Release dependencies

This release adopts SDK 0.2.6 and the App versions pinned in composer.lock. People and Time & Absence advance to 0.3.0 for their renamed CLI commands. Other App releases retain their existing stable or beta channel and declare Core 0.3 compatibility. Update external command invocations before deployment; old command aliases are unavailable.

0.3.0

Added

A new Data Center lists registered migration targets from Core and installed Apps, offers only supported source systems, and opens the existing import workspaces. The original Data Migration page remains available.
A dedicated getting-started workspace offers legal-entity document upload, Agent-assisted visible drafts, and user-confirmed saving. Required setup fields appear first; additional information can be expanded. The top-bar progress button resumes setup, while the provisioning welcome opens it once after the first login. Existing setup evaluators and domain write permissions remain authoritative.
nexia-demo:seed-apps --app=<key> seeds selected installed App fixtures without resetting the demo tenant or rerunning organization, access, and report seeders. Omit --app to run all supported installed App contributions. The existing demo fixture key and primary domain are both required; outside local environments, pass --force. No SDK upgrade or migration is required.
Developers explains the Agent request lifecycle, model roles, tool lanes, and isolated analysis, with flowcharts and local/production configuration maps.
Central Analysis Executions provides a read-only view of analysis outcomes, dataset sizes, collected stage timings, and execution limits. Diagnostics are separate from token billing and retain no prompts or dataset contents.
Central Agent Role Models assigns ordered catalog models to Router, Analysis Worker, and Research Worker through relational model references. Primary grades keep their existing assignments.
Ordinary Agent requests use the Router model to interpret intent and required capabilities. Roles without usable candidates inherit the request’s Primary grade and full fallback chain. Explicit role failures still stop the affected execution; ambiguous requests return a clarification question. Registered commands retain their deterministic path. The command menu is labeled Commands.
Access Management now starts with a user search and presents only the roles and grant history the acting administrator may manage. Tenant, Legal Entity, and Operating Unit assignment authority remain independently scoped. The new access-user read endpoints do not broaden the tenant user directory.
Direct role synchronization rejects new Protected Access roles; the existing request and approval workflow remains required.
Legal Entity and Operating Unit scoped lists now expose only legal_entity_roles or operating_unit_roles; their detail and history payloads are limited to the exact selected scope. Clients must stop inferring tenant authority from the former merged scoped-role responses. Operating Unit updates now return only operating_unit_roles. Empty Legal Entity and Operating Unit role sets now revoke the last direct role while preserving assignments from other sources.
Agent preparation can provide optional, operator-configured Router and Worker model candidates without changing the user's Primary grade. Authorized analysis datasets are bound to the current tenant, actor, chat task, live Resource/row permissions, and a short-lived controller export token.
Reports now separate the Create action from the All, Mine, Shared, and Archived collections in the sidebar. The library rows align with their headers, and report details prioritize Edit or Repair while keeping secondary actions quiet.
The report library supports title search within the current access and organization scope. Authoring groups analysis scope with data selection and clears source-specific conditions when that scope changes.
Report authoring now guides data selection, typed values, configuration, and preview recovery. It also includes a relationship graph for choosing connected data sources and saves only supported relationships.
The report composition catalog now includes additional owner-published communications and App business resources and labeled fields, while retaining each resource's existing read authorization.
Core directory and organization resources now publish safe composition fields through their existing read authorities; Operating Unit rows retain their effective Legal Entity and unit-level scope.
Provider-published filter operators are preserved in the catalog when they are supported by the shared allowlist; query execution remains the authority that validates a report.
Added tenant-user favorites for server-authorized pages and durable records. The Shell resolves current titles and destinations on every read, so a favorite never preserves access after a permission or installation change. Standard Resources expose favorite metadata only when their existing catalog, read permission, stored-owner authority, and Show route are all present. The shared primary-cell and pane-local detail header show the same server-authorized record affordance; related record links do not inherit the row's favorite identity. Active favorites use a filled yellow star without raised pressed chrome while retaining keyboard focus visibility.
Eligible Resource lists now publish a Favorites-only filter and canonical meta.list_schema.resource_key. After a record is added or removed, the Shell refreshes matching cached lists, including an active Favorites-only view, without guessing an App query key. The shared control updates its current state optimistically, then reconciles with the server.
The full Favorites view uses stable cursor pagination for page and record targets. Apps do not register a favorite-only contributor or a separate favorites list path: standard records use the shared stored-owner checks, while custom records remain available only through an owner-provided, authorization-safe summary and canonical route.
Added the /dashboards library with server-side title search, sorting, favorites filtering, and cursor pagination. Existing dashboard list clients retain the dashboards array and can adopt next_cursor; saved dashboard creation is no longer capped by the rail preview.
Adopted App SDK v0.2.4, People v0.2.5, and Talent v0.2.0-beta.4. The SDK's ResourceListFieldsAugmenterRegistry lets the host add the Favorites-only filter without an App-specific query contract.
Added centrally managed Agent Models and fixed Agent Grades. Users choose Auto, Light, Standard, Deep, or Deeper for the next request; operators assign ordered model fallbacks and a requested reasoning effort without exposing provider model choices. A model that does not support the requested effort now uses its model default automatically.
Added platform-only OpenAI and Anthropic model discovery with manual and monthly sync. Newly discovered models are disabled until an operator reviews them.
Clarified central Agent Model settings: the provider-reported output maximum is read-only, while the service output limit remains configurable per API call. Pricing now distinguishes unregistered rates from zero and applies changes only to future calls.
Agent Model sync feedback now names each provider, shows the actual safe failure reason and recovery step, and explains when a catalog row is absent from a successful provider sync.
Agent Models now show Central platform configuration readiness and safe setup reasons without making paid provider health checks or assumptions about tenant BYOK.
Adopted People v0.2.4 to fix Composer post-install frontend dependency validation. Its frontend SDK peer follows the host policy.
Fixed shell panel, loading-state, combobox, and dashboard spacing consistency.
Dashboard widget catalogs activate their owning Apps' translations before rendering, including when changing language without visiting an App page first.
Fixed local Developers navigation to retain its local development port.
Select options use their highlighted background as the focus indicator, avoiding an extra blue ring on pointer hover.
ResourceTable accepts defaultView="cards" when a card renderer is provided; saved user preferences still win. The Developers live and copyable examples now show both initial views with localized controls and shared row actions. App consumers need the corresponding SDK defaultView frontend contract.
Added an optional ResourceTable card view for complete list renderers. Table remains the default; saved view and card-column choices are scoped to the tenant, user, and stable table ID. Cards omit the outer collection frame and padding, use tighter inner padding, and match the WorkDesktop content surface.

