djust 1.2.0rc8 Release Notes

This is a pre-release. djust 1.2.0 has shipped since.

Part of djust 1.2 — read the 1.2 release notes.

Added

djust.T018 system check — undefined template variable detection: manage.py check (and djust_check) now warns when a LiveView template references a variable that resolves nowhere and silently renders as an empty string. Compares template variable references against the view's statically-determinable context (public class attributes, self.x = ... assignments, literal get_context_data() dict keys, template-declared loop vars, and framework/Django-injected names), reusing the same extraction manage.py djust_typecheck has shipped since v0.5.1. Covers both template_name (file) and inline template = "..." views. Advisory (WARNING, not ERROR); suppress project-wide with DJUST_CONFIG = {'suppress_checks': ['T018']} or per-name with the existing {# djust_typecheck: noqa name #} template comment. Known v1 limitation: {% extends %} templates are skipped entirely (block-override context is inheritance-scoped and not statically checkable here). Also fixes a pre-existing djust_typecheck false positive where a dotted {% if x.y %} expression incorrectly reported the attribute tail (y) as its own undefined top-level reference.
LiveComponentTestClient — test a LiveComponent in isolation: the component counterpart to LiveViewTestClient. A component has no view, no template_name dispatch and no URL, so it could not be mounted through the view client; testing one previously meant wrapping it in a throwaway parent LiveView. LiveComponentTestClient(MyComponent).mount(**props) returns the client, send_event(name, **params) calls the component's own handler method, get_state() returns its public (non-underscore) attributes, and render() delegates to LiveComponent.render. A missing handler raises NoHandlerFoundError rather than silently no-opping, matching LiveViewTestClient.send_event (#2823) — a renamed handler should fail loudly, not leave a test passing while the feature is broken. Like the view client it calls handlers directly rather than routing through the WebSocket consumer, so @event_handler metadata is not enforced; use LiveViewTestClient when the routing contract itself is under test.
dj-mouseenter / dj-mouseleave are now real directives (#2869). The attributes were previously stamp-listed by {% live_render %} — the framework asserted they existed — but the client never bound them, so a developer writing <div dj-mouseenter="highlight"> got a stamped attribute, no listener, and no warning. They are now wired as first-class event directives: mouseenter/mouseleave do not bubble, so the client attaches listeners DIRECTLY to the declaring element through the existing scoped-listener machinery instead of the delegated root-level shape dj-click uses. Nesting semantics are the platform's own: moving the pointer from an element into one of its children fires neither the element's dj-mouseleave nor a second dj-mouseenter (entering a child is not leaving the parent — the reason these event types exist over mouseover/mouseout). Bind-pass safety follows the house rules: the #2845 skip-on-unchanged-value / evict-and-rebuild rule prevents double-attach across morphs, the #2832 sweep detaches listeners when the template stops declaring the attribute, and the handler closure re-reads the attribute at fire time (#2858) so a value change under a surviving element is honoured. dj-debounce / dj-throttle, dj-confirm, data-* params, inline handler args, and {% live_render %} embedded routing (view_id from the stamp) all work as on sibling directives. Documented in docs/website/core-concepts/events.md ("Mouse Enter / Leave") and the template cheatsheet quick-reference card.
Theming: the domain model is now importable, and the request-scoped helpers exist: from djust.theming import ThemePack previously raised ImportError. The types (ThemePack, DesignSystem, SurfaceTreatment, TypographyStyle, LayoutStyle, SurfaceStyle, IconStyle, AnimationStyle, InteractionStyle, PatternStyle, IllustrationStyle) were defined in theming/_types.py — a private module holding public types, with its own __all__ advertising them — but were never re-exported from djust.theming, and register_theme_pack accepts a ThemePack, so the documented model was unusable end to end. The theming documentation also instructs from djust.theming import get_active_pack, set_active_pack, get_active_mode, set_active_mode, reset_to_defaults, get_theme_css_url, none of which existed. Those six are now implemented in theming/api.py as thin, request-scoped wrappers over ThemeManager — theme state is per-request (session-backed, with a cookie fallback for anonymous users), so every one takes request, and get_active_mode returns the resolved mode so a caller on 'system' learns which one they are actually seeing. Supporting additions: ThemeManager.set_pack() (the missing sibling of set_theme/set_preset/set_mode) and ThemeManager.reset(), which removes the session entry rather than writing defaults into it so a later change to configured defaults still applies.

