jskelet 0.2.2 → 0.2.3

package/CHANGELOG.md

@@ -1,315 +1,368 @@
-# Changelog
-
-All notable changes to this project are documented in this file.
-
-The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
-While the project is on `0.x`, minor releases may contain breaking changes; each
-one is listed under a **Breaking** heading.
-
-## [Unreleased]
-
-### Added
-
-- A cache admin panel at `/_jskelet/cache`, turned on with
-  `cache().panel: { enabled: true }` or `JSKELET_CACHE_PANEL=1`. It lists what
-  the in-process tier holds (key, size, status, remaining TTL, dependency count,
-  precompressed bodies for HTML; key and TTL for data), reports whether the
-  Redis tier is connected or bypassed, and runs the operations you would
-  otherwise hand-write an admin route for: targeted invalidation with an
-  optional hard mode, dropping a single entry, clearing either cache, unlinking
-  the shared keys and triggering a prewarm pass. Unlike the dev overlay it does
-  not look at `NODE_ENV`, because "why is this page stale" is a production
-  question — but nothing is mounted until it is explicitly enabled, so the path
-  does not exist by default. Access is a 32-character password regenerated on
-  every process start and printed once to the server log; there is no persistent
-  secret to leak and a deploy revokes old access on its own. The password is
-  never accepted in a query string, three failed attempts ban the IP for 24
-  hours, and every banned or unauthorised response is a `404` rather than a 401
-  that would confirm the panel exists. The panel is excluded from indexing,
-  prewarming and navigation speculation.
-- `dropHtmlCacheKey()` and `dropDataCacheKey()` drop one exact cache key.
-  `invalidateHtmlCache()` matches a path pattern and takes down every query
-  variant of a path, which is the right default for a webhook but wrong when you
-  want `/list?page=2` gone and `/list?page=3` left hot.
-
-- An adaptive per-host rate limit for upstream calls, `cache().upstream`. It sits
-  in the `fetch` wrapper rather than in the prewarm pass, because what spends the
-  quota is the API call, not the page: one render may make one call or twenty, so
-  `prewarm.rps` could never bound the real thing. A token bucket caps the average
-  rate, a concurrency limit caps the calls in flight, and `rate` is treated as a
-  ceiling that the limiter pulls down on its own — a 429 or 503 halves the rate,
-  `Retry-After` stops the bucket for exactly as long as the upstream asked, and
-  clean windows climb back one step at a time. A host that returns
-  `breakerFailures` rate limits in a row is bypassed for `breakerCooldownMs`,
-  which stops the worst waste: because a 429 counts as transient, the HTML
-  produced by a throttled call is never stored, so a pass in that state spends
-  quota and keeps nothing. Only 429 and 503 penalise the rate; a 400 or 500 is
-  not a quota problem. Off by default — set `rate` to turn it on.
-- `getUpstreamLimiterStatus()` reports the current rate, calls in flight, 429
-  count and breaker state per host. The dev panel's Server tab shows the same.
-- `getDataCacheStats()` counts how the data cache was used: fresh hits, stale
-  hits, misses, coalesced concurrent reads, values promoted from the shared tier
-  and — the only number that reaches the quota — real producer runs. A prewarm
-  pass now prints its own share of that (`12 upstream calls for 430 data reads
-  (97% from the data cache)`), which is what tells you whether the fix is a
-  longer TTL or a rate limit. The dev report has a Data cache card for it.
-
-- The dev overlay header now shows the installed JSkelet version next to the
-  title, labelled `latest` when it matches npm and `outdated` with the newer
-  version when it does not, so you can tell at a glance which version the
-  project runs without opening the Server tab.
-- An optional Redis tier behind both caches, turned on with
-  `cache().redis: { enabled: true, url }` and `npm install ioredis`. The
-  in-process cache stays primary and every request still reads it; Redis only
-  does the two things a single process cannot. An instance that has never seen a
-  path finds the HTML another replica already produced, so a fresh container or a
-  post-deploy replacement does not re-render and re-fetch everything from
-  scratch. And `invalidateHtmlCache()`, `clearHtmlCache()` and
-  `clearDataCache()` now reach every replica over pub/sub instead of only the one
-  that received the webhook — until now the others waited out the TTL and a
-  visitor saw old or new content depending on where they landed. Keys live under
-  `_jskelet:{namespace}:{buildId}:…`, where the build id makes HTML from a
-  previous deploy expire on its own rather than pointing at asset files that no
-  longer exist. Personalised (`storable: false`), degraded and non-200 responses
-  are never shared. If `ioredis` is missing, Redis is unreachable or it goes down
-  mid-flight, a warning is printed and the site keeps serving from memory.
-- `getRedisStatus()` reports whether the shared tier is connected, which key
-  prefix and build id it is using, and how many command failures there have
-  been — usable from a healthcheck endpoint. The same summary appears in the dev
-  panel report.
-- Servers started with `startServer()` now shut down on `SIGTERM`/`SIGINT`
-  instead of being killed: the listener is closed and the Redis connection is
-  drained so in-flight writes are not cut mid-command.
