先登记表，再归档。

All notable user-visible changes are recorded here, newest first. This is an aggregated historical archive, not a release-by-release record; details for individual versions are recoverable from Git history. Current product truth lives in:

• docs/architecture/product-model.md
• docs/release/release-state.json
• docs/release/public-interface-snapshot.json

1.0.0-alpha.5

Rendered as data resolves: part-level streaming, WC admission tiers, runtime convergence. ADR-0157/0158/0159 are accepted (2026-09-25 owner ruling). The qualification evidence boundary — including honest negatives (T1 offline snapshot falsified, third-party components verified shell-born only) — is recorded in docs/release/alpha5-qualification.md.

• Element / Router — rendered as data resolves: opted-in dynamic document GETs may defer eligible compiled text/Region Parts after status, security, headers and Cookie decisions. The same Part Program still drives SSR, fresh DOM and claim; server Part backfill is independent of island hydration. Actions and non-opted-in/static routes retain their non-streaming paths.
• Element — ownership and keyed identity: a single internal tree-shaped lifetime scope owns component activations and child Part/Region resources; it preserves DOM on disconnect and removes owned ranges on replacement. Keyed reorders retain stationary node identity/focus and use bounded moves. Old Part Program material is retained; no cross-alpha migration is implied.
• Router — internal renderer selection: Native/Lit imports and hydration ownership converge behind an internal adapter without a public renderer plug-in SDK or a third rendering path.
• Evidence, not support expansion: the Native reference route exercises a delayed Part, late failure, no-JS tail and ordinary POST action against a non-streamed comparison route. Local raw measurements are advisory. The WC snapshot experiment does not grant Lit-in-stream or T1/T2; a Workers module call is not a real deployment. Bun remains advisory.

1.0.0-alpha.4

Self-consistency train: the shipped package manifest stops lying about a module that installs itself, the candidate gate stops re-proving content it has already proved, and the config, error and navigation surfaces are converged. The source line is 1.0.0-alpha.4; npm publication is a separate, gated step, so docs/release/release-state.json records registry truth (@alpha = 1.0.0-alpha.3) and this train's own version stays a repository baseline until a post-publish sync. No migration steps beyond one retired spelling noted below, so an alpha.3 consumer upgrades with no action.

