Version 2.0.0-alpha.49

• Centralize 1.x Worker provider adoption and pending Automation script imports
  in dedicated upgrade modules; keep legacy fields out of runtime Job resources.
  Recover imported Jobs with the bounded Workspace rebuild when native thread
  files are missing, preserving queryable history and removing transcript replay.

• Recover knowledge bodies and property edits together after reloads, including
  invalid YAML, and preserve new input when an earlier save finishes. Display
  Page references inside business properties as links, report malformed knowledge
  panels, and keep graph and backlink titles in sync with confirmed Page updates.

• Fix nested transactions in Console promotion and merge reviews so review decisions commit and publish their changes.

• Remove unused Brain panel types and retired internal read/serialization helpers; use one Drive path contract for reads and writes.

• Read immutable Applet packages outside Page-lock transactions, batch Worker
  provider counts, and use deterministic tool-result journal keys. Remove unused
  VFS message handlers and coalesce the Worker's initial Drive refresh.

• Repair Skill lesson review and UUID-based memory retrieval, preserve forgetting
  provenance, and accept document search results without an entity subject.
  Publish scheduled, webhook, background Job and Workflow inputs immediately
  after commit. Recover Worker RPC credit on failed responses and report replay
  owner failures instead of leaving callers waiting.

• Correct Plugin restart dependencies, credential cache identities, binary
  attachment downloads, IMAP IDLE handling and fragmented HTTP body limits.
  Preserve reasoning spaces and completion labels, isolate WhatsApp webhook
  verification by matching credentials, reduce repeated cache and database work,
  and enforce tenant foreign keys on the Meta tables.

• Add published Applet Actions with caller and owner execution identities,
  persistent cursors, serial queues, coalescing and resumable slices. Route
  schedules, webhooks and Base changes through shared Signal Bindings and
  receipts, preserving historical deliveries. Add Base business-key upserts and
  suppress unchanged writes; provide streaming HTTP file transfers and isolate
  foreground and background Function Runtime capacity within the 2 GiB deployment budget.
  Let Applet frontends follow durable Run progress and recover the latest result
  after reconnecting through useActionRun. Retry Base deliveries independently
  by tenant and table, retain the event size limit on the production Run path,
  and reserve foreground capacity for PTC while honoring the total concurrency limit.
  Keep filtered Base deliveries in their correct concurrency groups and drain
  large change backlogs promptly. Wake queued Actions when a publication enables
  parallel execution, wait for busy runtimes, and distinguish completed function
  errors from interrupted attempts when deciding whether to retry.
  Allow tenant deletion to finish when an Agent owns published Applets.
  Fix Agent Applet publication lock order while keeping Action policies and
  files atomic and rejecting writes from completed runs.
  Preserve failed cron history when the final delivery attempt is rejected at admission.
  Hand off slices only after the current handler finishes, preserve event IDs and
  Human authorization on Applet emissions, and keep the original caller through
  owner-authorized child calls. Batch historical delivery migration and receipt
  cleanup, and reuse MCP tool discovery within each function invocation.

• Add Meta Ankole with a collapsed navigation tree, a Brain graph, editable Skill
  packages and per-Agent/global AGENTS and DESIGN contracts. Separate knowledge
  from collaboration roots and move Skills and contracts to independent tables
  with revision checks. Preserve positive Skill bindings under global enablement.
  Project Brain read-only by conversation disclosure and cache each Agent's
  read-only Skills and contracts across conversations; retain admitted Skill
  selection for background Jobs. Migrate existing packages, contracts and grants.
  Let members configure their assistant before starting a conversation, search
  Skills, import package metadata and edit attachments directly from the file
  list. Show effective Agent instructions and save feedback, with scrollable
  workspaces on desktop and narrow screens. Keep Worker
  connections alive during Page tree changes and use the internal Worker origin
  for Drive mounts.
  Persist Markdown supplied by Agent document creation so delivered links open
  the requested content.
  Isolate Skill drafts across package navigation and keep literal Agent UIDs
  separate from assistant and global settings routes. Read package directories
  without repeatedly scanning the capability catalog. Reject conflicting package
  file paths and preserve distinct legacy knowledge bodies and attachments during
  migration. Give each Agent list entry a separate, full-width click target.
  Correct Skill ZIP import tests to distinguish backend file limits from Drive VFS policy.
  Keep individual Skill bindings independent of package revisions and let editors
  reconcile conflicts without discarding drafts. Show separate global, colleague
  and assistant controls without creating an assistant during reads. Complete
  Brain favorites, moves, trash and semantic links, and derive object types from
  ontology categories. Handle archived legacy content, orphan contracts and slug
  collisions during migration. Refresh live contract files in long-running FUSE
  mounts and enforce admitted Job Skills across every runtime read path.
  Support interactive Workers without FUSE by materializing enabled Skill packages.
  Preserve installed package directories, manifest attachments, explicit enablement
  and inherited Agent defaults during direct upgrades. Remove the legacy Skill
  defaults setting, manage runnable packages in Meta Ankole, and grant members
  custom-Skill editing and deletion. Keep generated pattern pages inside Brain
  with their source's visibility, and load graph summaries without document bodies.
  Retire active lessons when deleting a Skill and complete Skill renames without
  reopening the discarded draft or requesting the old address.

• Restore Page as the source of truth for identity, content, subpages and access,
  and remove the folder Page type. Drive projects empty Pages as directories
  and populated Pages as body files with optional child directories. Preserve
  Page IDs, permissions, history and attachments during migration; clearing a
  body keeps its children. Require parent edit permission to create subpages.
  Use the Ankole logo for AnkoleBrain and a distinct book icon for knowledge Pages.
  Preserve empty Page subtrees during structured moves, allow editors to remove
  empty directories, and keep ordinary ChatDrive filenames writable. Delete
  Page-owned Actions in lock order when removing conversations or Agents.

• Hide unavailable Page actions across the canvas, sidebar, and child lists.
  System knowledge Pages display fixed titles and omit rename, move, reorder,
  and trash controls while keeping authorized content editing and child creation.

• Improve Brain memory selection so routine task completion, delivery and
  validation messages do not become facts or empty entity pages. Preserve
  substantive research conclusions, business outcomes and reusable knowledge
  when they appear alongside execution messages.

• Update the kernel's Rust dependencies to the latest releases that still build
  on Rust 1.91, including crc32fast 1.5.4. Update control-plane Elixir
  dependencies to llm_db 2026.11.0, Localize 1.6.1, attesto 2.10.0,
  attesto_client 2.13.0, and codepagex 0.3.0. The opaque Principal subject
  patch now applies to attesto 2.10.0.

• Fix native kernel stream completion and input boundaries: completed or failed
  AI streams no longer emit a later abort event, token truncation avoids excessive
  shortening across BPE merges, and OpenAI Chat records actual completion
  times. Provider numeric fields saturate instead of overflowing, JWT configs
  reject incorrectly typed fields, batch AuthZ rejects empty resources or actions,
  HTTP specs require HTTP(S), zero-capacity caches stay empty, and WebDAV requests
  are capped at 1 MiB and 128 XML levels.