-- Targeted HTML invalidation: `invalidateHtmlCache(target, { hard })` takes a
-  path, the config pattern syntax (`/news/:slug`), a regular expression or a list
-  of them, and returns how many entries were affected. By default it **stales**
-  the entries rather than deleting them, so a webhook that touches hundreds of
-  pages does not turn into hundreds of cold renders at the worst possible moment:
-  visitors keep getting the old HTML while the refresh runs in the background,
-  once per key. Matching is done against the path, so every query variant of a
-  page is covered by one call, and a render already in flight when the purge
-  arrives is not stored.
-- `clearDataCache()` now refreshes the HTML too. The `withDataCache` keys read
-  during a render are recorded, so dropping `news:abc` stales every page that
-  actually read it — the article, the home page listing it and the tag page —
-  without the application declaring any tags. Turn it off with
-  `cache().trackDependencies: false`; `getHtmlCacheEntries()` reports the
-  dependency count per page as `deps`.
-- Invalidated paths go to the front of the next prewarm pass, so an updated page
-  is refreshed without waiting for a visitor, while still respecting the `rps`
-  limit. The pass summary counts them separately.
-- An upstream data cache: `withDataCache(key, ttlSeconds, producer)` and the
-  `dataCache(fn, { key, revalidate })` wrapper, with `clearDataCache(prefix?)`,
-  `getDataCacheSize()` and `getDataCacheEntries()` alongside them. It keeps JSON
-  rather than HTML, so its default limit is 10,000 entries: a long-tail page that
-  was never prewarmed still renders without touching the API. Concurrent reads of
-  the same key collapse into one upstream request, an expired entry is served
-  immediately while it refreshes in the background, a failing producer falls back
-  to the stale value, and empty answers (`null`/`undefined`) are not stored
-  unless `storeEmpty: true` is passed.
-- `cache().prewarm.priority` decides the warm-up order and accepts both the
-  config pattern syntax (`/news/:slug`) and plain regular expressions. Matching
-  paths are warmed on every pass.
-- Drip prewarming for large sites: `cache().prewarm.rps` (also `PREWARM_RPS`)
-  caps requests per second regardless of parallelism, and `rotate` (on by
-  default) makes periodic passes continue through the queue where the previous
-  one stopped instead of re-warming the same first slice. A pass is skipped while
-  the previous one is still running.
-- `cache().prewarm.retryDelayMs` (also `PREWARM_RETRY_DELAY_MS`) waits before the
-  retry pass, since rate limit windows are measured in seconds.
-- `cache().maxEntries` configures the HTML cache limit, which used to be a fixed
-  500.
-- Transient upstream failures are now detected without any application code:
-  `globalThis.fetch` is wrapped during startup and `429`, `5xx` and network
-  errors raised inside a render are reported on their own, so rate limits stop
-  turning existing pages into 404s even when the data layer never calls
-  `reportUpstreamFailure()`. Requests outside a render and requests to the
-  server itself are ignored, deterministic answers such as `404` are not
-  reported, and the wrapper can be turned off with `cache().trackUpstream:
-  false`.
-- `cache().transientRetry` (`{ attempts: 1, delayMs: 300 }` by default) retries a
-  page that called `notFound()` while upstream was failing. Each attempt runs in
-  a fresh upstream and per-request cache scope, so a page whose data arrives on
-  the second try is served and cached as usual instead of degrading to an error.
-- The dev report now includes the data cache entry count under `cache.data`.
-
-### Changed
-
-- The prewarm retry pass no longer retries permanent failures. A `400`, `403` or
-  `404` does not get better on the second try, so those paths are dropped from
-  the retry round and counted as `N not retried (permanent)` in the summary. The
-  wait before the round now also honours the upstream rate limit: if a
-  `Retry-After` or an open circuit breaker is holding calls back, the pass waits
-  that out instead of retrying into the same 429.
-- Errors and warnings raised during a prewarm pass are no longer logged one per
-  page. Request errors and the per-page render warnings (`was produced with
-  missing data`, `returned notFound() while upstream is failing`, `could not be
-  produced`) are counted while the pass runs and printed as a single summary
-  block afterwards, grouped by message with the most frequent kinds first, so a
-  failing upstream can no longer bury the "warmed N/M pages" line under hundreds
-  of near-identical lines. Real traffic logs as before, and the dev tools panel
-  still shows the per-path detail.
-- The marketing example's changelog page is now a timeline: releases are laid out
-  along a rail with a sticky version column, each change group gets its own card
-  with a coloured rule and item count, and a row of version chips at the top
-  jumps straight to a release.
-- The dev tools panel is now fed over a WebSocket (`<devBasePath>/ws`) instead of
-  polling `/stats` every two seconds. The server pushes statistics as they change
-  and sends live reload and CSS hot-swap events over the same connection, so an
-  open tab no longer keeps hitting the server while the panel is closed. No new
-  dependency is involved; if the socket cannot be opened, the panel falls back to
-  the previous SSE plus polling path.