• Release — the packed Element entry declares its claim seam (#1425): the published sideEffects: false on @openelement/element was wrong, and it shipped a real defect. The default entry installs the compiled claim executor through an import-time seam (src/index.ts bare-imports internal/compiled/runtime/claim-install.ts, which calls installClaimExecutor(...) at module scope), so a consumer's own bundle dropped both halves: the bare import reads as an unused statement in a side-effect-free module and the installer body has no used exports. Every island that had to adopt server-rendered DOM then threw [compiled-kernel] existing DOM in the resolved root needs the claim executor — the packed-starter browser matrix failed 3/3 browsers, and the failure was reproduced in the consumer's own client bundle (island-app-shell-*.js), not inferred. The manifest now names both sides of the edge (./src/index.js and ./src/internal/compiled/runtime/claim-install.js; naming only one still drops the other, pinned by test), and pack-surface.ts gained the fail-closed findUndeclaredSeamInstalls rule: a module-scope installX(...) in a package whose sideEffects does not name that module is exactly the shape that ships looking correct and breaks at runtime. The client-only reduction from #1416 is unaffected — that entry imports neither module. This is the one item here that changes what a consumer downloads, and it takes effect only at the next publish: every 1.0.0-alpha.3 already on npm still carries the wrong manifest, so a starter installed from it today silently does not hydrate.
• Release — the browser-matrix skip is gone (#1425): with the root cause fixed, OPEN_ELEMENT_SKIP_BROWSER_MATRIX and every workflow injection of it are deleted. The packed-starter matrix, the dev continuation probe and the start continuation probe are unconditionally required again, so the qualification that caught this defect runs on every candidate. No tracked file mentions the variable any more (the alpha.3 entry above and ADR-0155 keep their historical narration, which is a record, not a guard).
• CI — tree-SHA evidence reuse (#1425 follow-up, ADR-0156): a green evidence package proves a TREE, not a commit — every record carries sha and tree, every log is bound to the tree by its clean-proof line, and every tarball is byte-hashed. Two commits with the same tree have byte-identical content, so a merge that stales dev no longer forces ~1–2 hours of runner time re-measuring content the suite has already measured. A new read-only reuse job (with exactly contents: read and actions: read) resolves the newest in-window successful run whose API-computed commit tree equals the candidate's and that carries every lane's artifact; the four lanes needs: reuse and, when it resolves, replay a verified no-op that downloads the source run's evidence and stamps reused: {runId, sha}. Reuse is strictly tree equality, and the paths out are fail-closed and total: no match, any API error, no token, or a run outside the retention window all resolve to reused=false and every lane then runs its full gate. The aggregate does not trust the resolver — it still requires each record's tree to equal the checked-out tree and additionally rejects a stamp naming the candidate's own commit, a stamp disagreeing with the record's sha, and a bundle whose reused jobs name more than one source run; the clean-proof lines, step argv and the Site E2E candidateSha are all bound to the record's own sha. A resolver job failure fails the aggregate rather than degrading silently. The decision record is docs/adr/ADR-0156-tree-sha-evidence-reuse.md, including its P7 four-question review and the rejected alternatives (message-scoped reuse, per-lane resolution, a checked-in tree→run index).
• CI — every remaining lane is bounded, and a red Site E2E run now stages its evidence (#1402, #1409): dependency-review, codeql#analyze and both published-consumers jobs had no ceiling, so a stalled extractor or a hung published-starter qualification could hold a runner for the platform default (six hours for CodeQL); each now carries one sized to its real work. The fresh-clone lane used to run its Site E2E suite before staging the report, so a failing suite threw past the staging step and its bundle carried logs but no report — the failing test's name was then unrecoverable without a live repro, which is exactly the class the alpha.2 closeout could not diagnose. A red run is now recorded before it is raised: the raw report plus sidecar are staged whenever the runner managed to write them, ran: false when it did not, and the failure is re-raised after result.json so the exit status and the validator's verdict are unchanged. A red run still cannot smuggle a pass — the aggregate fails closed on a red or absent sidecar and rebinds the report bytes to the candidate commit.
• Router — the accepted config surface grows to eleven keys, and head becomes structured (#1411 follow-up): dirs: { routes, islands, components } moves the source roots (the tokens / app-shell / head conventions follow the three roots' shared base, so a full src/* move carries them); packageIslands additionally derives ssr.noExternal so a listed package's islands are bundled, with no public ssr.noExternal key; head.scripts and head.stylesheets join the config file, so a config entry and an inline inject.scripts entry emit the same bytes through one serializer; speculation, viewTransition, build.manifestBudget and i18n validate as data. Raw markup stays out on purpose — inject is deliberately not a config key. The new app/head.tsx convention carries the structural head a URL list cannot express (preloads, icons, feed links, inline CSS); it is compiled into the app's module graph, not read as text, because loadConfigFromFile refuses a ?inline CSS import. Entries are validated at the boundary — unsafe attribute names, non-string values, javascript: URLs, @import and unclosed <style> all fail the build instead of being silently dropped. An accepted head key that no code reads fails closed too. Two latent build-context bugs surfaced and are fixed: config-file i18n options never reached the integration (the context derived them before the config resolved, so the SSG emitted a single-locale site with locales="[]"), and the head compile now keeps bare specifiers external instead of bundling package sources into a data module.
• Retired — inline openElement({ html }) (#1411 follow-up): the flat html option is gone and fails closed with CONFIG_RENAMED_KEY naming head. This is the one authored-spelling change in this train: six fixtures, three packed-consumer harnesses, apps/saas and the Site itself are migrated in the same train. middleware.use deliberately stays inline-only (the config file's middleware block carries corsOrigin), and /guide/configuration now says the two are either/or rather than mixing them, which remains a hard error.