Fixed

LiveViewTestClient.mount() never took the WebSocket branch, so WS-only setup code gated on hasattr(self, "_websocket_session_id") — the pattern djust's own #1612 presence guard documents — had zero coverage under a green test suite (#2821).mount() now stamps the same identity attributes the real WS/SSE mount path sets on a live connection (_websocket_session_id, _websocket_path, _websocket_query_string, _djust_mount_view_path, mirroring ViewRuntime.dispatch_mount) by default, making the documented "simulates the WebSocket mount process" claim true. Pass mount(via_websocket=False) to instead exercise the HTTP-prerender branch; the branch taken is recorded as client.via_websocket so a test can assert it explicitly rather than the branch being implied. Because that branch is now the default, a view's mount()-time track_presence() really runs under the client — which exposed a degenerate identity: with the default presence_unique_per_connection = False the identity is f"anon_{session_key}", and the client's session is deliberately unsaved, so every client collapsed to anon_None and two clients counted as ONE presence. A missing session_key no longer produces a degenerate identity. 8 regression cases in python/djust/tests/test_testclient_websocket_fidelity_2821.py.
Tick path now honors _skip_render: _tick_once previously called _snapshot_assigns() unconditionally before AND after every handle_tick(), even when the handler explicitly set _skip_render = True (e.g. a non-host session's tick that early-returns). The tick path now checks _skip_render right after handle_tick() runs — mirroring the event paths (server_push, db_notify, runtime.dispatch_event) — and skips both the render AND the second (expensive) _snapshot_assigns()/deep_fingerprint call entirely when set. _force_full_html (#1981) still wins over _skip_render, so an explicit forced re-render is never silently dropped. Closes #2822.
LiveViewTestClient.send_event() returned a {"success": False, ...} envelope instead of raising when no handler existed for the event name, so a typo'd or renamed handler could pass silently in a suite that doesn't inspect every return value (#2823).send_event() now raises djust.testing.NoHandlerFoundError by default — matching the production WebSocket consumer, where a missing handler is an error frame, not a silent no-op. Pass send_event(name, raise_on_missing=False) to opt back into the old envelope-return behavior, e.g. for a test that deliberately probes the error shape. The sibling param-validation-failure branch is unchanged. 4 regression cases in python/djust/tests/test_testclient_send_event_raises_2823.py. Existing coverage in TestLiveViewTestClient (the testing-utils suite) was updated for the new default, with new cases covering the escape hatch.
V008's return-annotation escape hatch never fired for methods or for dotted calls, so the documented non-noqa remedy — annotate the helper -> str — silently did nothing (#2825). The collector only inspected module-level functions and stored bare names, while every attribute call resolves to a dotted name; together the two gaps covered both shapes a real mount() call site takes. Methods of module-level classes are now collected and the comparison uses the call's final name segment, so a same-module helper or method annotated -> str (or any primitive) settles the check without # noqa or suppression. Unannotated helpers still report; annotations on helpers imported from other modules are still unresolved, and the check's hint now says so. 7 regression cases in python/tests/test_checks_v008_escape_hatch_2825.py.
T018 / manage.py djust_typecheck false-positived on every template reference to a framework mixin's injected state — {{ form_data }} / {{ field_errors }} on any FormMixin-based view reported as resolving to nothing (#2827). The static context extraction deliberately skips djust.* modules when AST-walking the MRO, so the mixin's runtime self.form_data = ... assignments were invisible to it. Framework mixins can now declare their injected template-visible keys in a _djust_injects_context class manifest that the extractor reads without needing to AST-walk framework source; FormMixin ships the manifest, and an anti-drift test pins it to cover every public self.x = ... assignment in the mixin so future additions cannot silently go blind again. New cases in python/djust/tests/test_djust_typecheck.py cover the manifest read, the end-to-end _check_view path, and the pin.
A server-initiated (tick/async) patch arriving while a user event was in flight forced a spurious full-HTML recovery morph (#2829). The client buffers such a frame while events are pending, and never applied its version, so the next event response failed the strict clientVdomVersion !== data.version - 1 check and logged VDOM version mismatch! Expected vN, got vN+2 — the recovery storm #1677 fixed for the push_to_view self-broadcast path only; the tick and async paths never went through server_push. The version is now consumed when the frame is buffered, but ONLY when it is contiguous with the cursor (it has arrived and will be applied), and the deferred replay is marked so it can neither move the cursor backwards nor trigger a recovery. Contiguity is the guard that keeps detection intact: consuming a NON-contiguous version would vouch for versions the server allocated and never shipped — the drop class _hotreload_broadcast_suppressed exists for — leaving the client permanently diverged with recovery never firing. A dropped patch, whether outside the buffering window or INSIDE it, still forces recovery — and the decision to decline a version survives to the flush, because a server noop carries no version of its own and would otherwise close the window with no strict check at all, letting the replay vouch for the gap. The client-owned flags are stripped by a shared helper each transport calls at its inbound entry (WebSocket, SSE, HTTP fallback), so a wire-supplied flag cannot suppress detection on any of them. 7 regression cases in tests/js/tick_buffer_version_desync_2829.test.js, plus a wiring case in the SSE suite (LiveViewSSE).
