our-space

name

our-space

description

Engineering guide for Our Space — a private two-user couples app at github.com/t7sen/our-space (deployed t7senlovesbesho.me, Android me.t7senlovesbesho). LOAD for ANY task in this repo: features, fixes, review, deployment, push routing, biometric gate, dom/sub permissions, Capacitor, Redis, server actions, presence, SSE. Triggers: OurSpace, t7senlovesbesho, Tasks/Rules/Ledger/Notes/Mood/SafeWord, FloatingNavbar, BiometricGate, FCMProvider, PushToast, Honor device, no-GMS, dom/sub, T7SEN, Besho, Sir, kitten. Stack: Next.js 16, React 19, Capacitor 8, Upstash Redis, Firebase Admin, shadcn/ui, Tailwind v4. Hosted-webapp Capacitor shell — server.url loads t7senlovesbesho.me, NOT bundled (no offline, no PWA). Enforces patterns invisible to training: globalThis casts, deferred setState, void vibrate, server-side role checks, FCM-only push, Cairo TZ keys. Skipping produces code that breaks on Android, leaks state, hallucinates removed deps (Serwist, VAPID, web-push), or violates the dom/sub model.

Our Space — Pre-Flight Skill

This skill complements AGENTS.md (the working summary and architectural overview). When this skill is loaded, run the pre-flight checklist below before any code or proposal. Heavy details live in AGENTS.md and references/*.md — load on demand.

The two highest-value sections are kept inline below: the anti-hallucination inventory (Section 2) and the abridged refusal catalog (Section 3). The full refusal catalog is in references/refusal-catalog.md.

---

0. Agent Pre-Flight (Run Every Request)

Before writing code or proposing changes, complete this checklist:

1. Banned scope check → Does the request mention gallery, bucket list, voice notes/audio recording on /permissions, or Telegram / WhatsApp / Signal / any third-party messenger as a notification channel? If yes → refuse, propose /notes//timeline//tasks//rules//ledger for the first two, text + markdown body for the third, and FCM-only (existing pushNotificationToHistory path) for the fourth — there is no sanctioned out-of-band messenger.
2. Architecture conflict check → Does the request imply offline support, PWA features, service workers, web push, or removing server.url? If yes → refuse with rationale from AGENTS.md Section 3.7. Do not implement.
3. Anti-hallucination check → Read Section 2 below before writing imports or env-var references.
4. Role-context identification → Does this involve a state mutation? If yes → identify which author (T7SEN/Besho) is allowed and ensure server-side role check (AGENTS.md Section 3.1; references/auth-and-security.md). For /permissions specifically, getAutoRules is Sir-only and must return [] for Besho — don't relax.
5. Reference routing → Use the table in AGENTS.md Section 13 to decide which references/*.md file to load. Don't skim the body when a reference has the answer. For any /permissions work, load references/permissions.md.
6. Date math check → Need a date key, day boundary, or windowing math? Import from @/lib/cairo-time. Never reinvent Intl.DateTimeFormat helpers at a callsite. Never hardcode +02:00 — DST drift half the year.
7. FCM nullability check → Does this touch push delivery? If yes → treat push:fcm:{author} as nullable per AGENTS.md Section 3.3, never as guaranteed-present.

When unsure, ask the user one targeted question rather than guessing. Guessing on this codebase produces runtime failures.

---

1. Critical Patterns to Apply Automatically

Apply without prompting. Full examples in references/coding-patterns.md.