• Element — one error dialect, mechanically enforced (#1386): the package used to raise failures through four conventions (OEC diagnostics, OpenElementError with codes, ~40 bare Errors carrying [compiled-*] prefixes, and three dedicated exception classes), so a consumer could not catch a failure by code without importing every class and pattern-matching the rest. OpenElementError and its per-surface code catalogue now live in internal/protocol/errors.ts — import-free, so the ADR-0148 semantic core and the Part Program protocol raise framework errors without gaining host state — and every throw in packages/element/src speaks it. The released values (OPEN_ELEMENT_COMPILED_CLAIM_MISMATCH, OPEN_ELEMENT_COMPILED_PROGRAM_INVALID, OE_PROGRAM_MISSING, SSR_DOM_ACCESS_UNSUPPORTED) are pinned so a later cleanup cannot rename them; new codes use the OE_ prefix. The dialect is proved by an AST walk over every module under src/ that fails on any bare throw or leftover TypeError, plus a duplicate/off-namespace catalogue check.
• Element — the emitted module is type-checked at build time (#1386): the compiler emits the compiled module as TypeScript text and the bundler lowers it, but nothing checked that text, so a compiler change could emit a program that fails deno check or a consumer's tsgo run and the first evidence was a consumer's build. typeCheckEmittedModule() compiles the emitted text as a real program and returns a checker's diagnostics for it; the Vite adapter gains an opt-in typeCheckEmitted build gate (off by default, because it runs a program per emitted module — a dev-server transform must not pay that). It is a pure function of its inputs and only reports diagnostics belonging to the emitted files, so a caller's incomplete resolution map cannot masquerade as an emitted-module defect.
• Element — claim-vs-fresh is decided from content, not a child count (#1381): the kernel chose claim by root.childNodes.length > 0, so a compiled light-root element carrying a whitespace-only text node — the shape an HTML formatter or a parsed file produces — threw PartProgramClaimError on upgrade for a node that is invisible in the rendered page. Formatting whitespace is now normalized before a fresh mount. The fail-closed direction is pinned as hard as the fix, because widening it would be the correctness regression: a real element, a serializer anchor comment, visible authored text and whitespace preceding real content all still throw the structured, catchable mismatch, and a serialized light host still claims in place so server node identity survives.
• Router — navigation state has one owner (#1385): the client router's three pieces of shared mutable state (the monotonic latest-wins ticket, the dedup key of the last browser-landed URL, and the one-shot marker of the router's own guard-veto restore) lived across four functions. They now live in internal/router/navigation-state.ts behind three questions — issue/owns, isDuplicateLanding/recordLanding, armRestore/consumeRestore — so cancelled pending execution is a side effect of owning intent, never of attempting a navigation; a guard veto, a stale ticket, an aborted traversal or a disposal all leave the current route's in-flight render alone. The machine is DOM-free, so its transitions are pinned without a browser, and the four interaction cases that motivated the extraction are pinned in the router suite.
• Router — browser-shaped action POSTs must present an Origin (#1382): the generated action POST floor let a multipart/form-data or application/x-www-form-urlencoded body through when it carried neither Origin nor Fetch Metadata, because that allowance exists for non-browser callers. A browser-shaped body can present the same absent headers, so such a body is now rejected with the same 403 and RFC 9457 problem document when it also carries browser navigation evidence (Upgrade-Insecure-Requests: 1, or the text/html Accept a form navigation sends). A scripted client sends neither marker and is unaffected; Origin: null is untouched, since null is an origin the browser did send. The residual window is written down rather than implied: Origin: null without Fetch Metadata, text/plain form bodies, custom API routes and ambient-auth apps, and the OPEN_ELEMENT_DISABLE_CSRF=1 opt-out, which disables this rule too.
• Docs — how i18n actually works (#1383): /guide/i18n (plus its zh twin) states the boundary that had no page: catalogs are application code. It covers the two file conventions — declarative locales resolved through PagePropsContext.locale, and a loader's file-suffix scheme over the content tree — and the three rules that keep them honest (one link helper, never infer a locale from a two-letter segment, the default locale stays unprefixed). For catalogs it names the boring standards and when each fits (ICU MessageFormat for tooling compatibility; paraglide once volume grows), both build-time because a static-first framework prerenders its pages. What stays out is explicit: no message format, no Accept-Language negotiation, no content translation inside the framework.
• Repo — the docs figures and the release bookkeeping follow the tree: www#check:doc-figures pins comparison.md/.zh.md against a fresh measurement of www/dist, so this train re-baselines them (one page per locale from /guide/i18n, plus the runtime growth from the Element error dialect: html 64→66, sitemap locs 62→64, manifests 64→66, per-locale pages 31→32, fragments 62→64, manifest entries 304→314, rail pages 54→56, code-block pages 42→44, island-open-cinematic-scroll 81,984→85,765 raw and 25,215→26,855 gzip, / payload 217,362→221,626 B). The six version points move through deno task version-bump rather than by hand, the state machine admits v1.0.0-alpha.4, and the README, the Site version source and the registry-truth constants follow.
• Tracked, not fixed: typeCheckEmitted is implemented and tested but not yet enabled in the Router build CLI (packages/router/src/cli/{build-ssg,build-client}.ts), because that tree belonged to another lane in this run; enabling it there is a follow-up. The #1387 portable-host migration (moving the build-time Deno.* calls in packages/router/src/{vite,cli} and packages/create/src to node:*, and retiring the existsSync seam) is still a deferred roadmap item, not part of this train — the build-time Deno-host requirement documented in the READMEs therefore still stands. The www/tools/generate-api-reference.ts config-type anchor still points at OpenElementOptions (which no longer carries the config-file keys); the API reference covers the config surface through /guide/configuration and the interface snapshot, and re-pointing the anchor at OpenElementUserConfig is a follow-up.

