Cache Usage Guide
July 25, 2026 · View on GitHub
Concept and invariants: cache-overview.md. Adding keys: cache-schema-guide.md.
React Hooks
Import from @data/hooks/useCache.
| Hook | Tier | Signature |
|---|---|---|
useCache | Memory | (key: UseCacheKey, initValue?: V) => [V, (next: V | ((prev) => V)) => void] |
useSharedCache | Shared | (key: SharedCacheKey, initValue?: V) => [V, (next: V | ((prev) => V)) => void] |
useSharedCacheValue | Shared | (key: SharedCacheKey) => V | undefined — read-only observer |
useSharedCacheSelector | Shared | (keys: SharedCacheKey[], selector: (values) => S, isEqual?) => S — multi-key read-only aggregate |
usePersistCache | Persist | (key: RendererPersistCacheKey) => [V, (next: V | ((prev) => V)) => void] |
Value type is inferred from the schema. The writable hooks pin the cache entry (refcounted) — the key cannot be deleted while any hook is mounted; useSharedCacheValue does NOT pin (and never writes a default), so an owner's deletion always passes through. Hooks do not accept a TTL option; using TTL under a writable hook logs a warning and is discouraged (see Design Invariant #4).
Pick the shared hook by writer provenance. If this window writes the key, use useSharedCache. If another process owns it (typically a Main service publishing via setShared) and this window only displays it, use useSharedCacheValue — mounting the writable hook seeds the schema default back into the cache and broadcasts it, which can clobber the owner's value during the mount race. Apply ?? fallback with a reference-stable default (module-level const, or an unconditionally evaluated useMemo — never a hook call on the right side of ??).
Write-only call sites take no hook. If a component only ever writes a key, call cacheService.setPersist(key, ...) directly. const [, setX] = usePersistCache(key) still registers useSyncExternalStore, so every write to that key rerenders a consumer that never reads it. The updater form is what makes dropping the hook safe — it resolves prev at write time, so no render-time snapshot is needed. The test is the data flow, not the destructuring: a site is write-only only if the value it writes never derives from that key's rendered value (a concrete value recomputed from other state qualifies).
The setter accepts a concrete value or a functional updater (prev) => next, like React's useState. The updater resolves against the latest stored value at write time (not the render-time snapshot), so read-modify-write stays correct across an await — prefer it whenever the next value derives from the current one. prev is shallow-readonly: the updater MUST be pure and return a new value (mutating prev in place is short-circuited by isEqual and silently skips the re-render — see Design Invariant #1). Keep it side-effect-free too: don't smuggle a derived value out of the updater (e.g. into an outer variable) to drive post-write work, and don't rely on how often or when it runs — to react to what changed (e.g. dispose resources for removed items), derive it in a useEffect that watches the value. For useSharedCache the updater resolves against the local window's value only; it is not cross-window atomic.
import { useCache, useSharedCache, useSharedCacheValue, usePersistCache } from '@data/hooks/useCache'
// Memory — single renderer
const [generating, setGenerating] = useCache('chat.web_search.searching', false)
// Shared — all windows
const [activeSearches, setActive] = useSharedCache('chat.web_search.active_searches')
// Shared, main-owned — read-only observation with a reference-stable fallback
const EMPTY_JOB_PROGRESS: JobProgress = { progress: 0 }
const progress = useSharedCacheValue(`jobs.progress.${jobId}`) ?? EMPTY_JOB_PROGRESS
// Persist — survives restart via localStorage
const [pinned, setPinned] = usePersistCache('ui.tab.pinned_tabs')
// Template key (schema: 'scroll.position.${topicId}': number)
const [scrollPos, setScrollPos] = useCache(`scroll.position.${topicId}`)
CacheService Direct Usage (Renderer)
Import the singleton:
import { cacheService } from '@data/CacheService'
Memory
// Schema keys (Fixed or Template) — type-inferred
cacheService.set('chat.web_search.searching', true)
cacheService.set('chat.web_search.searching', true, 30_000) // with TTL (ms)
cacheService.get('chat.web_search.searching') // boolean
cacheService.has('chat.web_search.searching')
cacheService.hasTTL('chat.web_search.searching')
cacheService.delete('chat.web_search.searching')
// Casual (Memory tier only, no schema match allowed)
cacheService.setCasual<TopicCache>(`topic:${id}`, data, 30_000)
cacheService.getCasual<TopicCache>(`topic:${id}`)
cacheService.hasCasual(`topic:${id}`)
cacheService.hasTTLCasual(`topic:${id}`)
cacheService.deleteCasual(`topic:${id}`)
Shared
// Fixed key
cacheService.setShared('chat.web_search.active_searches', map)
cacheService.getShared('chat.web_search.active_searches')
// Template key (schema: 'web_search.provider.last_used_key.${providerId}': string)
const k = `web_search.provider.last_used_key.${providerId}` as const
cacheService.setShared(k, 'api-key-id-1')
cacheService.getShared(k)
cacheService.hasShared(k)
cacheService.hasSharedTTL(k)
cacheService.deleteShared(k)
To observe a shared value reactively, use a hook; for an imperative one-shot read, use the TTL-aware getShared. There is no consumer API for TTL-blind physical reads — the hooks' internal snapshot reader is not for business code.