_run_async_work rendered without the consumer's _render_lock (#2830). Three of the four render paths — server_push, db_notify and _tick_once — acquire it; the async-work path did not, while calling the same render helper whose docstring states that "the caller MUST already hold self._render_lock". That let a background result re-render concurrently with an event-path render on the same PyO3 view, whose VDOM baseline is not thread-safe. It is also the server-side enabler of the client-side version desync fixed in #2829: a lock-free async render is what can put a server-initiated frame in flight mid-event-turn. BOTH render arms of the async path now serialise on the same lock — the success arm and the error arm, which re-renders to display the error state and was missed by the first version of this fix — and unlike the sibling paths they WAIT rather than taking a bounded wait and skipping — a skipped async render is a result the client never receives, whereas a delayed one still lands. 2 regression cases in python/djust/tests/test_async_work_render_lock_2830.py.
Dotted dj-keydown.enter / .escape / .space (and dj-keyup.*) never fired (#2831). A dot is a legal attribute-name character, so dj-keydown.enter is ONE literal attribute that closest('[dj-keydown]') cannot match and getAttribute('dj-keydown') cannot read — meaning the modifier-parsing block below it was dead code by construction, and the directive was silently inert. The framework documents the form as legitimate (_warnUnrecognizedDjModifiers deliberately does not warn about it) and it is now discovered by attribute NAME, the same prefix-matching scan that warning function uses. The required key comes from the attribute name, mapped through the module's own _normalizeKeyName, so the whole key set works instead of a three-entry allowlist — including the bare direction words: docs/website/guides/tutorials.md documents dj-keydown.right, which fell through the name map to the raw string while e.key is ArrowRight, so that documented example never fired either.
Keyboard delegation now dispatches every matching binding, not just the nearest one (#2831). Three defects fell out of the old single-closest() dispatch: an element carrying several dotted bindings (dj-keydown.enter + dj-keydown.escape, documented as a pair in core-concepts/events.md, core-concepts/templates.md, guides/template-cheatsheet.md and ai/templates.md) had only the first reachable; a binding whose key did not match returned early and swallowed the event before a container-level handler could see it; and the per-element rate-limit cache was keyed by element alone, so a morph swapping dj-keydown.enter for dj-keydown on a surviving node kept dispatching the stale binding forever. That same single slot also rebuilt the wrapper on every keystroke for an element carrying two bindings, resetting the timer inside debounce() — so dj-debounce fired N events for N keystrokes — and, for dj-debounce="blur", leaking one blur listener per keystroke, since the deferred form attaches its listener when the wrapper is created and cancel() clears timer state only. The cache is now keyed by (element, matched attribute): one persistent wrapper per binding.
Behaviour change: two bindings for the same directive on one path both fire (#2831). With the old nearest-closest() lookup a descendant carrying dj-keydown shadowed an ancestor's, so <div dj-keydown="outer"><input dj-keydown="inner"> fired only inner; it now fires both. This matches the scoped dj-window-keydown / dj-document-keydown delegation, which already dispatches every matching registry entry. No documentation pins the nesting semantics either way, so it is called out rather than changed silently. dj-key is deliberately NOT read as a key: it is the VNode list-identity attribute (heading "dj-key / data-key — Stable List Identity" in docs/website/advanced/vdom-architecture.md; the gloss "Analogous to React key" is in docs/website/guides/template-cheatsheet.md — two different files), and consulting it silenced handlers on keyed list rows. 15 cases in tests/js/keydown_dotted_modifier_2831.test.js.
A scoped event listener (dj-shortcut, dj-click-away) no longer keeps firing after the template stops declaring the directive on an element that survives the VDOM patch (#2832). The invariant — a scoped listener is evicted when its element detaches, when the server drops the declaring attribute, or when no LiveView root governs the element anymore — was implemented twice, once per scoped-listener path, and the #2108 fix landed on only one of the copies (the same parallel-path drift as #1646/#2110). Both paths now evict through one shared predicate, eviction is per declaring attribute (an element may carry several scoped directives and lose only one), and eviction clears the bound-handler marker so a re-declared attribute re-binds with the current value instead of serving the old closure.