1.0.0-alpha.3

Door-handle train: the surfaces a consumer actually touches — the config file, the packed npm facade, the error vocabulary, the documentation pipeline, and the shipped bundle — are converged and gated. The source line is 1.0.0-alpha.3; npm publication is a separate, gated step, so docs/release/release-state.json keeps its registry block at the alpha.2 truth until the post-publish sync. No migration steps: the config file is an optional overlay and nothing here is a breaking change, so an alpha.2 consumer upgrades with no action.

• Element, Router — framework options get one home (#1411): openelement.config.ts is optional and near-empty; with an empty object every option comes from a file convention — design tokens from app/styles/tokens.css, the application shell from app/islands/app-shell.tsx, site title from package.json. A non-empty config file next to inline openElement(...) options is a hard error ("framework options have two homes"), and an unknown or wrong-typed key throws with the accepted-key list; there is no silent merging. The starter template's vite.config.ts is now plugins: [...openElement()] with zero CSS strings, <app-shell> is registered by convention and stays deletable, favicon and og tags reach the built document, and the first deno task dev run no longer prints the production CORS advisory. The documented consumer scaffold command is the short all-permissions form the owner ruled on 2026-09-21 (the full command, with its concrete dist-tag, lives in the create README and the getting-started guide), admitted as an exact path-and-line exemption in the check-no-allow-all tripwire while every first-party invocation stays scoped.
• Release — packed npm facade (#1412): per-package keywords, engines (node >=24; deno >=2.9 on the two Deno-driven toolchains) and sideEffects (false for element/router/ui, ['./src/cli.js'] for create) are written from one source and re-read from the real tarball bytes by the new pack-surface:check gate, which also fails closed on a repository path or ADR citation in shipped text, on an export subpath the shipped README never names, and on a module-scope global write in a package declared sideEffects: false. Every Element facade subpath is documented in the package README — the eight that existed when this landed, plus client-only below; 54 shipped files lost their repository-internal references; the create README states the current "do not run the bin under npx" limitation (the Node entry is tracked by #1387).
• Element, Router — error experience (#1413): the compiler hands the build a structured diagnostic ({id, loc, frame, message, diagnostics}) instead of a pre-joined display string, so an overlay can underline the authored line and a consumer can read the location without parsing the message apart. The three user-facing runtime failures name the compiled module, the tag and the authored this.<property> instead of an index the author cannot map back; router authoring throws carry stable codes with phase and severity plus the remediation sentence they were missing; the start CLI prints one actionable line built from the message and cause chain, with --debug expanding the full stack. /errors is a derived artifact — every fail(…, 'OEC9xxx', …) literal plus the element error protocol and router code maps — and 38 codes render per locale from the source that raises them; a hand-written list fails check:errors.
• Docs — pipeline before content (#1414): five pipeline pieces landed first, so the facts are rendered rather than retyped. The API reference renders each export's declared signature and its option-bag members; an undocumented public export is red, and all 51 empty summaries are now documented at their declarations; the config option table is generated from the application options type; /errors is generated from the definitions; and the install command has one owner (packages/create/src/install-command.ts) which the site interpolates and starter-smoke asserts against the packed CLI's own output. Content followed the pipeline: the routing-and-data Metadata and Data boundary sections are written out, delegatesFocus/formAssociated/converter are documented on core-concepts, and getting-started's Build section is user-facing.
• Repo — legacy cleanup (#1415): deno task version-bump <version> owns all six version points — 4× package deno.json, CREATE_VERSION, and the four fixture deno.lock files — with a dry run by default and --write applying the edits, regenerating the locks, and then failing closed unless all six agree; hand-editing them burned two alpha.2 release rounds. A workflow_run companion re-runs the FAILED jobs of AutoFlow CI exactly once (attempt 1 only), because webkit fails deterministically per runner and only a fresh machine can change the outcome (#1409). A workspace-alias-hijack guard refuses to alias @openelement/* onto a checkout's sources when an app is scaffolded inside a framework clone — a build that would otherwise report success while resolving a different framework version (#1371 family). The evidence/ directory (0.43.3 tarballs) is gone and its ignore rule is documented as staying; the per-fixture Nitro rules collapsed to **/.output*|.wrangler|.nitro; and the ten fixture and e2e directories that had no README have one.
• Element — bundle and update cost (#1416): @openelement/element/client-only is a fresh-DOM entry without the existing-DOM claim executor, and the Router selects it only when no island on a page can hydrate server DOM (an island that says nothing keeps the full entry, because guessing the other way breaks hydration). Measured on the JFB keyed-table harness: 77,557 B → 68,630 B, −8,927 B (−11.5%), with the claim diagnostic strings at grep count 0 in the final bundle. Keyed updates also skip the slot walk when an entry's item reference is unchanged, turn the created-entry lookup into a Set, and resolve the fallback insertion reference on demand: instrumented on 05_swap1k, updateItemValues calls drop from 1000 to 0 per swap and the synchronous segment from 3.0 ms to 2.7 ms median (−10%, 480 samples).
• Honest measurement — the swap1k target was not met: swap1k did not reach single digits, and this section records why rather than restating the goal. On the same runs the afterframe protocol floor (one requestAnimationFrame plus one MessageChannel task with no DOM work at all) measured ~13–15 ms — 15.2 ms baseline against 14.5 ms with the change, and ~12.8 ms median in the separately documented floor measurement — while moveEntries' 997 real insertBefore calls cost ~2.3–2.7 ms. moveEntries is unchanged byte for byte (Svelte's algorithm, ruled out of scope for this train), so the reported total contains the protocol floor plus that walk. The reactive boundary this exposes is documented in both locales on core-concepts: mutating an item in place is outside the contract — the same reference comparison a signal() holding an object draws — and the supported shape is a new item or a new array.
• Tracked, not fixed: the packed-starter browser matrix runs behind OPEN_ELEMENT_SKIP_BROWSER_MATRIX=1 (#1425). It fails 3/3 browsers with a waitForFunction timeout in the consumer harness while the identical probe against a manually scaffolded packed starter passes all three browsers. The rest of packed qualification stays required; the guard is to be deleted when the investigation closes, not carried forward.

1.0.0-alpha.2

Constitutional-enforcement train: the alpha.1 review's fail-closed and single-source findings are fixed, and the compiled when grammar grows to the full admitted set. Published to npm as 1.0.0-alpha.2 under the @alpha dist-tag.