Before the initial sync from Main completes, getShared() returns undefined. Writes before sync are applied locally and broadcast; Main-priority override applies at sync time (see Shared Cache Ready State).
Persist
cacheService.setPersist('ui.sidebar.width', 300)
// Updater form — `prev` is the latest persisted value, resolved at write time.
// This is how write-only call sites write correctly without subscribing.
cacheService.setPersist('ui.tab.pinned_tabs', (prev) => [tab, ...prev.filter((t) => t.id !== tab.id)].slice(0, 10))
cacheService.getPersist('ui.sidebar.width')
cacheService.hasPersist('ui.sidebar.width') // "is overridden", not "is stored" — every key is seeded
cacheService.deletePersist('ui.sidebar.width') // resets to the schema default; keys are fixed and never removed
Persist writes are debounced (200ms) and flushed on beforeunload. localStorage is limited to ~5MB per origin — keep Persist values small.
Main Process Usage
import { application } from '@application'
const cacheService = application.get('CacheService')
Main does not expose casual methods. Main has its own persist storage — an independent JSON file ({userData}/cache.json) accessed via getPersist / setPersist / hasPersist, separate from the renderer's localStorage persist and never shared with it. Renderer-origin persist sync still goes through Main as an IPC relay only.
Internal and Shared Access
// Internal cache (Main-only; free-form string keys)
cacheService.set('myService.scratch', value, 30_000)
cacheService.get<MyType>('myService.scratch')
// Shared cache (schema-typed; authoritative at Main)
cacheService.setShared('chat.web_search.active_searches', map)
cacheService.getShared('chat.web_search.active_searches')
cacheService.hasShared('chat.web_search.active_searches')
Subscribing to Changes
// Exact key, internal cache
this.registerDisposable(
cacheService.subscribeChange<number>('myService.counter', (newValue, oldValue) => {
logger.info('counter changed', { oldValue, newValue })
})
)
// Exact key, shared cache
this.registerDisposable(
cacheService.subscribeSharedChange('chat.web_search.active_searches', (newValue, oldValue) => {
// reacts to writes from any window and from Main itself
})
)
// Template key — fires for every matching concrete instance
const tpl = 'web_search.provider.last_used_key.${providerId}' as const
this.registerDisposable(
cacheService.subscribeSharedChange(tpl, (newValue, oldValue, concreteKey) => {
const providerId = concreteKey.split('.').pop()!
logger.info(`provider ${providerId} rotated`, { from: oldValue, to: newValue })
})
)
Fire semantics, re-entrance rules, and the placeholder / character-set contract are listed in cache-overview.md → Design Invariants. In short:
- Fires only on explicit
set/delete/setShared/deleteSharedand renderer-origin writes relayed via IPC - Never fires immediately on subscribe — call
get()/getShared()yourself for initial state - Same-value writes are suppressed (
isEqualfrom es-toolkit/compat) - Callback errors are caught; other subscribers still fire
Shared Cache Ready State
if (cacheService.isSharedCacheReady()) {
// Initial sync from Main has completed
}
const unsubscribe = cacheService.onSharedCacheReady(() => {
// Fires immediately if already ready, otherwise once sync completes
})
Hooks (useSharedCache) work correctly before ready — they return the local initValue / schema default until Main's state arrives, then update.
Cache Statistics (debugging)
cacheService.getStats() // summary: entry counts, TTL status, hook refs, estimated bytes
cacheService.getStats(true) // per-entry details for every tier
Common Patterns
Cache an expensive computation
function useExpensiveData(input: string) {
const [cached, setCached] = useCache(`entity.cache.input_${input}`)
useEffect(() => {
if (!cached.loaded) setCached({ loaded: true, data: expensiveCompute(input) })
}, [input, cached, setCached])
return cached.data
}
Cross-window coordination
// Window A — functional updater derives from this window's latest local value
const [active, setActive] = useSharedCache('chat.web_search.active_searches')
setActive((prev) => ({ ...prev, [searchId]: state }))
// Window B re-renders automatically on next Main relay. It only displays the
// value, so it observes read-only — no default seeding, no pin.