• Browser globals via globalThis as unknown as { ... } cast — never window/document/navigator directly.
• setState inside mount-time effects or Capacitor callbacks → wrap in setTimeout(() => setState(...), 0).
• vibrate() always prefixed with void.
• Date.now() in render → useState(() => Date.now()).
• 'use server' files export only async — constants live in src/lib/*-constants.ts.
• cookies() and headers() are async (Next.js 16) — await them.
• useSearchParams() requires a <Suspense> boundary at the page level (Next 16 prerender bailout).
• Optimistic UI mutations on existing records → snapshot, mutate, server, rollback-on-error, refresh-on-success. Skip for create-paths.
• Redis writes that depend on each other → redis.pipeline().
• Date-derived keys → Cairo time via MY_TZ, never UTC.
• Mobile-first page padding: p-4 md:p-12 on outer wrappers, gap-4 md:gap-6 on grids, pb-28 md:pb-32 for floating-navbar clearance.
• <TabsContent> that holds form-bearing children must forceMount — Radix unmounts inactive tabs and FormData ignores DOM-absent inputs.
• Cards needing 1Hz/60s ticks own their own setInterval — never tick the dashboard parent.
• Custom interactive surfaces (raw <button>, <Link>) get active:scale-[0.95]. The shadcn <Button> already has its own press feedback.
• Every page subscribes to pull-to-refresh. Wire useRefreshListener(yourFetchCallback) for client pages with state, useRefreshListener(() => router.refresh()) for client pages without, or drop <RefreshListenerForServerPage /> inside Server Component pages. Only src/app/login/page.tsx is exempt (unauth gate). New pages must include it from the first commit.
• Every page has a header back link. Top-level → href="/" "Back". /admin/* sub-pages → href="/admin" "Admin". Canonical shape: ArrowLeft + group-hover slide on the icon. Only src/app/page.tsx (home) and src/app/login/page.tsx (unauth) intentionally omit it — every other page must include it from the first commit. Full snippet in AGENTS.md § 4.
• Never animate filter: blur() and never use mode="popLayout" in hot paths. Android WebView doesn't composite filter — every frame is a full repaint. Stick to opacity + transform. Lone exception: src/app/login/page.tsx (once-per-session).
• Glass cards = backdrop-blur-md + shadow-xl shadow-black/30. Not -xl / -2xl. Heavier glass tanks framerate on Samsung/Honor mid-range. Use will-change-transform only on elements that transform every navigation or every second (page-transition wrapper, navbar pill, CounterCard spans). Overuse explodes GPU layer count.
• Icon-only buttons need ≥24dp effective hit area. Use p-1.5 minimum for inline icons; p-2 for primary actions like panel close. Never opacity-0 group-hover:opacity-100 for actions a mobile user needs to reach — there is no hover.
• Form submit success effects call void hideKeyboard() from @/lib/keyboard so the soft keyboard dismisses with the form.
• Mobile-friendly form inputs: inputMode, enterKeyHint, autoComplete, autoCorrect, autoCapitalize, spellCheck — set them deliberately. <input type="search"> for search; autoComplete="current-password" for the login passcode.
• Every text-bearing element gets dir="auto". Both inputs (where user types) AND display surfaces (where user-entered fields render). Bilingual app — content auto-flips RTL/LTR by first strong directional character. <Textarea> primitive + <MarkdownRenderer> default to it; everything else — raw inputs/selects/textareas, plus <p>/<span>/<div> rendering .title / .body / .description / .reason / .reply / .terms / .notes / .message / .label / .feedback etc. — MUST include dir="auto" as the first attribute. Skip browser-chromed inputs (checkbox/radio/hidden/date/time/number/range/color/file), static labels, timestamps, IDs, status pills, and constants. New components rendering user text include this from the first commit.
• <body> carries suppressHydrationWarning to absorb browser-extension-injected attributes; don't remove.
• Per-page purge + per-item delete are Sir-only on both client (currentAuthor === "T7SEN" gate) and server (role check in the action). Use <PurgeButton> from @/components/admin/purge-button for purge UI; mirror the two-step / heavy-haptic pattern for new per-item destructive controls.
• Soft-delete is the boundary, not del. Every delete* / purgeAll* action calls moveToTrash / moveManyToTrash from @/lib/trash BEFORE the deletion pipeline. 7-day TTL, restorable from /admin/trash. Auxiliary state (reactions, audits, occurrences, streak/count keys, pin-set membership) is intentionally not preserved — only the primary record + index ZSET entry come back.