• Element — conditional grammar (#1372): when regions accept >, >=, <, <= against a finite numeric literal; strict ===/!== against number, string, or boolean literals (no cross-type coercion — 1 never matches "1"); and bare this.<property> truthiness with negation. Ternaries were already two-branch regions and now compose with the widened operators. All three executors evaluate conditions through one shared module (condition-holds.ts); the operator/literal space is closed by one predicate shared by both wire validators; the convergence guard pins the new invariant. The showcase island re-baseline is recorded honestly (78,176 → 79,199 raw bytes, +1.31%).
• Element — security (#1373): meta.tags attribute names are validated against the canonical isSafeAttributeName and fail the render closed (UNSAFE_META_ATTR_NAME); CMS-fed metadata can no longer smuggle attribute injection through name grammar.
• Element — correctness (#1374, #1375): the each-region item-key derivation is one shared typed function (number 1 and string "1" keep distinct identity from SSR through claim); runtime update-phase errors route into the kernel error boundary instead of escaping into the signal writer's stack, with retain-and-retry isolation semantics.
• Tooling (#1376, #1377, #1378): deno task verify is gate:ci-equivalent and pinned by a contract test; local agent-workspace directories are ignored and git hooks are a documented setup prerequisite; the Router npm README states the build-time Deno-host requirement for ./vite and ./cli/*.
• Site (#1379, #1380): the roadmap publish-state derives from release-state.json (P6); the README comparison's Fresh row states the niche argument with dated stall facts.
• CI hygiene (#1400, interim #1402): doc-figures chunk-raw rows carry the ±1% cross-runner tolerance; site-e2e retries a single flake per test. Investigations open: doc-figures determinism (#1401), job timeouts and flake management (#1402), content-dates pre-push hook (#1405).

1.0.0-alpha.1

New repository baseline for Element and Router, published to npm as 1.0.0-alpha.1 under the @alpha dist-tag (npm latest stays on the stable 0.43 line). This is not an upgrade of the 0.x lines and no migration path from 0.x is offered: new projects start from @openelement/create.

• Packages: exactly four public packages — @openelement/element, @openelement/router, @openelement/create, and the experimental @openelement/ui. @openelement/app and @openelement/adapter-vite are retired; their responsibilities now live in Element and Router tooling subpaths.

• Element: one mandatory compiler lowers supported TSX into a Part Program; OpenElement subclasses are authored with the @element decorator and @property state. Server serialization, fresh DOM, and existing-DOM claim all consume the same compiled artifact. Element installs without Router.

• Router: Route Mode (explicit route records) and Framework Mode (file routes, loaders/actions/forms, SSR/SSG, Vite integration, Nitro mount) ship from @openelement/router, ./vite, ./nitro-mount, and ./cli/*. Route Mode installs without Element. The unimplemented client-side RouteConfig.loader/action fields and the public SpaLoader*/SpaAction* types are removed: the client router never ran them, so the public API no longer promises it. Data fetching stays on the route modules' server loader/action.

• Breaks from 0.x: package names and import paths changed with the new baseline; @openelement/element/sanitize is gone and trustedHtml is the trust boundary; raw head fragments are passed through verbatim with only fail-closed invariants (<script> and executable <style> rejected); served Content-Type values derive from the maintained mime package; runtimes without the Web Standard URLPattern fail fast; the generated standalone dist/server/serve.mjs is replaced by the portable fetch(Request) -> Response entry plus the start CLI; the ADR-0120 hard rule "pages with actions cannot be prerendered" is repealed — a page may be hybrid: prerendered static GET plus a request-time action POST, with dist/server/ emitted whenever any route has an action (ADR-0120 amendment, 2026-09-16).

• Install (Alpha):

  deno run --allow-read --allow-write --allow-env --allow-net --deny-ffi --no-prompt --minimum-dependency-age 0 npm:@openelement/create@alpha my-app
  

  @alpha tracks the 1.0 prerelease line; a versionless install resolves the stable 0.43 line. Deno 2.9+ is required.

• Known limitations: Alpha APIs may still change. @openelement/ui is experimental and outside the stable API promise. Hosted/deployed SaaS qualification, production SMTP, and real scan-engine qualification are external pending.

Review-round corrections (post-baseline, this branch)

• Correctness: the reading-page heading-id allocator now probes for the first free suffix against every id already in the document (a page with an element id="foo-2" plus two Foo headings no longer emits a duplicate DOM id); RSS pubDate validation is a UTC calendar round-trip, so impossible dates (2026-02-29, 2026-04-31) fail closed instead of being silently normalized by Date.
• Public API (Router): head.structuredData entries and values are typed as a recursive JsonValue (StructuredDataEntry = { readonly [key: string]: JsonValue }) instead of Record<string, unknown>; JsonValue is exported. Runtime validation is unchanged.
• Public API (UI): openPropsRootSheet and the toRootCss transform are removed. The remaining openPropsTokenSheet token block selects :root, :host, so one generated sheet serves document and shadow adoption; the structural fallback stays :host-only.
• Build: the production Site build is hermetic — source dates come from the committed www/lib/content-dates.json manifest and the retired-URL baseline from www/tools/site-baseline-routes.json, so site:build needs no network, no remote and no .git. Git-based refresh/verify tasks (content-dates:*, retired-url:check --refresh) run in CI, not in the build.
• Release: @openelement/ui now ships THIRD_PARTY_NOTICES.md inside its tarball (open-props MIT text), asserted by the packed-artifact gate.
• Tokens: the open-props adapter lives at packages/ui/tools/generate-ui-tokens.ts (task deno task --cwd packages/ui generate:ui-tokens); build-artifact emitters are named emit-* and no longer fake --check tasks.