Changed

Employees in another legal entity can receive, claim, and complete assigned process tasks using active membership and exact task permissions in their own entity. Task inboxes and notifications include those assignments without granting general access to the executing entity. Foreign task context exposes explicitly mapped inputs and evidence, and reassignment or loss of eligibility revokes active access. Existing local permissions and assignment records are retained; no database migration or SDK update is required.
Agent task history, process assignment groups, and signature operations use the shared organization selector for Legal Entity scope. Existing scope URLs and permissions are retained; no migration is required.
Business process and decision-rule creation opens the authoring screen directly. Company-wide definitions offer assignees from all authorized legal entities with their affiliations, without a reference-entity selector. Legal-entity-specific definitions retain an owner selector. Shared definition references and Approval presets now use their common catalog and existing tenant permissions. Publishing a shared business process links to legal-entity usage settings; per-entity versions and execution rules are unchanged.
Core Artisan commands now use functional sections such as nexia-apps:, nexia-agent:, nexia-access:, nexia-runtime:, and nexia-demo: instead of one nexia: section. Update external command invocations using the mapping in docs/reference/ARTISAN-COMMANDS.md; old names have no aliases. Repository scripts, Composer hooks, and schedules use the new names. The standalone contribution-cache clear command and Agent usage demo seed command were removed; runtime cache clearing and the underlying demo seeder remain available. People and Time & Absence commands now use nexia-app-people: and nexia-app-time-absence:; adopt the matching App changes and update external calls from people: and time-absence:. No SDK upgrade or database migration is required.
Reference-picker refresh buttons use a smaller 24px control and 12px icon.
Resource forms can use the optional NxResourceInformationForm.legalEntity contract to place Legal Entity input inside the first section at a consistent field width. Edit shows the stored owner's label. Core and App resource templates adopt it; existing Apps must migrate their forms after adopting the matching SDK and Core host.
Primary-cell favorite buttons appear on hover or keyboard focus unless already favorited. Touch devices keep the buttons visible.
The Central App Catalog now groups entries by App Family by default, using localized group headings and the existing family order. No upgrade action is required.
Adopted App SDK v0.2.3 and People v0.2.3 in the committed Composer lock for shared resource cards and employee list cards.
Reports now open directly from the Workbench rail with library views in the Report sidebar. Dashboard editing can create a Report in a contextual Work Tab and returns the saved revision to the originating dashboard draft.

