullet when a transcript carries no assistant line. Cost is deliberately absent: measured 2026-08-03, result.total_cost_usd never reaches the transcript on disk, so it cannot be restored — shipping a guessed number would be worse than none.

Honest residuals

• UXA-4: with a permission pending and text typed, Enter still SENDS a chat message (unchanged, deliberate — mid-turn steering is a real workflow); only the Deny button consumes the text as a refusal reason. The placeholder says which button completes the deny flow.
• UXA-7: on a machine where the floating window cannot open (exotic layouts), the title badge remains the only signal — the same fallback the old code had AFTER its force-switch, minus the yank.
• UXIA-2: the five advanced foldouts indent their body by Unity’s default foldout content margin; accepted rather than fighting the control’s internal layout.
• UXIA-5: rows from withText:false builds hydrate ModelName only for VISIBLE rows (HydrateVisibleRows), same staging as preview/title — an off-screen row briefly has no model label, by design.