0.44.0-beta.1

First public v0.44 prerelease (dist-tag beta; npm latest stays on the stable 0.43 line). The TSX-to-Part Program compiler and page-route SSR bound to the compiled program. The 0.41.0-era npm beta.1–beta.3 artifacts remain withdrawn partial publishes, unrelated to this line.

0.43.3 / 0.43.2 / 0.43.1 / 0.43.0

Stable maintenance line (npm latest). Compatible bug, security, runtime, documentation and release-truth patches under ADR-0140 — no 0.44 feature train.

0.42.0

WC light fullstack, stable. The stable cut of the 0.42 alpha line: the request-time Application Loop (ADR-0120) frozen on top of the 0.41.x static freeze (ADR-0119), with freeze scope and non-goals in ADR-0122.

• Frozen scope (ADR-0122 §1–§4): the loop contract (loader/action signatures, fail()/redirect() algebra + HTTP encodings, PRG revalidation, no-JS baseline), the action protocol (x-openelement-action, morph client contract, channel symmetry), the fail-closed CSRF same-origin default, and first-mile start semantics. Breaking changes to these require an amendment ADR.
• Breaking changes since 0.41.x: IslandOptions.strategy renamed to hydrate with no alias (ADR-0127); shape-1 SSR markup gains one fallback-tag wrapper element (ADR-0128); head-injection tightening — <base> / <meta http-equiv> / raw <script> rejected in headExtras; unfrozen alpha exports removed (i18nStaticPaths, switchLocale, AppIslandOptions, OPEN_PROPS_TOKEN_CSS, the OpenElementRouteNode re-export); pure-static builds no longer emit dist/server/; minimum Deno version is 2.8.
• Not frozen / not claimed: session/flash, cache/ISR (revalidate stays inert forward-compat data), streaming SSR, performance SLOs, third-party WC SSR corpus, production runtime recovery.

0.42.0-alpha.17

Registration-decoupling train (ADR-0128): route modules whose default export is definePage(...) register the page class under the route-path-derived fallback tag; export const tagName on a definePage route only names a content element. Shape-1 pages gain the fallback-tag page element as an outer DSD wrapper around the content element; user CSS targeting the old root tag must target the new fallback tag or the content element. Plain element routes keep their tagName export as the registration tag.

0.42.0-alpha.16

Starter-first remediation train.

• Island runtime: hydrate: 'only' islands bind events and signals instead of rendering inert; function-mode defineIsland islands re-render on signal change.
• Routing/SSR: notFound() from a page render propagates to a real 404; unmatched request-time paths render the styled 404 page with Cache-Control: no-store; successful GET pages relax to private, no-cache while POST responses keep no-store.
• element runtime: <style>/<script> text children serialize as raw text; data: URIs are allowed on img src only; keyed For semantics locked and its teardown leak fixed.
• Breaking (unfrozen alpha surface): head-injection sanitization tightened (<base> and <meta http-equiv=...> stripped from headExtras/head fragments); IslandOptions.strategy renamed to hydrate with no alias (ADR-0127).
• CSRF and enhanced forms: cross-site POSTs rejected while same-origin native form posts pass; enhanced forms morph correctly into slotted light-DOM pages.
• Starter surface: working blog routes, pinned dev vite version, styled 404 route.

0.42.0-alpha.15

Backlog-zero train: the element runtime gains a built-in allow-list HTML sanitizer sanitizeHtml (later removed in the 1.0 baseline in favor of the trustedHtml boundary) and keyed <For each key> reconciliation via an optional key prop (ADR-0124), with displaced entries disposed on duplicate keys. URL safety decoding tightened and _blank links get rel neutralized. The SPA router gains a URLPattern fallback for Firefox, and framework throws converge on OpenElementError.

0.42.0-alpha.14

Simplification and consumer-packaging train: packed-package consumers of JSR dependencies unblocked; client runtimes are bundled via virtual modules; jsonc parsing delegated to @std/jsonc; no-skip version continuity in release tooling.

0.42.0-alpha.13

Standards-as-seams train (ADR-0123).

• Morph robustness: the enhance/island client is a real tested module; focus, scroll and form-control state survive enhanced updates; nested DSD templates instantiate recursively; open:ready fires for every strategy bucket.
• Route standards: all route matching is WHATWG URLPattern — the SPA client router and the generated server matcher share one semantics; renderIntent.mode is 'static' | 'dynamic' with the 'auto' alias removed.
• Server seams: a WinterCG-shaped fetch middleware contract runs identically in dev, start, fixtures and Nitro; cli/preview merged into cli/start --mode=preview.
• Protocol: fetch-channel error responses are RFC 9457 application/problem+json, including the CSRF 403.
• Components: open-input becomes a real native-form citizen; open-dropdown moves to the Popover API with CSS anchor positioning.
• Site search: full-text, ranked, bilingual via Pagefind.

0.42.0-alpha.12

Maintenance train; no new product surface beyond the fixes.

• Correctness: a guard-vetoed redirect() from a post-action loader re-run no longer wipes page data; the SPA client router matches Hono-style :param{.+} catch-all patterns; malformed percent-encoded URLs answer 400 instead of hanging.
• CSRF floor proven: the generated action POST same-origin floor has real deny/allow coverage, with OPEN_ELEMENT_DISABLE_CSRF=1 as the documented opt-out.
• Honest claims: ISR labelled forward-compat/inert wherever it appears; the security guide documents the built-in CSRF floor; the starter headerNav config renders.