-- The server now binds to `::` instead of `0.0.0.0` when no `HOST` is given, so a
-  single dual-stack socket answers both IPv6 and IPv4. Browsers resolve
-  `localhost` to `::1` first and, unlike ordinary requests, a WebSocket handshake
-  does not fall back to IPv4 — which made the dev panel's live channel fail on an
-  IPv4-only socket. Where IPv6 is unavailable the bind falls back to `0.0.0.0`.
-- Prewarming no longer holds up the rest of the dev server. In development it now
-  runs with a single worker and a default limit of 4 requests per second
-  (`prewarm.rps` / `PREWARM_RPS` still override it), so page requests and the dev
-  panel stay responsive while a warm-up round is going on. Production behaviour
-  is unchanged.
-- `notFound()` is no longer served as a 404 when a transient upstream failure
-  (`429`, `5xx`, network error) happened during the same render. The page is
-  retried first and, if upstream is still failing, responds with an uncached
-  `503` and `Retry-After`. A temporary rate limit is no longer frozen into "this
-  page does not exist" for the whole TTL. A retry that gets a clean answer saying
-  the page is gone still returns a normal 404.
-- Responses produced with missing data are no longer offered to shared caches:
-  a `degraded` render is sent with `private, no-store` instead of
-  `public, s-maxage=…`. The `X-JSkelet-Cache` diagnostic header is still written.
-- The prewarm summary distinguishes paths left for the next pass
-  (`700 deferred to the next pass`) from paths dropped entirely
-  (`700 over the limit`).
-- The changelog page of the marketing example is generated from the project's
-  `CHANGELOG.md` instead of a hand-written list, and shows the version published
-  on npm next to the installed one.
-- The marketing example reads its markdown (documentation and changelog) from
-  the repository over GitHub's raw endpoint, falling back to the installed
-  package when the network is unavailable, so a deployment that ships without
-  `node_modules` can still serve the docs. In development the local file wins
-  and nothing is cached. The branch is overridable with `DOCS_REF`.
-
-### Fixed
-
-- The dev panel's WebSocket handshake was answered with a `Sec-WebSocket-Accept`
-  value derived from a mistyped protocol constant. Browsers verify that value and
-  closed the connection immediately with "Incorrect 'Sec-WebSocket-Accept' header
-  value", so the panel silently fell back to polling.
-
-## [0.1.2] - 2026-08-30
-
-### Added
-
-- `route(fn, { private: true })` for pages that depend on the visitor. The HTML
-  cache is bypassed, `cache.html` patterns can no longer turn caching on for
-  that route, and the response is sent with `private, no-store`, `Vary: Cookie`
-  and no ETag.
-- A runtime guard against identity leaks: when a cacheable route reads
-  `Cookie`, `Authorization` or a session field, the rendered HTML is never
-  stored. In development the request fails with an explanation, in production it
-  is served with `no-store` and logged.
-- `fragment()` for layout-less partial responses, with `no-store` and cache
-  bypass built in.
-- CSRF protection. Cross-site state-changing requests are rejected based on
-  `Origin` and `Sec-Fetch-Site`; requests carrying neither header still pass, so
-  webhooks keep working. An optional double-submit token layer is enabled with
-  `security.csrf.token` and rendered into forms by the new `csrfField()` helper.
-- Signed cookie helpers under `jskelet/cookies`: `parseCookies()`,
-  `setCookie()`, `clearCookie()`, `setSignedCookie()`, `getSignedCookie()`,
-  `randomToken()` and `safeEqual()`. Defaults are `HttpOnly`, `SameSite=Lax` and
-  `Secure` outside development.
-- A `security` configuration section: `trustProxy`, `cookieSecret` and `csrf`.
-- `seeOther()` for the post/redirect/get flow, which needs 303 rather than the
-  method-preserving 307 that `redirect()` sends.
-- Island cleanup. A `mount()` function may return a teardown callback; it is now
-  stored and called by the new `unmount(root)` export when the subtree leaves
-  the DOM.
-- Client helpers for partial updates: `swap()` and `startSwapLinks()` for
-  fetching and replacing a region, `enhanceForm()` and `startForms()` for
-  submitting forms without a full page load while keeping the no-JavaScript
-  path working.
-- A fourth example, `examples/dashboard`: sign-in with a signed cookie session,
-  a private page, a paginated table fragment, a CSRF-protected mutation and an
-  island with cleanup, covered by its own `smoke.mjs`.
-- An npm version badge in the `README`, linking to the package page.
-- An English edition of the documentation under `docs/en/`, mirroring every
-  chapter of the Turkish `docs/`.
-- The dev overlay now compares the installed version against the `latest` tag on
-  npm: the Server tab shows the version, marks an `update` chip when a newer
-  release exists and offers the upgrade command. The lookup is cached for six
-  hours, never blocks the server and can be turned off with
-  `JSKELET_VERSION_CHECK=0`.
-
-### Changed
-
-- `trust proxy` is now configurable through `security.trustProxy` instead of
-  being always on. The default is unchanged, but a server exposed directly to
-  the internet should turn it off: while it is on, a client can forge its own
-  `X-Forwarded-For` and rate limiting or audit logs see the wrong address.