• Activity feed is logger-driven. logger.interaction / warn / error / fatal automatically write to the activity:log ZSET (capped at 500). Don't call recordActivity directly; let the logger do it. Sir reads at /admin/activity.
• Force-logout = bump session:epoch:{author}. decrypt() checks JWT iat against the epoch (5s in-process cache). All admin destructive actions go through the inspector / push-test / sessions / export / trash / activity surfaces under /admin, gated by src/app/admin/layout.tsx redirect + requireSir() in src/app/actions/admin.ts.
• summonKitten() is the only Sir → Besho push that mirrors safeword's bypass: bypassPresence: true + channelId: "safeword" + priority: "max" + sound: "default". Possessive/dominant copy is fixed in-action. Surfaced as <SummonButton> on the /admin landing.
• <DeviceTracker /> (root-layout-mounted) is the sole writer of device:*. pingDevice runs on mount + 60s heartbeat. Author claim is sticky. Sir reads at /admin/devices. Don't call pingDevice from feature code; don't conflate with usePresence.
• Restraint mode = mode:restraint:Besho. Every Besho-writable server action MUST call assertWriteAllowed(session.author) from @/lib/restraint and return its error if non-null. Sir is never restrained. Safeword is intentionally exempt. Toggled from <RestraintToggle> on /admin landing.
• auth:failures is the failed-login ZSET (capped 100). Written exclusively from login() in actions/auth.ts; read via getAuthFailures() (Sir-only). Don't reuse for unrelated security events.
• Project is on Vercel Hobby; there is intentionally no vercel.json. Hobby rejects builds with cron schedules more frequent than once-per-day. Every cron trigger goes through cron-job.org (configured in the operator's account, minute cadence) hitting /api/cron/* with Authorization: Bearer ${CRON_SECRET}. New cron-style features should add a new /api/cron/* route + an entry in cron-job.org — never via vercel.json.
• /admin/redis is read-only — inspectRedisKey exposes type, TTL, capped preview. No set/del/expire. Don't add mutation; per-feature admin actions are the boundary.
• Finalization is cron-only. getRewardsBundle and any other page-read path MUST NOT call finalizeWeek or catchUpFinalizations — they fire recap FCMs as a side-effect. The obedience-sweep cron is the sole trigger.
• Four functional crons total (plus heartbeat-watch meta-cron). /api/cron/ritual-windows (minute), /api/cron/obedience-sweep (daily), /api/cron/review-window-open (daily, internally gated to Saturday), /api/cron/timer-expire (minute — sweeps directive timers via expireDueDirectives emitting directive_missed, punishment timers via expireDuePunishments emitting punishment_bailed, AND Truth or Dare challenges via expireDueChallenges emitting tod_expired for Kitten recipients only). All bearer-authed via CRON_SECRET. All triggered from cron-job.org. The heartbeat-watch cron watches the others and FCMs Sir if any go stale.
• Test mode (obedience:test-mode, Sir-only) opens current-week claims for diagnostic flow verification. Week itself stays unfinalized; streak + multiplier untouched. Toggle: /admin/rewards Status tab.
• Both RewardItem AND RewardTier carry optional emoji. Both snapshot onto the claim at claim time. Default tier seeds use medal/trophy/crown progression.
• trash:retention-days is the Sir-tunable TTL (1-90, default 7). @/lib/trash reads via 5s cache; new TTL applies only to future deletions.
• Obedience score = recordObedienceEvent from @/lib/obedience at the action site, never from UI/components/hooks. Wired on: tasks.approveTask (on-time/late), rules.acknowledgeRule (within deadline only), rituals.submitOccurrence (Besho-owned, in-window), permissions.createPermission (auto-approve + re-ask penalty) + decidePermission (manual approve), mood.submitMood (state submissions intentionally NOT scored), admin.setRestraintState (off→on transition only), ledger.createLedgerEntry (punishment → ledger_punishment default −10; violation → severity-scaled rule_violation_minor / _moderate / _major defaults −3 / −8 / −20, with the standard ledger_punishment emit suppressed for violations to prevent double-counting; reward entries are pure log, no emit), directive.completeDirective (directive_completed default +5), punishment.completePunishment (punishment_completed default +2) + punishment.bailPunishment (punishment_bailed default −20 — also writes a "bailed" ledger:{id} entry inline, intentionally bypassing createLedgerEntry so the standard ledger_punishment emit does NOT fire), plus the Sir-only manual_adjust from admin.adminAdjustScore. Member key {eventType}:{eventId} makes retries idempotent. Negative missed-event emissions are owned by /api/cron/obedience-sweep (task_missed, ritual_missed, rule_unacked) and /api/cron/timer-expire (directive_missed + punishment_bailed from expireDuePunishments) — don't duplicate them in feature code. Tier-unlock + week-close FCM fire automatically from inside recordObedienceEvent and finalizeWeek respectively — don't wire FCM at the action site for these. Tier count is fixed at 5; reward catalog items carry optional emoji snapshotted onto claims; reward claims are append-only (no delete* exposed). ClaimStatus is five values (pending/delivered/denied/revoked/rerolled); only revoked is fully terminal — rerolled accepts a flip to delivered or denied via the normal deliverClaim / denyClaim paths (Sir reversing his reroll), but can't be re-rerolled or revoked directly. rerollClaim is Sir-only, pending-only, DELs the per-week slot so kitten can re-pick; the slot is NOT re-set on a subsequent rerolled → delivered/denied flip. No obedience emit on reroll. The history pill is labelled "History N×" (not "Changed N×"); audit row note captures sirNote (delivered/denied prior) or rerollReason (rerolled prior).
• Rule-violation ledger entries snapshot the rule body at write time. LedgerEntry.type === "violation" carries ruleId + ruleSnapshot: { title, body? } (body capped at MAX_VIOLATION_RULE_SNAPSHOT_BODY_LEN = 1000) + severity. Future rule edits/deletes do NOT rewrite the entry — same pattern as the reward/tier emoji snapshot. Per-rule index ledger:violations:by-rule:{ruleId} ZSET is pipelined with ZADD ledger:index on create. Sir-only "Log violation" affordance on /rules cards links to /ledger?prefill=violation:${rule.id} (Suspense-wrapped useSearchParams()). Violation count line surfaces on every rule card for both authors. Active rules only — completed rules can't have new violations logged. The auxiliary index is intentionally NOT preserved on restore from trash (orphan-on-restore matches the reactions/occurrences pattern).
• Real-time directive overlay = <DirectiveDialog> mounted in root layout, self-gates on Besho. Single-slot via directive:active:Besho STRING sentinel; reissue while active refuses. State machine: issued (forced full-screen modal) → acknowledged (non-modal pinned strip with countdown) → completed (auto-dismiss after 2s). FCM payload data.kind === "directive" + data.directiveId; <FCMProvider> foreground listener branches on kind and dispatches ourspace:directive-arrived instead of the standard PushToast. Cron-only expiration via /api/cron/timer-expire — page reads MUST NOT mutate state. Obedience: directive_completed (+5) at action site, directive_missed (−10) from cron only. The extraData parameter on sendNotification is the sanctioned way to pipe non-standard FCM data fields (string values only). Suppression: when an active punishment exists, the dialog renders null and re-checks on ourspace:punishment-cleared.
• Punishment timer = <PunishmentOverlay> mounted in root layout BEFORE <DirectiveDialog> (precedence rule). Single-slot via punishment:active:Besho STRING sentinel (TTL = durationSec + PUNISHMENT_ACTIVE_TTL_BUFFER_SEC). State machine: issued (full-screen modal with Begin button — clock paused) → running (full-screen kneel-timer; Complete button locked until now >= endsAt; Bail = two-tap with 1.5s auto-revert) → completed / bailed (brief confirm + dispatch ourspace:punishment-cleared). Background-grace: appStateChange.isActive: false while running starts a 60s timer; expiry calls bailPunishment(reasonTag: "background-grace"). FCM uses data.kind === "punishment" with bypassPresence: true + priority: "high" + channel default (NOT safeword). Cron-only timeout sweep via /api/cron/timer-expire — expireDuePunishments walks running records past endsAt + 60s and bails them with reasonTag: "timeout". Auto-creates a ledger:{id} punishment entry on completion AND on bail with linkedPunishmentId back-reference — written inline in punishment.ts, NOT via createLedgerEntry, intentionally bypassing the standard ledger_punishment emit (the typed punishment_completed / punishment_bailed events are the canonical signal). Bail is terminal — no re-start.
