feat(settings): add toggle to disable in-app toast notifications

[pauldambra]

Summary

• Adds a toastNotifications setting (default on) alongside the existing push/dock notification toggles
• Gates all non-error toasts on it in the shared toast wrapper's emit() — a single choke point every toast.* call in the app goes through, so no per-feature changes were needed
• Error toasts are exempt and always show
• Adds an "In-app toasts" toggle to the Notifications settings page, wired into "Reset to defaults"

Test plan

• pnpm --filter @posthog/ui typecheck
• pnpm --filter @posthog/ui test (28/28 passing, including new parameterized suppression cases)
• biome check on changed files

---

Created with PostHog Code

[github-actions]

React Doctor found no issues in the changed files. 🎉

_{Reviewed by React Doctor for commit b039f86.}

[greptile-apps]

_{Reviews (1): Last reviewed commit: "feat(settings): add toggle to disable in..." | Re-trigger Greptile}

[pauldambra]

QA Swarm review complete. See inline comments.

[pauldambra]

Note

🤖 Automated comment by QA Swarm — not written by a human

Multi-perspective review: qa-team (specialists + generalists), paul-reviewer, xp-reviewer, security-audit

Verdict: 💬 APPROVE WITH NITS

Small, well-scoped change. The centralized gate in toast.tsx's emit() is the right call — one choke point instead of sprinkling guards across every feature. No CRITICAL/HIGH findings; one MEDIUM UX papercut plus several LOW/NIT items, none blocking.

Key findings

• 🟡 MEDIUM (convergent: qa-team/frontend + qa-team/generalist-b + qa-team/copy) — The settings-page "Test" harness ("In-app toast" / "Deep-link toast" Send buttons, NotificationsSettings.tsx:503 & :512) silently does nothing when the new toggle is off, since it routes through the now-gated NotificationBus. Couldn't post this inline since those lines aren't part of this PR's diff (pre-existing code). Suggest disabling those buttons with a hint, or having the harness bypass the gate.
• 🟢 LOW (convergent: paul + xp) — toast.tsx (a primitive) now imports the settings feature store — a mild layering inversion within @posthog/ui. Works today since getState() is read lazily, not a real cycle risk, but worth a thought.
• 🟢 LOW (convergent: qa-team/generalist-a + qa-team/generalist-b + paul + xp) — The suppressed-emit early return (toast.tsx:51) drew independent attention from all four reviewers: it returns a fabricated/unregistered id, gives no observability into suppression, can strand a stale toast if the setting is toggled mid-upsert, and raises the question of whether suppressing loading toasts is deliberate. None of these bite today.
• 🟢 LOW (qa-team/generalist-a) — notifications.ts:49's comment ("the in-app toast always shows") is now stale post-PR. Not inline-commentable (file untouched by this PR).
• 🟢 LOW (qa-team/frontend) — useArchiveTask.ts:200's "Task archived" toast carries the only Undo action; disabling toasts silently drops undo-archive. Worth a deliberate call (exempt action-bearing toasts, or accept the tradeoff). Not inline-commentable (file untouched by this PR).
• 🟢 LOW (qa-team/generalist-b) — Async store hydration means a brand-new session briefly serves the default (true) before the persisted value loads. Benign direction (over-shows, never under-shows).
• ⚪ NIT (xp) — toast.tsx:49's gate condition could be an extracted isSuppressed(level) predicate instead of a comment.

Convergence

• Layering inversion (toast.tsx importing settingsStore) — paul + xp, independently.
• Suppressed-emit return-value cluster (toast.tsx:51) — qa-team/generalist-a + qa-team/generalist-b + paul + xp, four independent angles on the same line.
• Test-harness silent-no-op — qa-team/frontend + qa-team/generalist-b + qa-team/copy, independently within the qa-team pass.

Reviewer summaries

Reviewer	Assessment
🔍 qa-team	APPROVE WITH NITS — one MEDIUM UX papercut (test harness), rest LOW/NIT; security and performance sub-agents found nothing.
👤 paul	"stamp anyway" — clean, well-scoped, loves the parameterized tests; flags the primitive/feature layering and silent-swallow observability as thoughts, not blockers.
📐 xp	Textbook OnceAndOnlyOnce — one choke point instead of ShotgunSurgery across call sites; minor extract-method nit on the gate condition.
🛡 security-audit	0 findings — purely client-side local preference, no auth/network/data boundary crossed.

---

Automated by QA Swarm — not a human review

[stamphog]

The greptile bot raised a valid, unresolved concern on the current head: when toasts are suppressed and no caller-supplied id exists, emit returns "". A caller doing const id = toast.loading("Saving"); toast.dismiss(id) then passes "" to dismiss, which doesn't hit the undefined guard and forwards quillToast.dismiss("") — potentially dismissing unrelated toasts or behaving unexpectedly. The diff doesn't address this, and bot comments with valid concerns that are ignored require escalation.

[pauldambra]

QA Swarm round 2 review complete (follow-up fix commit). See inline comment.

[pauldambra]

Note

🤖 Automated comment by QA Swarm — not written by a human

Round 2 — reviewing follow-up commit b039f86 (fixes the round-1 fabricated-id finding)

Verdict: ✅ APPROVE

All six review passes converged: emit()'s string | undefined return type change is correct and complete, verified against every call site in the repo that captures a toast return value (connectivityToast.ts, ConfigureAgentsSection.tsx:295, useInboxCloudTaskRunner.ts:158 — all safe, two are toast.loading/exempt levels that never hit the suppressed branch). Tests pass (13/13 including the new regression case), typecheck clean.

Key findings

• ⚪ NIT (convergent: qa-team/generalist-a + xp) — the new parameterized test's expected tuple element always equals id; could collapse to one parameter. Optional, posted inline.
• 🟢 LOW (qa-team/frontend, re-surfaced) — the pre-existing suppression-before-upsert-lookup edge case (a live stable-id toast created before the setting is toggled off won't receive its upsert/dismiss) was already flagged and resolved as informational in round 1's review — not re-posting as a new thread, just noting it's the same known, deferred item.
• ⚪ NIT (qa-team/generalist-b) — toast.error's signature now advertises string | undefined even though it can never actually return undefined (error is exempt from suppression); a narrower per-level overload would encode the invariant, but for one call site the current shape is a defensible simplicity trade-off.

Nothing blocking. Ready from qa-swarm's side.

---

Automated by QA Swarm — not a human review

[github-actions]

Clean feature addition with tests. The original empty-string ID bug (flagged by bots on an older commit) was fixed — suppressed emit() now returns undefined or the caller-supplied stable id, and connectivityToast.ts was updated to match. All inline comments are resolved. The soft layering concern (primitives importing a feature store) is acknowledged and non-blocking given the lazy getState() call.