-- Every message the framework prints is now English: config, router, render,
-  cache, prewarm, asset and build warnings, CLI output, the project `jskelet
-  init` scaffolds, and the devtools overlay and report interfaces. Visitor-facing
-  status pages still follow `brand.lang` and keep their Turkish translations.
-- The dev overlay and report now show the current JSkelet logo, served with a
-  cacheable response instead of being re-fetched on every navigation.
-
-### Fixed
-
-- A page rendered without `revalidate` used to be sent with no `Cache-Control`
-  at all, while still carrying a strong ETag. HTTP treats such a response as
-  heuristically cacheable, so an intermediate proxy or the browser's back button
-  could store a response meant for a single visitor. Dynamic pages now send
-  `private, no-store` and no ETag.
-- A redirect thrown from a route that reads the session is no longer cacheable
-  either; a stored "you need to sign in" redirect used to follow the visitor
-  even after signing in.
-
-## [0.1.1] - 2026-08-30
-
-### Added
-
-- English `README`, plus `CONTRIBUTING.md`, `CODE_OF_CONDUCT.md`, `SECURITY.md`,
-  a `LICENSE` file, issue and pull request templates, and a CI workflow.
-- An English-first `examples/marketing` with a Turkish translation, serving the
-  package documentation under `/docs` and reading its version, dependencies and
-  bundle sizes from the installed package.
-
-### Changed
-
-- The install instructions point at the npm package instead of the git
-  repository.
-
-### Fixed
-
-- No more white flash between pages: the page background moved onto the root
-  element, so it applies before the body paints. Reduced-motion preferences now
-  switch off the decorative animations as well, not just page transitions.
-
-## [0.1.0] - 2026-08-30
-
-Initial release.
-
-### Added
-
-- Express 5 server with EJS rendering: `createApp()`, `startServer()`,
-  `route()`, `renderPage()`, `renderView()`, `renderNotFound()`.
-- In-process HTML TTL cache with stale-while-revalidate, plus prewarm at boot.
-- Island runtime with visibility, eager and idle hydration strategies, a small
-  cross-island store, and DOM helpers.
-- Configuration through `jskelet.config.mjs`: `brand`, `paths`, `navigation`,
-  `icons`, `fonts`, `clientEnv`, `redirects()`, `rewrites()`, `headers()`,
-  `cache()` and `hooks`.
-- Build pipeline: fonts, SVG sprite from used icons, Tailwind v4 CSS, esbuild
-  bundles with code splitting, webp variants, hashed output and brotli/gzip
-  precompression.
-- Dev server with watch build, CSS hot-swap, automatic restart and a devtools
-  overlay (requests, errors, upstream calls, cache dump, Web Vitals).
-- CLI: `jskelet dev`, `jskelet build`, `jskelet start`, `jskelet init`.
-- Documentation under `docs/` and three examples: `minimal`, `blog`,
-  `marketing`.
-
-[Unreleased]: https://github.com/ayberkenis/jskelet/compare/v0.1.2...HEAD
-[0.1.2]: https://github.com/ayberkenis/jskelet/compare/v0.1.1...v0.1.2
-[0.1.1]: https://github.com/ayberkenis/jskelet/compare/v0.1.0...v0.1.1
-[0.1.0]: https://github.com/ayberkenis/jskelet/releases/tag/v0.1.0
+# Changelog
+
+All notable changes to this project are documented in this file.
+
+The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
+While the project is on `0.x`, minor releases may contain breaking changes; each
+one is listed under a **Breaking** heading.
+
+## [Unreleased]
+
+### Added
+
+- `cache().query`, a pattern → allowlist mapping that decides which query
+  parameters belong to the HTML cache key. An allowlist caches one entry per
+  distinct value of the listed parameters and ignores the rest, so every
+  `?utm_source=…` variant of a path shares one copy; `true` puts the whole query
+  in the key and `[]` ignores it entirely. Parameters enter the key sorted, so
+  `?a=1&b=2` and `?b=2&a=1` are one entry.
+- A cache admin panel at `/_jskelet/cache`, turned on with
+  `cache().panel: { enabled: true }` or `JSKELET_CACHE_PANEL=1`. It lists what
+  the in-process tier holds (key, size, status, remaining TTL, dependency count,
+  precompressed bodies for HTML; key and TTL for data), reports whether the
+  Redis tier is connected or bypassed, and runs the operations you would
+  otherwise hand-write an admin route for: targeted invalidation with an
+  optional hard mode, dropping a single entry, clearing either cache, unlinking
+  the shared keys and triggering a prewarm pass. Unlike the dev overlay it does
+  not look at `NODE_ENV`, because "why is this page stale" is a production
+  question — but nothing is mounted until it is explicitly enabled, so the path
+  does not exist by default. Access is a 32-character password regenerated on
+  every process start and printed once to the server log; there is no persistent
+  secret to leak and a deploy revokes old access on its own. The password is
+  never accepted in a query string, three failed attempts ban the IP for 24
+  hours, and every banned or unauthorised response is a `404` rather than a 401
+  that would confirm the panel exists. The panel is excluded from indexing,
+  prewarming and navigation speculation.