• Truth or Dare game (/games/truth-or-dare) is symmetric — both authors can issue; obedience-scored on Kitten direction only. Game-launcher architecture at /games driven by src/lib/games/registry.ts — adding a new game registers a descriptor + scaffolds its own slug page + actions + Redis prefix; the launcher renders the new tile automatically. Each TOD challenge is a single submission carrying BOTH a truth prompt AND a dare prompt; recipient picks which type to engage with. Single-slot per issuer via tod:active:T7SEN / tod:active:Besho STRING sentinels (each author can have at most one outgoing in flight). State machine: pending (waiting for recipient pick) → picked (recipient locked in truth/dare, response phase) → completed (response submitted) OR refused (recipient bailed, Kitten penalty) OR safeworded (hard-no on this specific prompt, no penalty) OR expired (cron sweep past TTL — pending = 72h, picked = 48h) OR withdrawn (issuer cancelled, no penalty either direction) OR cancelled (Sir admin force-cancel, no penalty). Obedience emits (Kitten recipient only): tod_truth_answered (+3), tod_dare_completed (+6), tod_refused (−6), tod_expired (−10). Safeword and withdrawn fire no obedience event by design. Sir's recipient-side responses move stats counters but no score (no Sir obedience axis). Cron-only expiration via /api/cron/timer-expire — expireDueChallenges walks past expiresAt. Per-author stats HASH at tod:stats:{author} — issued / truthsAnswered / daresCompleted / refused / safeworded / expired / withdrawn; both authors' stats visible to both. Restraint blocks Kitten from issuing AND withdrawing her own challenges (initiating writes) but allows pick / submit / refuse / safeword on a challenge Sir already issued (responsive — otherwise game stalls until restraint lifts; differs from the directive analog where acknowledge also requires assertWriteAllowed). Soft-delete via moveToTrash with feature: "tod_challenges" — Sir-only, terminal-state-only. FCM payload uses data.kind === "tod_challenge"; <FCMProvider> foreground listener dispatches TOD_ARRIVED_EVENT ("ourspace:tod-arrived") — the page subscribes and re-fetches on receipt (no dedicated overlay component — page-as-listener pattern). When recipient is on a different page, the standard PushToast also fires; same-page → dispatch only. Admin lives under /admin/games/truth-or-dare (Active / History / Stats tabs) — force-cancel (no penalty), mass-cancel, per-stat edit, stat reset, purge-all. The 4 new event weights surface in the existing /admin/rewards Weights tab automatically (registered in TUNABLE_EVENT_TYPES via OBEDIENCE_EVENT_TYPES). Active count surfaces on the /admin landing dashboard strip alongside pending perms / claims / crons / errors. New games go in src/lib/games/registry.ts; the file-map convention is src/app/games/{slug}/, src/app/actions/games/{slug}.ts, src/app/admin/games/{slug}/page.tsx, src/app/actions/admin/games.ts (shared admin bucket). Page-bodies extract into src/components/games/{slug}/* (user-facing) and src/components/admin/games/{slug}/* (admin); shared display utilities (STATUS_CHIP, formatRelative, formatCountdown) live in {slug}-constants.ts + {slug}-utils.ts. TOD v2 additions: reactions on completed rows via the generalized <TodReactions> (reactions:tod:{id} HASH); currentStreak + longestStreak fields in TodStats (HASH at tod:stats:{author}) — recipient-side, reset on refused/expired, free pass on safeword/withdrawn; CHANGE_PICK_WINDOW_SEC = 60 lets the recipient flip pick types within 60s of pickedAt (suppressed issuer notification on the flip); per-author prompt library at tod:prompts:{author} (TodPrompt[] capped at 50, private to each author) with getTodPromptLibrary / saveTodPrompt / deleteTodPrompt; cron-driven pre-warning push via warnExpiringChallenges (dedup at tod:fcm:expire-warn:{id} SET NX EX); windowed admin stats via getTodWindowedStats(windowDays?) aggregating the index ZSET by createdAt, surfaced as an All/30d/7d toggle in the Stats tab. TOD v3 additions: FCM response excerpt — submitResponse embeds ~80 chars of the answer in the issuer-side push body (RESPONSE_EXCERPT_LEN, word-boundary truncation via formatResponseExcerpt); admin History tab gains a client-side substring filter input over prompts + response + reason fields (no refetch needed to clear); CLI surface at /api/admin/cli/tod + /api/admin/cli/tod/cancel driven by packages/cli/src/commands/tod.ts (ourspace tod status|issue|cancel) — CLI issuance is hardcoded Sir-to-Kitten; the cancellationReason field on TodChallenge was renamed to adminCancelReason for clarity alongside refuseReason and withdrawReason.
• Reactions are feature-agnostic via reactToTarget(type, id, emoji) + getReactionsForTargets(type, ids[]) in @/app/actions/reactions keyed on ReactionTargetType = "note" | "tod". Notes preserve the legacy reactions:{id} key shape (no migration); new types use reactions:{type}:{id}. The reactionKeyFor(type, id) helper is the single source of truth for storage. UI: <ReactionStrip> is the generic component with variant: "default" | "compact"; <NoteReactions> and <TodReactions> are thin wrappers. Reactions are restraint-gated writes (Kitten blocked when restrained). Adding reactions to a new feature: add the type to ReactionTargetType, write a thin wrapper component, plumb reactions through the bundle reader.