T018's {% extends %} skip was invisible: manage.py check / djust_check reported "All djust checks passed!" for an app whose extends-based templates the check never examined — indistinguishable from a run that examined everything and found nothing (#2833). The skip itself is unchanged (the documented v1 trade-off: block-override context is inheritance-scoped), but a run that skipped one or more views now emits one Info-level djust.T018 message with the skipped-view count, so a pass is falsifiable. docs/system-checks.md also no longer claims the two entry points "can never drift apart" — extraction is shared, but coverage deliberately differs: T018 additionally covers inline template = "..." views and skips extends templates, while manage.py djust_typecheck covers extends template_name templates but not inline ones. 2 regression cases in python/tests/test_checks_t018_extends_skip_info_2833.py.
set_changed_keys() can no longer be silently dropped by _skip_render on any render path: when a handler set _skip_render = True alongside set_changed_keys() (the _force_full_html hatch, #1981), most render paths resolved the contradiction as "skip" — silently dropping the explicitly requested render and leaking the flag into a later unrelated turn (the #1646 silently-dropped-hatch class) — while the tick path rendered. All render paths (events, broadcasts, DB notifies, ticks) now resolve the two flags through one shared decision where an explicit forced render always wins and both flags are consumed on the turn that serves them. A handler setting only _skip_render is unaffected. Closes #2834.
check-changelog-test-counts no longer counts prose as JS tests (#2839). The JS test-function regex was "deliberately loose" — any whitespace/;/{-preceded it/test before ( counted — so a comment sentence like "…block below it (requiredKey)…" was counted as a test and a correct changelog claim was rejected as drift (the workaround was rewording the prose, which is backwards, #2238); a stray mention could equally inflate a count so a genuinely wrong claim passed. The regex now matches test declarations in statement position (line start, any indentation) and also counts the previously-missed it.each / test.each declarations (one per declaration, mirroring the Python side's parametrize-as-one convention). Validated against all 178 files in tests/js/ with per-file ground truth from vitest run: the old regex was wrong on 10 files (8 prose phantom matches across 6 files, 4 missed it.each declarations across 3 files — two files wrong in both ways); the new count is correct everywhere the count is statically knowable. 4 new cases in tests/test_changelog_test_counts.py cover the issue's exact prose example, the false-negative direction, the real-corpus comment shapes, and it.each counting.
handle_async_result now runs under the render lock on both arms of the async-work paths (#2840). A background-work handler that mutates view state (the documented self.result / self.error pattern) could previously interleave with a concurrent lock-holding render of the same view. LiveViewConsumer._run_async_work awaits the handler inside its existing _render_lock region (success + error arms; the error arm's handler → identity re-check → re-render ordering is preserved), and the runtime twin ViewRuntime._execute_async_task — which serves WS events' start_async work after the ADR-022 flip — borrows the consumer's lock via transport.event_context() for its handler + render and gained the missing #1940 teardown identity-guards on both arms. 6 regression cases in python/djust/tests/test_async_result_render_lock_2840.py.
Embedded-view stamping missed dotted event attributes (dj-keydown.enter was never stamped, #2841).{% live_render %} stamps data-djust-embedded on every event-bearing element so the client can route the element's events to the embedded child view, but the compiled matcher required = immediately after the bare attribute name — and a dot is a legal attribute-name character (#2831, #1999), so dj-keydown.enter="go" is ONE literal attribute that could not match. An element whose ONLY event attribute was dotted got no stamp, and its events routed to the parent instead of the embedded view; a bare event attribute elsewhere on the same element masked the failure. The matcher now accepts dotted in-name modifiers (dj-keydown.enter, the multi-dot dj-keydown.enter.shift the runtime reads as .enter, and whitespace before = per the HTML tokenizer). The scoped family (dj-window-keydown / dj-document-* for keydown, keyup, click, scroll, resize) was missing from the attribute list entirely — missed dotted AND bare — and is now stamped too; its dispatch path consumes the same view_id from addEventContext. 8 cases in TestDottedKeyboardEventStamping (tests/unit/test_live_render_tag.py).