Upgrade required

The expanded composition catalog requires App SDK 0.2.4, Assets 0.2.0-beta.3, Payroll 0.2.2, People 0.2.6, Recruiting 0.2.0-beta.3, Talent 0.2.0-beta.5, and Time & Absence 0.2.4.

Fixed

Task assignment and completion commands recheck current membership and permissions before replaying a previous result, so suspended membership cannot reuse cached authority.
Signed-document lists correctly validate sorting instead of returning a server error.
Site lists sort site types and statuses by their displayed translated labels, with stable ordering for equal labels.
Shared approval forms appear once in the creation catalog. Shared forms open with the first eligible entity selected above the title. New personal lines use the same inline selection above their name. Entity-specific forms and saved documents retain their owning entity. Form scope and document ownership are displayed separately.
Template and shared approval-line administration now open across readable entities, with a header entity filter. Tenant-wide records appear once and entity-owned records show their owner.
Personal approval lines now list the caller’s lines across authorized legal entities, with an optional entity filter. Start an approval from the inbox: available forms appear across authorized entities, with an entity column and an optional header filter. Personal lines sort from name and modification-date headers; table resize dividers are thinner. Existing creation links redirect to the nested inbox route while retaining the target and draft.
Approval, Signature, and Process side menus separate personal work, activity and history, and configuration. Receiving and tracking work appear first; shared approval routes and workflow activity have clearer names. Existing routes and permission requirements are retained.
Generated App resource lists distinguish pending read authorization from access denial. Demand Planning lists now show loading while the context and read scope are being resolved, and keep list queries gated on confirmed permission.
New demo fixtures use business-specific names, item specifications, periods, and request details across Core and Apps. Maintenance examples advance through real task completion and cancellation. The primary demo account receives all business permissions except authorization administration and account ownership transfer, with approvals and tasks assigned to it; the admin retains full access. Existing fixture records are preserved where seeding already skips them. Recreating the demo tenant with the matching App fixture sources applies the full dataset.
Demo report seeding reuses installed App lookups within each account's seeding operation and prints account-level timings. Installation changes still invalidate the lookup; report authorization and execution checks remain in place.
Toast action buttons now use the shared primary button and follow the theme's accent color instead of using the body text color as their background.
Agent approval history now says “Approved” without a permanent “running” label. Approval records no longer imply that a finished tool is still executing.
Image and other analysis requests no longer send the analysis delegation tool twice when it is already in the selected palette, preventing provider rejection for duplicate tool names. Restart Agent Gateway to load this fix.
OpenAI stream overload refusals before any output are retried once, then use the configured fallback chain. Partial responses and unknown failures are not retried.
Composition tools describe v4 predicates, calendar grouping, and numeric aggregates explicitly. Aggregate sorting resolves grouped output columns; the shared SQL deadline starts after authorization and planning.
OpenAI models use the Responses API for reasoning with function tools, with server-side response storage disabled. Router classification also accepts structured text blocks returned by provider adapters.
Model administration supports provider and enabled-state filters. Grade model choices include only globally enabled models, regardless of API-key readiness. Each linked model can be paused in its grade without losing its priority; existing links remain enabled after the Central migration.
Report and dashboard analysis uses Resource-generic discovery and queries instead of enumerating each installed App's endpoints, while retaining report, dashboard, and attached-file tools. This avoids tool-count overflow as Apps grow. Internal Router JSON is withheld from chat, and provider request rejections retain their original error instead of being reported as missing usage.
Anthropic credit exhaustion before any response chunk permits the configured fallback model to run, while generic invalid requests remain non-retryable.
Agent request interpretation allows up to 15 seconds per model within a 30-second total deadline, avoiding premature fallback after roughly 1.5 seconds. Router logs identify the provider, model, and elapsed time for each failed or completed attempt; deadline failures retain a distinct timeout reason.