0.42.0-alpha.11

Maintenance train.

• Security: validateSafeUrl tab/newline bypass closed (module-script data: XSS); desktop examples bind loopback only; the Mastodon example no longer caches API errors as data.
• Correctness: a guard-vetoed redirect during navigation keeps the current page's loader data; useLoaderData<T>() types T | undefined; theme broadcast no longer clobbers host-owned data-theme; disconnect→reconnect no longer resets non-reflected prop state; JSX callback ref is consumed; open-button anchor-mode disabled sync works both ways; open-input / open-badge observe dynamic attribute changes.
• Honesty over surface: the homepage flagship example is real compilable API; dead options, config keys and misleading types removed across element/app/adapter-vite.

0.42.0-alpha.10

Cleanup and hardening train; no new product surface.

• Runtime correctness: element update() routes re-render errors to onRenderError; SPA loader failures take the __openElementError channel and SPA loader/action honor redirect()/notFound() with real navigation; UI double-escaping removed, open-tabs accessibility rewrite with instance-unique ARIA ids, open-dialog SSR open sync; generated-data writes fail closed in build mode; getStaticPaths errors honor dynamicRouteFailure: 'fail'; the start CLI static server shares the test fixture and unwraps request-time responses.
• Breaking (unfrozen alpha surface): @openelement/app drops i18nStaticPaths/switchLocale/AppIslandOptions and the OpenElementRouteNode re-export; @openelement/ui/open-props-tokens no longer exports OPEN_PROPS_TOKEN_CSS; the generated ui manifest corrects open-tabs slots to ['tab','panel']. @openelement/element gains @experimental exports: ISR cache types and the third-party client runtime hydrateOpenElement/disposeOpenElement.

0.42.0-alpha.9

Cleanup train; no new product surface. <open-button> binds its click handler so shadow-DOM submit events reach the outer form; request-scoped context is passed explicitly through render/hydrate entry points.

0.42.0-alpha.8

Incomplete release (npm-unpublished). Tagged as a mechanical version bump, but the packages were never published; superseded immediately by 0.42.0-alpha.9.

0.42.0-alpha.7

Patch release; all five packages published to npm under the alpha dist-tag (historical five-package line).

0.42.0-alpha.6

Second independent review of the application loop.

• Morph client correctness: an explicit <form action> wins over the page URL on enhanced submits; the popstate guard survives reloads and bfcache restores; morphed-in islands show the server render; morph matching is an ordered walk with exact deletion and relocation; nested DSD compares normalized on both sides; forms inside late-hydrating islands get the enhancement listener; a cancelable open:action-error hook precedes the network-failure reload.
• Detection: hasEnhancedForms follows relative imports.
• Protocol tail: malformed form bodies answer 400 on both channels; the redirect duck type honors the 3xx whitelist; 405 responses carry no-store/Vary.

0.42.0-alpha.5

First hardening pass on the 0.42 line (ADR-0121).

• Root cause: the alpha.3 morph enhancement never fired because submit is not composed and page content lives inside page-element DSD shadow roots. The client is rewritten around shadow-root submit interception and shadow-content morphing; island-survival claims are mechanically true.
• Protocol (ADR-0121): named-action dispatch is own-key gated; one x-openelement-action header (true = ActionResult JSON, enhance = HTML morph) with Vary; an action returning a Response is a contract violation; the default PRG strips the ?/name marker; every 3xx coerces to 303 on POST and redirect() validates its status; fetch callers always receive ActionResult JSON; request-time responses carry Cache-Control: no-store; action POSTs get a 10 MB body limit; non-GET/POST methods answer 405.
• Morph continuity: form-scoped data-open-region targeting with navigation fallback, id-keyed + lookahead identity matching, popstate reload, cancelable open:action-failure, submitter name/value preserved, 500 / cross-origin responses navigate instead of morphing, double-submit guard, fragment preservation, <details>/media state protection.
• Also fixed: dev SSR crash on every route; [...path] request-time routes; zero-island apps with enhanced forms; the starter's /contact route now builds and is POST-smoked.

0.42.0-alpha.4

Hardening and recipes (ADR-0120).

• Validation recipes verified in tests: zod (/register) and valibot (/subscribe) run inside fixture actions with 422/303 asserted in three engines. better-auth and Drizzle recipes are published as doc-level, honestly marked unverified.
• The create starter gains a request-time /contact route exercising the full loop (rendering: 'dynamic' + action + data-open-enhance).
• Fix from alpha.3: the morph no longer replaces an island whose light DOM carries whitespace-only text around the DSD template.

0.42.0-alpha.3

Revalidation continuity (ADR-0120): enhanced forms (data-open-enhance) morph the returned document into place instead of reloading — submission returns the same HTML the no-JS path renders (303/422), the client morphs it, and history.pushState follows the PRG target. A hydrated island whose light-DOM surface is unchanged keeps its shadow state; data-open-preserve exempts any subtree; the island client script is never re-executed by a morph. A data-open-region container limits the morph to the matching region.

