moving

docs/ux-refactor/ux-wave2-plan-2026-07-04.md

@@ -0,0 +1,87 @@
+# UX Wave 2 Plan — 2026-07-04
+
+Source audits: `ux-audit-2026-07-04.md` (F1–F12) + `ux-audit-2026-07-04-part2.md` (F13–F32).
+Wave 1 (F1–F5, F8–F11) shipped in working tree; this wave targets the
+trust-critical consistency cluster + part-1 leftovers.
+
+## Clusters
+
+### W2-A — Workspace context stability (F13, P1)
+Visiting `/` or `/workspaces` must NEVER silently switch the active workspace.
+Find what mutates active-workspace state on those routes (likely a "most recent"
+auto-select effect in Home/grid mount) and make selection explicit-only (click).
+Breadcrumb/switcher must agree with the workspace the user last chose.
+
+### W2-B — One truth for workspace counts + switcher (F14 partial, P1)
+Home "4 workspaces waiting", switcher ~14, grid "All 55", notification "54".
+Diagnose each source (endpoint/filter). Likely: grid counts ALL incl. dev-noise
++ archived; home counts a "recent" subset; switcher caps at N without saying so.
+Fix: single selector/source-of-truth for workspace counts; switcher gets
+current-workspace checkmark + overflow indicator ("14 of 55 — view all").
+Card metadata "—" placeholder rows (F21): hide missing metadata instead of "—";
+filters that can never match (Virtual/Local/Team all 0 while All=55) either get
+real predicates or are removed; drop the "Table — soon" chip.
+
+### W2-C — One truth for the current model (F15, P1)
+Top-bar chip `claude-sonnet-4-6` vs Settings Model Pilot "Claude Opus 4.6" vs
+raw IDs in New Agent. Diagnose: chip probably shows per-chat model; Model Pilot
+shows global default; New Agent lists raw router IDs. Fix: consistent display
+names via one formatting util (grep for existing model-name formatter); Model
+Pilot fallback must not default to == primary; New Agent model list uses same
+labels as everywhere else.
+
+### W2-D — Memory numbers agree (F16, P1)
+Timeline "617 entities · 9420 relations" vs Graph tab "No knowledge graph data";
+top-bar brain 543 vs Trust "452 memories"; Trust "0 high confidence & fresh"
+above rows labeled "fresh". Diagnose each pair: different endpoints? stale
+cache? per-workspace vs global scope? Graph-tab emptiness while KG has data is
+the highest-value bug (feature looks broken). Fix root causes, not copy.
+
+### W2-E — Onboarding polish batch (F6, P2)
+(a) "English (US)" disabled-button pill → plain text badge (or working picker if
+one exists); (b) remove duplicate footer "Back" (keep header ← Back);
+(c) "help with" chips: default NONE selected (placeholder greeting preview until
+≥1 chosen); (d) template picker: rank/badge "Recommended" from who-are-you
+role/work-type (pure mapping + test); (e) final step: restore Back + step
+counter; (f) final-step suggestion chips filtered to the chosen template.
+
+### W2-F — IA naming + empty states (F7, F12, F19 partial, P2)
+Align: sidebar "Agents & tasks" ↔ page "Agent Center" → both "Agents"; sidebar
+"Library" ↔ page "Artifacts" → page header "Library" (artifacts remain the
+content type); "EVERYTHING ELSE" section label → "Workspace-independent" intent,
+pick something plain ("General"). Agent Center sparse state (1 row + honeycomb
+void): add template suggestion cards under the row. Files tab empty state gets
+an Upload action; Artifacts empty-state "Browse in Library" becomes a real link.
+
+### W2-G — Briefing modal a11y + chat error history parity (F27 + wave-1 residual)
+Login-briefing modal: Escape closes; ✕ gets aria-label; "Start Working" disabled
+until content ready; drop duplicated greeting (visible identically on Home
+behind it). Chat: persisted failed assistant turns ("Generation failed: …")
+re-render as ErrorBlock on reload (reuse isAuthShapedError), and the
+server-persisted failed pair must not duplicate after a local Retry + reload.
+
+## Deferred to wave 3 (product decisions or heavier)
+- F17 Erase-all placement, F18 Never-ask friction (safety UX — quick but opinion-laden)
+- F20 marketplace triplicates (needs catalog-merge design)
+- F22 dedup, F23 wiki quality floor, F24 weaver explainer
+- F25 notifications rework, F26 team invite moment, F28 responsive top bar
+- F29 locale, F30 copy nits, F31 tier naming (Solo vs Free), F32 stale drawer/breadcrumb
+
+## Incidental defects from wave-1 QA (2026-07-04, not UX-wave scoped)
+- Ollama cloud models fail every send with server error "The database connection
+  is not open" (reproducible across Retry; ErrorBlock renders it correctly).
+  Smells like the LiteLLM no-DB bearer-key class of bug — backend triage needed.
+- Onboarding template pick that collides with an existing workspace slug 409s
+  and silently continues INTO the existing workspace — user never learns their
+  "new" workspace wasn't created. Needs suffixing or a visible choice.
+- Transient F2 window: during the auto-sent first task's streaming, the text
+  briefly also sat in the composer before clearing — an Enter there would
+  double-send. Cosmetic, final state correct.
+- The "The database connection is not open" send failure is provider-agnostic —
+  it hit an Anthropic retry too, not just Ollama cloud. Intermittent, not
+  reproducible on demand. Backend triage (the LiteLLM no-DB class, above).
+- Probe truth-gap (deeper than the wave-3 copy fix): the model gate verifies a
+  PROVIDER KEY, not the workspace chat's configured default model. Product
+  decision whether the gate should probe the actual default model instead.
+- Workspace switcher: the active row lacks the "· time ago" disambiguator when
+  `lastActive` is null, so it can't be told apart from same-named rows.