---

2. Things That Do Not Exist (Anti-Hallucination Inventory)

These were removed or never existed. Do not import them, reference them, or write code that uses them. If you find yourself typing one of these — stop.

Removed / nonexistent	Replacement
@serwist/next, @serwist/sw, serwist, workbox-*	None — PWA removed
web-push package, VAPID_* env vars	None — Web Push removed
navigator.serviceWorker, sw.register(), public/sw.js, public/manifest.json	None
src/lib/offline-notes.ts, storePendingNote, getPendingNotes, removePendingNote, PendingNote	None — IndexedDB queue removed
/api/notes/sync endpoint	None — only /api/notes/stream (SSE) exists in notes/api/
push:subscription:{author} Redis key	push:fcm:{author} only
prisma, @prisma/client, SQL migrations	Upstash Redis is the sole datastore; /src/generated/prisma is a stale gitignore entry
Light-mode Tailwind variants	Dark theme is forced via forcedTheme="dark"
tailwind.config.ts / tailwind.config.js	Tailwind v4 is CSS-first; tokens live in src/app/globals.css
pages/ directory, getServerSideProps, getStaticProps	App Router only
VoiceNote type, voiceNote field, MediaRecorder + RECORD_AUDIO permission for /permissions	None — voice notes prototyped on permissions and explicitly removed
Telegram bot, Telegram fallback push, TELEGRAM_BOT_TOKEN / TELEGRAM_*_CHAT_ID env vars, WhatsApp / Signal / any third-party messenger as a notification channel	None — FCM is the sole sanctioned push transport (§ 3.2). Out-of-band messenger fallbacks were considered and explicitly banned
TZ primitives in @/lib/rituals (dateKeyInTz, todayKeyCairo, tzWallClockToUtcMs, previousDateKey, nextDateKey, weekdayOfDateKey)	@/lib/cairo-time — the migration relocated all TZ math; importing from rituals will fail
Inline todayInCairo(), secondsUntilMidnight(), or per-callsite Intl.DateTimeFormat date-key helpers	@/lib/cairo-time exports the canonical versions — never reinvent
Standalone routes /admin/inspector, /admin/sessions, /admin/cooldowns, /admin/time, /admin/timezone, /admin/activity, /admin/notifications, /admin/restraint-history, /admin/auth-log	/admin/devices (Right now + Sessions sections) for inspector + sessions; /admin/health (Cooldowns + Time tabs; Time tab also hosts the Cairo↔Tabuk converter) for the three time/cooldown routes; /admin/logs (Activity / Outbound / Restraint / Auth failures tabs) for the four append-only log routes
currentReviewWeekDate for rewards bucketing	currentWeekKey in @/lib/obedience (containing-week semantic, NOT just-completed)
beforeSend with stack-frame string check (fetchServerAction / server-action-reducer)	ignoreErrors in Sentry.init — message-based, invariant under Turbopack minification
void pingDevice(...) (or any void asyncCall()) for unhandled-rejection-safe fetch	someAsync(...).catch(() => {}) — void discards the return value but does NOT catch rejections