0.42.0-alpha.2

The form/action loop (ADR-0120): plain HTML forms work without JavaScript on rendering: 'dynamic' routes. Actions run before loaders; fail(4xx, data) returns take the 422 re-render channel with the echo; successful mutations answer 303 (PRG); redirects thrown from actions coerce to 303; POST without an action is a defined 404. Named actions dispatch via formaction='?/name'; unknown names are a defined 404. Fetch callers receive the ActionResult discriminated union.

0.42.0-alpha.1

First alpha of the 0.42 line (ADR-0120): request-time rendering gains semantics. rendering: 'dynamic' routes skip prerendering and are served per request by the generated dist/server/index.js, with dist/server/server-manifest.json recording the partition. Hard rule: pages with actions cannot be prerendered — a route module exporting an action without mode: 'dynamic' fails the build (repealed by the ADR-0120 amendment of 2026-09-16: hybrid static GET + request-time POST is now allowed). Pure-static projects emit no new artifacts.

0.41.2 / 0.41.1

Patch releases: release-tooling hardening and hygiene only — no public API, topology or runtime-default changes.

0.41.0

Stable five-package release (ADR-0119; historical line): the interface freeze covers defineElement, definePage, buildApp, the package graph, the supported subpaths and the static/SPA semantics of defineApp; request-time data, forms, sessions and cache stay explicitly unfrozen until 0.42/0.44.

• Breaking: adapter-vite internal subpaths pruned at the freeze (app-vite, build-context, head-injection, i18n-plugin, plugin, generated-data-resolver, plugin-mdx, route-manifest, cli/build-client, cli/build-ssg) — use the root, nitro-mount, cli/build and sitemap instead.
• ui control geometry is squared (--btn-radius, --badge-radius, --ui-control-radius: --radius-round → --radius-1, 6px) — visual breaking change; update screenshots and custom control CSS.
• The release verifier now supports stable x.y.z versions, and the version guards stay honest on a stable current line.

0.41.0-alpha.19

Cleanup sweep; no new product surface. Fixes the reflect removal suppression (Boolean default: true desync), the popstate redirect-then-block URL fork, and For drift-token separator collisions. Breaking type-surface changes: the element root switches to explicit type export lists, SafeHtml/UnsafeHtml/StyleSheetRule leave the root, and the internal open-element-render/open-element-hydration subpaths are pruned from element exports.

0.41.0-alpha.18

Fixes the reflect: true static-prop write loop and SSR attribute overwrite; resolves the root-level <Show>/<For> CSR edge; unifies prop attribute casing; makes For branch tokens content-sensitive; fixes client-runtime double hydration; runs router guards on history traversal; honors prefers-color-scheme in theme-init. Dynamic-route render failures now fail the build (opt-out 'warn'). Breaking removals of dead exports, fields and scripts.

0.41.0-alpha.17

Covers hydration and binding behavior in a real browser: signal text patching, event hydration, SSR/hydration mismatch fallback and form submission through shadow boundaries. A failing route render produces a defined 500 result with RenderError diagnostics. Breaking surface removals: element root build utilities (migrate to @openelement/element/build-utils), app root RouteConfig/RouterInstance/RouterMode types, adapter-vite ExternalManifest type and SsgPageOutput.hydrationHints.

0.41.0-alpha.16

Fixes unknown dynamic-route params serving 200: a notFound() thrown from a page element's render propagates through the DSD render chain as protocol control flow so the request-time server entry answers 404. SSR and hydration event markers align for custom-element hosts and Show/For branches; hydration validates marker counts and branch tokens and falls back to client re-render on mismatch. Static-props observedAttributes merge at class-definition time. Windows drive-letter island paths normalized.

0.41.0-alpha.14

Recovers the release line with an exact-version starter and verified published consumers; all five packages published under the alpha dist-tag (historical five-package line) with post-publish Deno, Node ESM, Nitro and third-party Web Component smoke coverage.

0.41.0-alpha.13

Publication failed; changes shipped in alpha.14.

• Removes the alpha-only defineLayout alias; use defineElement with the same arguments for layout elements.
• Restores declared static-prop defaults when reflected attributes are removed.
• Hardens SSR prop injection, custom-element hydration, params parsing, nested SSR depth, and adopted stylesheet composition.
• Stabilizes SPA action errors, caches same-route GET requests, bounds render data contexts, and compiles client routes into a declaration-ordered trie.
• Moves UI tokens to a CSS source of truth with generated-output drift checks; Create template generation becomes asynchronous, deterministic, and bound to the same-version release invariant.

0.41.0-alpha.12

Audit-remediation foundation release.

• Fixes core runtime issues, including the signal-context infinite loop and ErrorBoundary retry.
• Hardens SSG/build: command-injection closure, dynamic-route encoding, and pure-Node process.cwd() compatibility.

0.41.0-alpha.11

• Restores frozen-install and changed-path workflow truth.
• Fixes query decoding, SPA page-host data, dialog inert restoration, and theme propagation.
• Consolidates the Element/Adapter protocol seam and removes verified dead DSD, CEM, route-scanner, and UI escape code.
• Repairs clean Nitro Workers builds, semantic visual smoke, and package artifact allowlists.