Agent GET tools preserve array and nested query inputs when calling Core, fixing task inbox requests rejected with 422 when multiple organization targets are supplied. Empty query collections are rejected rather than silently omitted.
Process and decision draft editors accept null optional plan fields from the Agent while retaining required-field validation and explicit extension clears.
App Catalog shows deactivation impacts in a compact red popover trigger, so long warnings no longer stretch the card grid. Apps that are not active omit this notice.
Access catalog requests now have a 30-second timeout per attempt and are cancelled when their last consumer unmounts. The role list's existing error and retry state can recover a stalled catalog request without a page reload. Existing query retry rules still apply. No migration or SDK update is required.
Agent conversations read across authorized organization targets without a global Legal Entity selection. Process catalogs use unscoped routes, the personal inbox uses the unified task endpoint, and data/composition queries apply authorized targets before pagination. Explicitly scoped service tokens remain restricted; ambiguous assignment-group creation requires a target. No SDK or App upgrade is required.
Analysis datasets accept inclusive date_ranges on queryable date columns independently of list filters, including for historical records. No SDK or App update is required.
Approval administration now resolves Create targets and authoring catalogs from Create authority instead of requiring template or route-policy read access. Tenant catalog creators can create tenant-scoped configurations without supplying a placeholder Legal Entity, including fixed-user route steps. Generated Legal Entity-owned forms now keep their create target selector available for zero, one, and many authorized targets and do not treat malformed target URLs as tenant-wide access. Media edit availability now follows the server's record-level authorization, including permitted uploaders and Legal Entity-scoped updaters.
Tenant Resource Reference option lookup now requires matching tenant consumer and target authority through the SDK helper; the existing Legal Entity route remains compatible. Consumers using the new tenant route must send the tenant-owned reference context rather than a Legal Entity placeholder.
Moving multiple selected BPMN elements now retains every element's position and the selection after release. One Undo restores the whole move. No upgrade action is required.
Creating a decision rule from BPMN now starts with the available recommended template. Authors can review or change it before saving; catalog refreshes do not replace an edited draft. No upgrade steps are required.
Agent delegation token requests now reject malformed session UUIDs with a validation response instead of a PostgreSQL error. Callers must send the chat session's public UUID.
Search target resolution now distinguishes record ownership from permission scope, so authorized Operating Units remain searchable without exposing ungranted units.
Legal Entity Agent reads retain membership-based visibility. An explicitly targeted composition catalog only exposes permissions held in that Legal Entity, including its authorized Operating Units.
Package diagnostics now resolve Composer-installed Apps as well as local source checkouts. Generator commands retain their existing source-only scope.
Added the missing Agent model-grade assignment foreign-key index. Apply the new central migration during upgrade; existing assignments are preserved.
Local and production environment examples document optional Agent capacity and session-shard settings without overriding inherited Redis configuration.
Categorical resource filters now explicitly support multiple selections, with OR within a category and AND across categories. List authoring rules and generated model stubs preserve this policy without changing omitted runtime options. Report, approval, signature, directory, and usage lists add supported column sorting before pagination; filter clearing no longer restores an unwanted report archive constraint.
The resource toolbar orders Refresh, Filter, then More. Report source relationships use a draggable force graph with authorized catalog edges and controls for isolated nodes.
Report management uses a full-width table without a side menu. Ownership relationship and archive state appear as columns and independent toolbar filters; creation remains in the page header. Existing view links continue to select their previous filters.
Shared section cards vertically center their empty and error states, including Approval, Signature, and Process overviews and cards stretched by taller neighbors.
Reports and artifacts share favorite row controls and More menus. The Reports menu links to report management and exposes permitted edit, duplicate, and archive actions.
User imports accept the linked person's legal name, contact email, phone, and address. Account email remains the duplicate key; existing accounts are still skipped, and no HR records are created. File columns can be reassigned to a field already mapped elsewhere.