const EMPTY_SEARCHES: ActiveSearches = {} // module-level: reference-stable fallback
const active = useSharedCacheValue('chat.web_search.active_searches') ?? EMPTY_SEARCHES
Observe a main-owned key (read-only)
Main publishes; this window only displays. The writable hook would seed the schema default back (clobbering the owner during the mount race) and pin the key — use the read-only observer with a reference-stable local fallback:
const EMPTY_JOB_PROGRESS: JobProgress = { progress: 0 }
function useJobProgress(jobId: string): JobProgress {
return useSharedCacheValue(`jobs.progress.${jobId}` as const) ?? EMPTY_JOB_PROGRESS
}
// Fallback depends on props? Evaluate the hook UNCONDITIONALLY, then `??`:
const cached = useSharedCacheValue(key)
const fallback = useMemo(() => getDefaultStatus(isActive), [isActive])
return cached ?? fallback // never: cached ?? useMemo(...) — conditional hook call
Aggregate multiple main-owned keys (read-only selector)
A dynamic number of keys cannot be observed with per-key hooks (Rules of Hooks
forbid hooks in loops). When N values must merge into one derived result, use
useSharedCacheSelector: keys is both the subscription set and the only
snapshot read set; the selector receives the matching values tuple (same order,
undefined on miss) and must not touch cacheService itself:
const EMPTY_TOOLS: McpTool[] = [] // module-level: reference-stable fallback
function useMcpToolsByServer(serverIds: readonly string[]): Record<string, McpTool[]> {
// Derive keys AND the zip source from the same memoized array
const uniqueIds = useMemo(() => Array.from(new Set(serverIds)).sort(), [serverIds])
return useSharedCacheSelector(
uniqueIds.map((id) => `mcp.tools.${id}` as const), // no extra useMemo needed
(values) => Object.fromEntries(uniqueIds.map((id, i): [string, McpTool[]] => [id, values[i] ?? EMPTY_TOOLS]))
)
}
isEqual (default: Object.is plus one level of item-wise comparison for
arrays/plain objects) gates re-renders at the selection level;
Map/Set or domain-value selections need an explicit comparator. Same
zero-side-effect contract as useSharedCacheValue: no default write-back, no
pin. See the hook's JSDoc in useCache.ts for the full consumer discipline.
Bounded recent list (Persist)
const [pinned, setPinned] = usePersistCache('ui.tab.pinned_tabs')
// Functional updater derives from the latest stored value — correct even if
// pin() races another write (e.g. fires after an await).
const pin = (tab: Tab) =>
setPinned((prev) => [tab, ...prev.filter((t) => t.id !== tab.id)].slice(0, 10))
Observe every instance of a template key (Main only)
One subscription covers all providers, including ones registered at runtime:
const tpl = 'web_search.provider.last_used_key.${providerId}' as const
this.registerDisposable(
cacheService.subscribeSharedChange(tpl, (next, prev, concreteKey) => {
const id = concreteKey.split('.').pop()!
// react to rotation for provider `id`
})
)
TTL on a non-hook read path
// Main service or non-hook code path
cacheService.set('search.recent_query_hash', hash, 60_000)
// ... check before recomputing
if (!cacheService.has('search.recent_query_hash')) recompute()
Type-Safe vs Casual
| When | Use |
|---|---|
| Key is known at design time | Fixed key + type-safe method |
| Key has a recurring pattern with a variable part | Template key + type-safe method |
| Key is truly unknown until runtime | getCasual / setCasual (Memory only) |
| Need cross-window dynamic key | Template key on Shared tier — there is no getSharedCasual |
Casual methods type-error if the concrete key matches any schema pattern — that's intentional.
Best Practices
- Pick the tier by lifecycle, not by scope: Memory = regenerable, Shared = cross-window regenerable, Persist = nice-to-keep across restarts.
- TTL belongs on non-hook read paths; hook paths log a warn and may expire between renders.
- Pick the shared hook by writer provenance: this window writes →
useSharedCache; another process owns the key and this window only displays →useSharedCacheValuewith a reference-stable??fallback. Shared expiry is eventually consistent — an observed value may briefly outlive its TTL until Main's tombstone lands (see Design Invariant #4). - Prefer Fixed > Template > Casual. Promote recurring casual keys to Template.
- Keep Persist values small — localStorage is ~5MB per origin.
- For Main-process reactions to cache changes, always wrap the
subscribe*return inthis.registerDisposable(...)so teardown is automatic. - Same-value writes are free — don't add your own equality guards around
set/setShared.