+- Cloudflare cache management, from the panel and from code. Set
+  `JSKELET_CLOUDFLARE_KEY` and `JSKELET_CLOUDFLARE_ZONE_ID` (or
+  `cache().cloudflare`) and the panel gains the CDN tier next to the origin one:
+  purge everything, purge every URL currently held in memory with one button or
+  a single row with `cf purge`, purge by prefix, host or cache tag, toggle
+  development mode, cache level, browser cache TTL, query string sorting,
+  Always Online, Tiered Cache, Regional Tiered Cache and Cache Reserve, clear
+  Cache Reserve, and read the cache hit ratio. This matters because
+  `invalidateHtmlCache()` refreshes the origin while the copy your visitors get
+  keeps being served from the edge until its TTL expires. The same surface is
+  exported as `purgeCloudflare()`, `toCloudflareUrls()`,
+  `fetchCloudflareOverview()`, `fetchCacheAnalytics()`, `fetchPathEdges()` and
+  `getCloudflareStatus()`; none of them throw, so a CDN outage returns
+  `{ ok: false, error }` instead of breaking a publish flow. Long purge lists
+  are batched at Cloudflare's 100-keys-per-request limit and sent sequentially
+  to stay inside the rate limit. The token is read from the environment, is
+  never returned in a response, and only cache related zone settings can be
+  changed.
+- An edge breakdown for a single path: `fetchPathEdges()` reports which
+  Cloudflare colos served it from cache and which went to the origin. This is
+  observation, not inventory — Cloudflare has no endpoint that lists which
+  edges currently hold a URL, and no way to warm an edge you pick, so the panel
+  says as much rather than implying otherwise.
+- `getRedisDetails()` reports where the shared tier actually points — address,
+  TLS, database, namespace, which kinds are shared and whether the purge channel
+  is subscribed — because "connected" alone does not explain a Redis that shares
+  nothing because of a wrong namespace. The password is never part of the
+  output. `inspectRedis()` counts the keys per kind plus `DBSIZE` and
+  `used_memory`; it runs a `SCAN`, so the panel keeps it behind its own button
+  instead of the refresh loop. When Redis is off, the panel explains what a
+  shared tier would buy and shows the memory and disk state of the host instead,
+  which is the number that decides whether `maxEntries` is too high.
+- `dropHtmlCacheKey()` and `dropDataCacheKey()` drop one exact cache key.
+  `invalidateHtmlCache()` matches a path pattern and takes down every query
+  variant of a path, which is the right default for a webhook but wrong when you
+  want `/list?page=2` gone and `/list?page=3` left hot.
+
+- An adaptive per-host rate limit for upstream calls, `cache().upstream`. It sits
+  in the `fetch` wrapper rather than in the prewarm pass, because what spends the
+  quota is the API call, not the page: one render may make one call or twenty, so
+  `prewarm.rps` could never bound the real thing. A token bucket caps the average
+  rate, a concurrency limit caps the calls in flight, and `rate` is treated as a
+  ceiling that the limiter pulls down on its own — a 429 or 503 halves the rate,
+  `Retry-After` stops the bucket for exactly as long as the upstream asked, and
+  clean windows climb back one step at a time. A host that returns
+  `breakerFailures` rate limits in a row is bypassed for `breakerCooldownMs`,
+  which stops the worst waste: because a 429 counts as transient, the HTML
+  produced by a throttled call is never stored, so a pass in that state spends
+  quota and keeps nothing. Only 429 and 503 penalise the rate; a 400 or 500 is
+  not a quota problem. Off by default — set `rate` to turn it on.
+- `getUpstreamLimiterStatus()` reports the current rate, calls in flight, 429
+  count and breaker state per host. The dev panel's Server tab shows the same.
+- `getDataCacheStats()` counts how the data cache was used: fresh hits, stale
+  hits, misses, coalesced concurrent reads, values promoted from the shared tier
+  and — the only number that reaches the quota — real producer runs. A prewarm
+  pass now prints its own share of that (`12 upstream calls for 430 data reads
+(97% from the data cache)`), which is what tells you whether the fix is a
+  longer TTL or a rate limit. The dev report has a Data cache card for it.
+
+- The dev overlay header now shows the installed JSkelet version next to the
+  title, labelled `latest` when it matches npm and `outdated` with the newer
+  version when it does not, so you can tell at a glance which version the
+  project runs without opening the Server tab.
+- An optional Redis tier behind both caches, turned on with
+  `cache().redis: { enabled: true, url }` and `npm install ioredis`. The
+  in-process cache stays primary and every request still reads it; Redis only
+  does the two things a single process cannot. An instance that has never seen a
+  path finds the HTML another replica already produced, so a fresh container or a
+  post-deploy replacement does not re-render and re-fetch everything from
+  scratch. And `invalidateHtmlCache()`, `clearHtmlCache()` and
+  `clearDataCache()` now reach every replica over pub/sub instead of only the one
+  that received the webhook — until now the others waited out the TTL and a
+  visitor saw old or new content depending on where they landed. Keys live under
+  `_jskelet:{namespace}:{buildId}:…`, where the build id makes HTML from a
+  previous deploy expire on its own rather than pointing at asset files that no
+  longer exist. Personalised (`storable: false`), degraded and non-200 responses
+  are never shared. If `ioredis` is missing, Redis is unreachable or it goes down
+  mid-flight, a warning is printed and the site keeps serving from memory.