The artifact library opens inside Work Tabs. The Reports rail menu exposes favorites and recently updated reports. Dashboard ownership uses Legal Entity wording, and the signature template list uses the shared Legal Entity selector width.
Signature template authoring waits for a Legal Entity selection before requesting its data catalog or accepting source uploads. Process permission loading and failures no longer appear as access denials; the Approval inbox waits for read access before showing its submit-only state.
Report and dashboard creation use the same rounded action button as Business Process.
Business Process and decision-rule libraries show all authorized Legal Entities when no list scope is selected, instead of a permission error. Explicit selections remain permission-checked; existing singular Legal Entity links still work. Creation keeps a separate target choice, and row actions retain the definition's owner or authorized read context. No migration or SDK update is required.
Dashboard management now opens in its own Work Tab instead of retaining the previously selected surface after navigation.
Dashboard management now uses the standard table controls while retaining its organization scope, default status, and cursor loading. Title, created date, and modified date sort through column headers in either direction. The standalone favorites filter is removed; record favorite actions remain available.
Spotlight now keeps Favorites as a visible category, including an empty state and a link to the full favorites list when more saved items are available.
The Reports library now uses the standard table controls and a compact detail favorite action. Existing search, organization scope, and cursor-based loading are retained.
Agent Gateway checkpoint pools now open no idle database connection and cap each tenant saver at two connections. Startup rejects a configured per-container checkpoint budget that could exceed its explicit ceiling; operators must size the database for all Gateway instances and reserve connections for usage-outbox delivery.
Agent execution admission now applies central global and per-tenant task limits, bounded queues, lease renewal, and fenced release by task turn. Provider calls have separate global and credential-pool limits; platform credentials share a provider pool while tenant BYOK credentials remain tenant-scoped.
Agent Gateway container shard expansion keeps legacy session routes stable. New shards require the server-minted session route set, so an unmapped recovered session cannot move because a container-name list changed. A post-cutover session carries its server-issued route-name set in the delegation token; existing sessions retain the legacy route until drained.
Agent conversations no longer load Canvas guidance or active artifact source on unrelated turns. Explicitly selected artifacts and their bounded follow-ups retain their source context; attachment data remains untrusted reference material.
Table presentations can select authorized output columns without exposing an internal identifier. Unknown or duplicate columns remain rejected.
Saving or canceling a contextual Report create returns to the originating Work Tab instead of selecting another report. Saving preserves the dashboard draft and adds the new report to it.
Fixed local Agent Gateway startup when dashboard proxy tools share the dashboard namespace with Gateway-executed dashboard handlers.
Agent dashboard and overview requests now use authorized report discovery, preview, and confirmed creation to return editable Dashboard specs backed by real report revisions. Standalone report requests create or reuse a saved Report; normal access and confirmation rules remain in effect.
Product guide screenshots can be opened at their original resolution in a new tab using a mouse or keyboard, making small interface labels easier to read. Procedure numbering continues across screenshots. No upgrade action is required.
Product guide detail pages use the Developers documentation navigation style, with collapsible work areas with direct guide lists, a highlighted current guide, and remembered expanded sections. Table of contents links scroll to the selected heading, and long sidebars can scroll independently. No upgrade action is required.
Agent dashboard previews and report saves now describe the required presentation input fields. Composition tool failures show localized explanations before request details, and dashboard data and report tool names have localized labels. No migration is required.
Agent composition previews, report saves, and pinned-revision queries exchange the session credential for the required source read permissions. Saved report exchanges resolve the authorized revision on the server; metadata remains readable when its source is unavailable. Existing source grants and organization boundaries still apply. Restart a local Gateway after updating; production requires the matching Gateway rollout.
Agent tool history no longer presents accumulated failed attempts as user corrections. Data-request failures identify Nexia as responsible, and raw request details appear only when expanded. Individual failures remain visible in the history without implying that a retry succeeded.