Unknown dotted keyboard modifiers are no longer silently inert — debug warning added (#2842).dj-keydown.f1="go" or dj-keydown.esc="go" never fired and never logged anything, even in DEBUG: _normalizeKeyName maps a fixed set of names and falls back to the RAW name, so .f1 resolved to "f1" while e.key is "F1". The raw fallback stays (it is what lets single characters like .a fire), but _warnUnrecognizedDjModifiers — the #1999 debug-only warning channel, zero production cost — now also flags any multi-character key modifier on dj-keydown / dj-keyup / dj-window-keydown.* / dj-document-keydown.* that is not in the map, naming the attribute and the recognized names. Warning is once per bind, never per keystroke. Notably, HTML parsers lowercase attribute names, so dj-keydown.PageUp arrives as .pageup and cannot match "PageUp" either — casing cannot rescue a multi-character suffix, and .F1 / .PageUp-style spellings warn like every other inert name. The map moved from a local of _normalizeKeyName to a module-level _KEY_NAME_MAP so the warning pass reads the same source of truth. 11 cases in tests/js/keydown_key_warning_2842.test.js.
A stale handler closure kept serving dj-shortcut / dj-click-away after a re-render changed the attribute VALUE on an element that survives the VDOM patch (#2845). The bind loops skipped already-marked elements on the _isHandlerBound marker alone, and the #2832 eviction predicate judges attribute PRESENCE — so a present-but-changed attribute kept dispatching the old handler name (and, for dj-shortcut, the old key bindings and the stale dj-shortcut-in-input gate). The skip is now keyed on the value the listener was built from: an unchanged value skips exactly as before, a changed value evicts the old listener and rebuilds the closure from the new value — the sweep path's counterpart of the #2108 registry refresh. 5 regression cases in tests/js/stale-scoped-listeners-2845.test.js.
Both deferred-activity re-dispatch paths resolved _skip_render vs _force_full_html the pre-#2834 way (#2847). With both flags set by the same handler, ViewRuntime._dispatch_single_event and the WS consumer's _dispatch_single_event sent a noop — silently dropping the forced full-HTML render and leaking _force_full_html into a later, unrelated turn. Both now resolve through the shared _resolve_skip_render helper (force wins; _skip_render consumed whenever set), and the helper's docstring no longer falsely claims to own only the four paths #2834 named. New cases in TestDispatchSingleEventParity2847 (python/djust/tests/test_skip_render_force_parity_2834.py).
Release gates: a tagged release can no longer ship without its CHANGELOG section, and fragment path/class claims are now checked (#2854, #2849). v1.1.3 shipped to PyPI with no ## [1.1.3] section while the shipped-section pin reported OK against v1.1.2 — the pin's anchor silently fell back to the newest sectioned tag. scripts/check-changelog-tagged-sections.py now fails for every release tag above that anchor reachable from HEAD whose version has no working-tree section, and make release refuses to tag when CHANGELOG.md has no section for the target version (the pre-commit hook cannot see this: at the version-bump commit the tag does not exist yet). Separately, a new fragment reference check resolves backtick-quoted file paths and test-class names in pending fragments against the tree, closing the same #2652-shaped gap for fragments; count claims were already covered by scripts/check-changelog-test-counts.py.
Three remaining stale-closure sites of the #2845 class now rebuild when the declaring attribute changes on a surviving element (#2858).dj-poll (09-event-binding.js), _bindModel (20-model-binding.js) and bindUploadHandlers (15-uploads.js) skipped already-bound elements on the bind marker alone, so a morphdom-surviving element kept dispatching the old closure — the old poll handler and cadence, the old model field, the old upload slot. Each site now keys the skip on the value the closure was built from: an unchanged value skips exactly as before (for dj-poll, an unchanged value/interval pair does NOT restart the poll phase — the bind loop runs on every patch and a restart would reset the interval timer), a changed value evicts the old listeners and rebuilds, and the poll phase reads its data-* params at fire time like dj-click/dj-change do. 10 regression cases in tests/js/stale-closures-2858.test.js.
dj-shortcut key names that can never match now warn in debug mode (#2859). The comma syntax (pageup:handler) resolves through the same _normalizeKeyName helper as the dotted keyboard directives, and an all-lowercase multi-character name (pageup) can never equal a KeyboardEvent.key — the binding was dead on arrival with no warning. Attribute VALUES keep their casing, so the correctly-cased raw spelling (PageUp:handler) fires and is deliberately not warned about; each distinct dead name warns at most once per bind pass.
dj-document-scroll / dj-document-resize now warn in debug mode (#2859). Both were recognised by _scanScopedElements and their entries registered, but the document-level listener was deliberately never installed (resize never fires on document), so the attributes parsed and did nothing — the same silently-inert class as #2842. The warning points at the documented dj-window-scroll / dj-window-resize twins.