If a search result, training memory, or autocomplete suggests one of these — it is wrong for this codebase. This table is also mirrored in references/anti-hallucination.md for tools that load reference files but not this one.

---

3. Refusal Catalog (Abridged)

Refuse these immediately with a one-line rationale. Do not implement, do not ask for clarification, do not "try a workaround." Full table (14 rows) in references/refusal-catalog.md.

Request pattern	Why refuse
Add a gallery / photo feature, or bucket list	Banned feature surface
Add Telegram / WhatsApp / Signal / any third-party messenger as a push channel	Banned — FCM is the sole sanctioned push transport (§ 3.2)
Re-add PWA / Serwist / service worker	Removed intentionally; conflicts with server.url
Re-add Web Push / VAPID / web-push package	Removed with PWA; conflicts with server.url
Re-suggest voice notes on /permissions	Prototyped and explicitly removed
Use == / != instead of === / !==	Coercion masks bugs in this strict-mode codebase
Skip role check because "the UI hides the button"	Server actions are public endpoints; client is adversarial
Expose getAutoRules to Besho	Sir-only authoring artifacts; gaming risk
dangerouslySetInnerHTML user content	XSS vector — use MarkdownRenderer
Notification dedup / per-author cooldown	Banner pile-up is by design — every event surfaces
Rate-limit safeword or permission submissions	Already layered-protected; refuse without observed glitch
Universal MutationResult typing of all server actions	Preventive refactor with no observed drift; refuse pre-evidence

---

4. Agent Operating Procedure

When this skill triggers, follow this order:

1. Run Section 0 pre-flight. Refuse if banned or architecturally incompatible.
2. State a plan before code for any non-trivial change. Name the file paths you'll touch and the function/symbol you'll edit.
3. Load references on demand per AGENTS.md Section 13. Don't rely on memory of patterns when a reference is one tool call away.
4. Apply Section 1 patterns to every code change automatically. Re-check before submitting.
5. Cite file paths when proposing edits. Format: src/app/notes/page.tsx::handleFormSubmit.
6. Push back on bad ideas, including from the user. Refuse with rationale; offer alternatives. Do not sugar-coat. Examples: user asks for == → refuse, explain coercion. User asks to add Web Push → refuse, point to AGENTS.md Section 3.7. User asks to skip a server-side role check → refuse, explain client adversariality.
7. Surface uncertainty. If a request is ambiguous, ask one targeted question. Do not invent context.
8. No bugs. Re-read every block of generated code before presenting. "Probably works" is a failure mode.
9. Tone: formal, direct, technical. The user is solo-developing this. They want answers, not warmth.

When you finish a non-trivial change, suggest the relevant smoke-test step from references/deployment.md.

---

5. Where to Find Everything Else

• Architectural pillars (3.1–3.7): AGENTS.md Section 3
• Code style, React, TypeScript, naming, UI: AGENTS.md Section 5 + references/code-style.md
• Auth, security, error handling, accessibility: AGENTS.md Section 6 + references/auth-and-security.md
• File map: AGENTS.md Section 11
• Decision heuristics: AGENTS.md Section 12
• Reference index: AGENTS.md Section 13