+- `getRedisStatus()` reports whether the shared tier is connected, which key
+  prefix and build id it is using, and how many command failures there have
+  been — usable from a healthcheck endpoint. The same summary appears in the dev
+  panel report.
+- Servers started with `startServer()` now shut down on `SIGTERM`/`SIGINT`
+  instead of being killed: the listener is closed and the Redis connection is
+  drained so in-flight writes are not cut mid-command.
+- Targeted HTML invalidation: `invalidateHtmlCache(target, { hard })` takes a
+  path, the config pattern syntax (`/news/:slug`), a regular expression or a list
+  of them, and returns how many entries were affected. By default it **stales**
+  the entries rather than deleting them, so a webhook that touches hundreds of
+  pages does not turn into hundreds of cold renders at the worst possible moment:
+  visitors keep getting the old HTML while the refresh runs in the background,
+  once per key. Matching is done against the path, so every query variant of a
+  page is covered by one call, and a render already in flight when the purge
+  arrives is not stored.
+- `clearDataCache()` now refreshes the HTML too. The `withDataCache` keys read
+  during a render are recorded, so dropping `news:abc` stales every page that
+  actually read it — the article, the home page listing it and the tag page —
+  without the application declaring any tags. Turn it off with
+  `cache().trackDependencies: false`; `getHtmlCacheEntries()` reports the
+  dependency count per page as `deps`.
+- Invalidated paths go to the front of the next prewarm pass, so an updated page
+  is refreshed without waiting for a visitor, while still respecting the `rps`
+  limit. The pass summary counts them separately.
+- An upstream data cache: `withDataCache(key, ttlSeconds, producer)` and the
+  `dataCache(fn, { key, revalidate })` wrapper, with `clearDataCache(prefix?)`,
+  `getDataCacheSize()` and `getDataCacheEntries()` alongside them. It keeps JSON
+  rather than HTML, so its default limit is 10,000 entries: a long-tail page that
+  was never prewarmed still renders without touching the API. Concurrent reads of
+  the same key collapse into one upstream request, an expired entry is served
+  immediately while it refreshes in the background, a failing producer falls back
+  to the stale value, and empty answers (`null`/`undefined`) are not stored
+  unless `storeEmpty: true` is passed.
+- `cache().prewarm.priority` decides the warm-up order and accepts both the
+  config pattern syntax (`/news/:slug`) and plain regular expressions. Matching
+  paths are warmed on every pass.
+- Drip prewarming for large sites: `cache().prewarm.rps` (also `PREWARM_RPS`)
+  caps requests per second regardless of parallelism, and `rotate` (on by
+  default) makes periodic passes continue through the queue where the previous
+  one stopped instead of re-warming the same first slice. A pass is skipped while
+  the previous one is still running.
+- `cache().prewarm.retryDelayMs` (also `PREWARM_RETRY_DELAY_MS`) waits before the
+  retry pass, since rate limit windows are measured in seconds.
+- `cache().maxEntries` configures the HTML cache limit, which used to be a fixed 500.
+- Transient upstream failures are now detected without any application code:
+  `globalThis.fetch` is wrapped during startup and `429`, `5xx` and network
+  errors raised inside a render are reported on their own, so rate limits stop
+  turning existing pages into 404s even when the data layer never calls
+  `reportUpstreamFailure()`. Requests outside a render and requests to the
+  server itself are ignored, deterministic answers such as `404` are not
+  reported, and the wrapper can be turned off with `cache().trackUpstream:
+false`.
+- `cache().transientRetry` (`{ attempts: 1, delayMs: 300 }` by default) retries a
+  page that called `notFound()` while upstream was failing. Each attempt runs in
+  a fresh upstream and per-request cache scope, so a page whose data arrives on
+  the second try is served and cached as usual instead of degrading to an error.
+- The dev report now includes the data cache entry count under `cache.data`.
+
+### Changed
+
+- The release history page in `examples/marketing` now shows one release at a
+  time: the newest one is expanded and older releases collapse to a single
+  header row with their date, status and change count. Every release used to be
+  printed open in a two-column grid, which made the page an unreadable wall as
+  soon as a few versions piled up. Version links and the quick-jump strip still
+  work, and they open the collapsed release they point at.
+- The prewarm retry pass no longer retries permanent failures. A `400`, `403` or
+  `404` does not get better on the second try, so those paths are dropped from
+  the retry round and counted as `N not retried (permanent)` in the summary. The
+  wait before the round now also honours the upstream rate limit: if a
+  `Retry-After` or an open circuit breaker is holding calls back, the pass waits
+  that out instead of retrying into the same 429.