Tenant App installation and development

Added nexia-apps:install-dev-app <app> --tenant=<id> for local console development without publishing an App or changing its readiness. Production CLI and HTTP installation keep their distribution and readiness checks.
New tenants prepare Core schema only. App installation migrates the selected App and prerequisites before initialization; subsequent tenant migrations update installed Apps, including disabled installations. Deactivation preserves tables and data. Schema writes are serialized per tenant.
Existing tenant tables and migration history remain intact. New tenant provisioning no longer loads tenant schema dumps that may contain unselected Apps; older pending tenants retain their existing schema. Explicit migration paths remain maintenance overrides. Keep local package overlays and local databases separate from production.
Upgrade guidance: retain the normal tenants:migrate --force deployment step. Install new Apps through the tenant installer after Central catalog registration/publication and any private entitlements. Fix migration failures and retry installation; completed migration steps remain recorded.
Tenant migration --database selects a connection template without redirecting schema writes to that template database. --seed runs after all selected schema updates.
App scaffolding now rejects duplicate local/installed App keys and occupied package paths before writing files, including dry runs. Existing Apps may still add repository support files with --repository-files-only.

Report browsing and exploration

Report lists open an in-place preview; the full report keeps connections behind an explicit action. Resource selection now retains searchable App and Resource controls, confirms destructive source changes, and shows scope counts.
Relationship browsing includes unconnected resources in full-scope mode and visibly marks active toggles.
Saved single-source v1 and reference-graph v3 reports support result selections and source detail without rewriting revisions. Legacy shared-dimension reports retain their existing read/edit path and explain the exploration limitation. No reseeding or SDK upgrade is required.

Fixed

Resource generation uses translated organization validation messages and completes its translation check for tenant- and Legal Entity-owned resources.
Composition value aggregates honor App-declared required_dimensions, including derived and comparison queries. Select each same-source dimension or constrain it to a single value with a conjunctive equality filter.
App Process handlers receive the originating Token's elementId for durable Approval callbacks. Persist it with the exact Case and originating resource.

Upgrade

Run the central analysis-execution migration before using the monitoring view. Diagnostics default to 30 days of retention and use the application scheduler for daily pruning. Core rejects datasets exceeding the row or query-time limit instead of forwarding partial data for analysis. These are initial protection limits, not measured optimal capacity.

Run the central role-assignment migration before serving requests with the updated Gateway. Separate role assignments are optional and override Primary inheritance. The migration imports unambiguous legacy agent.role_models entries and logs skipped identifiers; the retained JSON is no longer the runtime source. No model is assigned automatically when legacy configuration is absent.

Local Agent/full mode now starts an analysis controller and builds an isolated SQL/Python runner. Set AGENT_ANALYSIS_CONTROLLER_TOKEN before using analysis. Production analysis requires a separately deployed Worker/Sandbox and matching service secrets. The production workflow now validates and deploys analysis, checks Worker liveness, deploys the Gateway bridge, then triggers Laravel Cloud. Configure the GitHub production Environment and install Core prerequisites before initial activation; CI does not populate Laravel Cloud secrets. See the analysis configuration guide.

Requires SDK 0.2.5 or later for ProcessWorkActionInvocation.elementId; the committed Composer lock adopts SDK 0.2.5. Existing generated controllers can replace their four organization validation literals with __('access.organization_target.select_one_legal_entity') and rerun generation without --force; existing customizations are preserved. Apps opt into aggregate dependencies through their field metadata. Procurement provider adoption remains a separate App integration step.