Deleting an already-shipped CHANGELOG.md section no longer passes the shipped-section pin silently (#2862). The #2028 pin iterated only the sections present in the working tree, so a shipped section deleted from the tree was never compared — removing the ## [1.2.0rc6] section exited 0 with OK: 133 shipped CHANGELOG section(s) match the newest release tag. The pin now iterates the union of the working tree's sections at or below the anchor and the anchor tag's frozen snapshot, so deletion and rewrite are two symptoms of one comparison; a deleted section fails by name with a restore instruction, a distinct message from the rewrite diff because the operator's next action differs. The multi-branch scoping from #2861 is preserved — the new demand reads the anchor's snapshot, never the tag list, so a maintenance branch is not failed for main's sections. Covered by synthetic tests in TestDeletedShippedSection and a real-tree canary in tests/test_changelog_tagged_sections.py.
A full CHANGELOG.md wipe passed the shipped-section gate silently (#2865).scripts/check-changelog-tagged-sections.py anchored on the newest tagged section and failed OPEN when it could not select one: deleting one shipped section was caught (#2854 absence, #2862 deletion-below-anchor), but deleting EVERY tagged section — the realistic tail of a cross-branch CHANGELOG.md merge resolved toward a branch without the shipped history (the v1.1.0rc5 consolidation class) — left the check with no anchor and exited 0. The no-anchor state is now disambiguated by the release tags reachable from HEAD, enumerated by the #2861 walk (kept --merged HEAD, extracted into one shared helper — no third tag call): with no release tag reachable, the fresh / pre-first-release pass stands; with some, every one of them lacks its section and the check fails naming the newest tag and its restore source (git show v<newest>:CHANGELOG.md). New cases in TestWipedTaggedSections in tests/test_changelog_tagged_sections.py.
dj-keypress, dj-viewport-enter, and dj-viewport-leave removed from the {% live_render %} stamp list (#2869). They sat in _LIVE_RENDER_EVENT_ATTRS — the list {% live_render %} scans when stamping view_id onto embedded elements — but no client module ever bound them: zero mentions in static/djust/src/, zero occurrences in the built client. An element carrying one of them inside an embedded child was stamped exactly like a working directive while no listener was ever installed — no error, no warning, no event. dj-keypress is also deprecated in the DOM in favour of keydown (which djust ships with a full modifier system); there is no viewportenter DOM event for the viewport pair to bind at all. dj-keypress, dj-viewport-enter, and dj-viewport-leave were all introduced in the same v0.6.0 stamp-list commit (e9907c7f) — the invariant test below caught the viewport pair on its first run. A new invariant test — new cases in tests/unit/test_live_render_event_attrs_invariant.py — fails whenever a stamp-list entry has no client-side binding: each entry must appear as a quoted string literal in some static/djust/src/ module, or belong to the dj-window-/dj-document- scoped cross product derived from the client's own scopedPrefixes × scopedEventTypes arrays. Its blind spots (a quoted mention inside a comment, a dead string reference that never dispatches, dynamic names outside the scoped path) are documented in the module docstring.
Template pipeline performance: Reduce template rendering and reactive update overhead by reusing exact immutable temporal values after protection checks, memoizing filesystem include selection per render, streaming VDOM serialization into one buffer, and reducing lazy tag-binding conversion overhead. Preserve live subclass and custom-timezone behavior, template reloads between renders, and escaped hydrated HTML output.
Avoid repeated dir() scans for numeric lookups on exact built-in sequences during template rendering, while preserving Django's lookup behavior for custom objects and subclasses.

Security

HTTP POST fallback (RequestMixin.post()) enforced none of djust's three authorization layers, so an unauthenticated or under-privileged caller could drive @event_handler methods on a login_required = True view with a plain POST.get() and every WS/SSE event path already enforced view-level check_view_auth, handler-level check_handler_permission, and the ADR-017 object-level check; the POST transport now runs all three before dispatch, with the same denial shapes (403 {"redirect": <login_url>} for the unauthenticated case, 403 {"error": "Permission denied"} for view/handler permission denials, 403 {"error": "Access denied for this object."} for object-level denials). Anonymous POSTs to views without auth requirements are unchanged. Also fixes an UnboundLocalError in post()'s own error path: a body that was not valid JSON raised from the except handler itself, masking the real exception as an unlogged 500. 7 regression cases in python/tests/test_http_post_authz.py.

All releases · Atom feed