+- Errors and warnings raised during a prewarm pass are no longer logged one per
+  page. Request errors and the per-page render warnings (`was produced with
+missing data`, `returned notFound() while upstream is failing`, `could not be
+produced`) are counted while the pass runs and printed as a single summary
+  block afterwards, grouped by message with the most frequent kinds first, so a
+  failing upstream can no longer bury the "warmed N/M pages" line under hundreds
+  of near-identical lines. Real traffic logs as before, and the dev tools panel
+  still shows the per-path detail.
+- The marketing example's changelog page is now a timeline: releases are laid out
+  along a rail with a sticky version column, each change group gets its own card
+  with a coloured rule and item count, and a row of version chips at the top
+  jumps straight to a release.
+- The dev tools panel is now fed over a WebSocket (`<devBasePath>/ws`) instead of
+  polling `/stats` every two seconds. The server pushes statistics as they change
+  and sends live reload and CSS hot-swap events over the same connection, so an
+  open tab no longer keeps hitting the server while the panel is closed. No new
+  dependency is involved; if the socket cannot be opened, the panel falls back to
+  the previous SSE plus polling path.
+- The server now binds to `::` instead of `0.0.0.0` when no `HOST` is given, so a
+  single dual-stack socket answers both IPv6 and IPv4. Browsers resolve
+  `localhost` to `::1` first and, unlike ordinary requests, a WebSocket handshake
+  does not fall back to IPv4 — which made the dev panel's live channel fail on an
+  IPv4-only socket. Where IPv6 is unavailable the bind falls back to `0.0.0.0`.
+- Prewarming no longer holds up the rest of the dev server. In development it now
+  runs with a single worker and a default limit of 4 requests per second
+  (`prewarm.rps` / `PREWARM_RPS` still override it), so page requests and the dev
+  panel stay responsive while a warm-up round is going on. Production behaviour
+  is unchanged.
+- `notFound()` is no longer served as a 404 when a transient upstream failure
+  (`429`, `5xx`, network error) happened during the same render. The page is
+  retried first and, if upstream is still failing, responds with an uncached
+  `503` and `Retry-After`. A temporary rate limit is no longer frozen into "this
+  page does not exist" for the whole TTL. A retry that gets a clean answer saying
+  the page is gone still returns a normal 404.
+- Responses produced with missing data are no longer offered to shared caches:
+  a `degraded` render is sent with `private, no-store` instead of
+  `public, s-maxage=…`. The `X-JSkelet-Cache` diagnostic header is still written.
+- The prewarm summary distinguishes paths left for the next pass
+  (`700 deferred to the next pass`) from paths dropped entirely
+  (`700 over the limit`).
+- The changelog page of the marketing example is generated from the project's
+  `CHANGELOG.md` instead of a hand-written list, and shows the version published
+  on npm next to the installed one.
+- The marketing example reads its markdown (documentation and changelog) from
+  the repository over GitHub's raw endpoint, falling back to the installed
+  package when the network is unavailable, so a deployment that ships without
+  `node_modules` can still serve the docs. In development the local file wins
+  and nothing is cached. The branch is overridable with `DOCS_REF`.
+
+### Fixed
+
+- The dev panel's WebSocket handshake was answered with a `Sec-WebSocket-Accept`
+  value derived from a mistyped protocol constant. Browsers verify that value and
+  closed the connection immediately with "Incorrect 'Sec-WebSocket-Accept' header
+  value", so the panel silently fell back to polling.
+
+### Breaking
+
+- A request that carries a query parameter is now dynamic by default: it is not
+  written to the HTML cache and the response is sent with `private, no-store`,
+  even when a `cache().html` pattern covers the path. Every query variant used
+  to become its own cache entry, which let campaign parameters
+  (`?utm_source=…`) mint unbounded keys and evict real pages from a 500-entry
+  store. Pages whose output genuinely depends on the query keep their cache by
+  listing the relevant parameters under `cache().query`.
+
+## [0.1.2] - 2026-08-30
+
+### Added
+
+- `route(fn, { private: true })` for pages that depend on the visitor. The HTML
+  cache is bypassed, `cache.html` patterns can no longer turn caching on for
+  that route, and the response is sent with `private, no-store`, `Vary: Cookie`
+  and no ETag.
+- A runtime guard against identity leaks: when a cacheable route reads
+  `Cookie`, `Authorization` or a session field, the rendered HTML is never
+  stored. In development the request fails with an explanation, in production it
+  is served with `no-store` and logged.
+- `fragment()` for layout-less partial responses, with `no-store` and cache
+  bypass built in.
+- CSRF protection. Cross-site state-changing requests are rejected based on
+  `Origin` and `Sec-Fetch-Site`; requests carrying neither header still pass, so
+  webhooks keep working. An optional double-submit token layer is enabled with
+  `security.csrf.token` and rendered into forms by the new `csrfField()` helper.
+- Signed cookie helpers under `jskelet/cookies`: `parseCookies()`,
+  `setCookie()`, `clearCookie()`, `setSignedCookie()`, `getSignedCookie()`,
+  `randomToken()` and `safeEqual()`. Defaults are `HttpOnly`, `SameSite=Lax` and
+  `Secure` outside development.