Custom forms can use NxResourceFormLegalEntityField inside their first input section to share the same full-width create selector and stored edit label as schema-based forms. Apps adopting it require the matching Core host and SDK; existing applicability and Operating Unit inputs keep their domain-specific behavior.
Added NxOperatingUnitSelector for full-width form inputs with host-localized defaults and the existing reference-picker loading, refresh, and create-continuation behavior. Apps supply authorized options and retain dependent-field resets and editability. This does not add an Operating Unit field automatically or change list/Shell scope.
Agent composition tools now describe the server's ASCII alias constraints, keeping translated table and chart labels separate from query identifiers. Exhausted model candidates fail ordinary tasks instead of reporting completion. Capacity refusals and Router interpretation failures retain distinct error codes in live chat and saved task history. Deploy Core before the Gateway so terminal callbacks accept the new provider_capacity and router_failed codes.
The App Launcher uses a compact three-column grid with smaller icons and selection areas. App families stay on separate rows with one heading per family.
The tenant App Catalog starts each app family on a new row with a single heading. Cards share a height within each row with compact description and action spacing. Publisher and visibility appear as secondary text instead of repeated badges. Relationship and deactivation details use consistent hover/focus hints; deactivation impact has red text without an alert background. Card actions are grouped in a More menu beside the aligned status badge. Beta badges are omitted from catalog cards to keep app titles on one line.
Central registration starts each app family on a separate row and shows its heading only once, including when the app grid wraps.
Agent conversations restore the selected grade, including Auto, when reopened. The requested policy is stored separately from the last resolved grade. Apply the tenant add_model_selection_to_agent_chat_sessions migration before serving this Core change; legacy conversations fall back to their last resolved grade. Terminal streams refresh the task feed after the durable callback, and delayed completion notifications are suppressed for results already viewed.
Router output has a larger bounded budget so classification JSON is not cut off by the former 256-token limit. Requests needing clarification reach the task model and its question-card tool; process recommendations reuse existing answers and offer concrete draft choices. Malformed or truncated routing output still stops execution.
Pasted and selected chat uploads can be stopped while transferring or waiting for server processing. The composer remains editable during upload, and cancelled or previous-conversation uploads cannot attach late results.
Approval creation places its Legal Entity selector in the page header, distinguishes target selection from permission denial, and skips catalog/draft prefetch until the destination specifies a target. Process instance lists expose the standard toolbar, multiple status filters, and server-side status/start-time sorting before pagination. No migration is required.
Process deployment and assignment-group administration prefetches use the destination’s selected Legal Entity and wait when no target is selected, avoiding premature requests with missing or unrelated scope.
ResourceTable data columns show resize handles and support dragging or keyboard adjustment. Double-click or Home resets one column; More → Reset column widths restores all widths. Narrowing columns keeps the table filling its viewport without expanding action columns. Widths are remembered in this browser per tenant, user, and table when a stable table ID is provided.
Resource-list filter counts reserve space around the trigger so badges remain visible in clipped containers. Active filter chips in approval, signature, task, and administration lists now sit beside the filter trigger instead of the search field.
Production Agent deployment reads CLOUDFLARE_ACCOUNT_ID and AGENT_ANALYSIS_CORE_URL from GitHub Variables, alongside AGENT_ANALYSIS_WORKER_URL. Register the existing AGENT_EDGE_TOKEN and AGENT_SERVICE_TOKEN in GitHub Secrets before adopting this workflow; every Gateway deploy now injects these plus the analysis token and URL. Keep matching Edge/Service/analysis-controller tokens in Laravel Cloud separately. Missing selected-deployment settings stop the workflow before remote deployment. Direct CLI deploys must preserve CI-managed URL variables with --keep-vars.
Approval template and shared approval-line creation now offer company-wide or Legal Entity-specific scope with an in-form Legal Entity selector. Changing the approval-line scope clears participant and step configuration for reselection while preserving its name and description. Existing authorization remains enforced; no migration is required.
Signature list pages use the shared organization selector width, and the bulk request list no longer repeats its page heading. Removed fixed percentage widths from the request status and received signature lists so the legal entity column keeps usable space and no longer inflates the header height.
NxRadio supports an optional variant="card" using the existing native radio and group contract. Labels accept descriptive React content; existing default radios are unchanged. Card consumers require the matching Core host and SDK; no data migration is required.