+- A `security` configuration section: `trustProxy`, `cookieSecret` and `csrf`.
+- `seeOther()` for the post/redirect/get flow, which needs 303 rather than the
+  method-preserving 307 that `redirect()` sends.
+- Island cleanup. A `mount()` function may return a teardown callback; it is now
+  stored and called by the new `unmount(root)` export when the subtree leaves
+  the DOM.
+- Client helpers for partial updates: `swap()` and `startSwapLinks()` for
+  fetching and replacing a region, `enhanceForm()` and `startForms()` for
+  submitting forms without a full page load while keeping the no-JavaScript
+  path working.
+- A fourth example, `examples/dashboard`: sign-in with a signed cookie session,
+  a private page, a paginated table fragment, a CSRF-protected mutation and an
+  island with cleanup, covered by its own `smoke.mjs`.
+- An npm version badge in the `README`, linking to the package page.
+- An English edition of the documentation under `docs/en/`, mirroring every
+  chapter of the Turkish `docs/`.
+- The dev overlay now compares the installed version against the `latest` tag on
+  npm: the Server tab shows the version, marks an `update` chip when a newer
+  release exists and offers the upgrade command. The lookup is cached for six
+  hours, never blocks the server and can be turned off with
+  `JSKELET_VERSION_CHECK=0`.
+
+### Changed
+
+- `trust proxy` is now configurable through `security.trustProxy` instead of
+  being always on. The default is unchanged, but a server exposed directly to
+  the internet should turn it off: while it is on, a client can forge its own
+  `X-Forwarded-For` and rate limiting or audit logs see the wrong address.
+- Every message the framework prints is now English: config, router, render,
+  cache, prewarm, asset and build warnings, CLI output, the project `jskelet
+init` scaffolds, and the devtools overlay and report interfaces. Visitor-facing
+  status pages still follow `brand.lang` and keep their Turkish translations.
+- The dev overlay and report now show the current JSkelet logo, served with a
+  cacheable response instead of being re-fetched on every navigation.
+
+### Fixed
+
+- A page rendered without `revalidate` used to be sent with no `Cache-Control`
+  at all, while still carrying a strong ETag. HTTP treats such a response as
+  heuristically cacheable, so an intermediate proxy or the browser's back button
+  could store a response meant for a single visitor. Dynamic pages now send
+  `private, no-store` and no ETag.
+- A redirect thrown from a route that reads the session is no longer cacheable
+  either; a stored "you need to sign in" redirect used to follow the visitor
+  even after signing in.
+
+## [0.1.1] - 2026-08-30
+
+### Added
+
+- English `README`, plus `CONTRIBUTING.md`, `CODE_OF_CONDUCT.md`, `SECURITY.md`,
+  a `LICENSE` file, issue and pull request templates, and a CI workflow.
+- An English-first `examples/marketing` with a Turkish translation, serving the
+  package documentation under `/docs` and reading its version, dependencies and
+  bundle sizes from the installed package.
+
+### Changed
+
+- The install instructions point at the npm package instead of the git
+  repository.
+
+### Fixed
+
+- No more white flash between pages: the page background moved onto the root
+  element, so it applies before the body paints. Reduced-motion preferences now
+  switch off the decorative animations as well, not just page transitions.
+
+## [0.1.0] - 2026-08-30
+
+Initial release.
+
+### Added
+
+- Express 5 server with EJS rendering: `createApp()`, `startServer()`,
+  `route()`, `renderPage()`, `renderView()`, `renderNotFound()`.
+- In-process HTML TTL cache with stale-while-revalidate, plus prewarm at boot.
+- Island runtime with visibility, eager and idle hydration strategies, a small
+  cross-island store, and DOM helpers.
+- Configuration through `jskelet.config.mjs`: `brand`, `paths`, `navigation`,
+  `icons`, `fonts`, `clientEnv`, `redirects()`, `rewrites()`, `headers()`,
+  `cache()` and `hooks`.
+- Build pipeline: fonts, SVG sprite from used icons, Tailwind v4 CSS, esbuild
+  bundles with code splitting, webp variants, hashed output and brotli/gzip
+  precompression.
+- Dev server with watch build, CSS hot-swap, automatic restart and a devtools
+  overlay (requests, errors, upstream calls, cache dump, Web Vitals).
+- CLI: `jskelet dev`, `jskelet build`, `jskelet start`, `jskelet init`.
+- Documentation under `docs/` and three examples: `minimal`, `blog`,
+  `marketing`.
+
+[Unreleased]: https://github.com/ayberkenis/jskelet/compare/v0.1.2...HEAD
+[0.1.2]: https://github.com/ayberkenis/jskelet/compare/v0.1.1...v0.1.2
+[0.1.1]: https://github.com/ayberkenis/jskelet/compare/v0.1.0...v0.1.1
+[0.1.0]: https://github.com/ayberkenis/jskelet/releases/tag/v0.1.0