v2.4.0

• Private cross-device vault: Added opt-in WebDAV and S3/MinIO sync with
  per-domain Automerge documents, client-side HKDF/AES-GCM encryption,
  opaque remote object names, encrypted OPFS chunks, device state, recovery
  codes, deterministic conflict handling, transactional local application, and
  no server-side credential or plaintext persistence. ZIP export remains
  version 3 and deliberately excludes sync keys, credentials, baselines, and
  device identity.
• Remote MCP compatibility: Added legacy SSE support alongside Streamable
  HTTP, preferring Streamable HTTP and falling back only when initial connection
  setup returns 404 or 405. The negotiated transport is persisted for later
  tool calls, authentication failures remain visible, and Registry servers that
  require header credentials now collect them before discovery and keep them in
  local encrypted-secret storage.
• Local MCP bridge: Added an optional hardened Docker profile that exposes
  deployment-admin allowlisted stdio MCP servers through authenticated
  Streamable HTTP. Commands and environment variables come only from a
  read-only configuration file; the bridge runs unprivileged with bounded
  output, timeouts, restart backoff, and redacted logs.
• Long-chat reliability: Virtualized message timelines with dynamic row
  measurement, stable end anchoring, screen-proximity rendering for expensive
  diagrams, durable streaming checkpoints, bounded pre-output retries, partial
  output preservation, and guarded continuation after interruptions.
• Message and session workflows: Added reply snapshots and jump navigation,
  explicit interrupted-generation ownership, model-selectable sibling-branch
  regeneration, per-chat composer drafts, and a token/context usage summary
  with a clearly marked estimate when providers omit usage. Message edits,
  deletion, retraction, and branch switches are guarded during generation while
  the existing context-compression pipeline remains intact.
• Agent mode: Added a per-chat, tool-capability-gated client orchestration
  mode with localized web search, knowledge search, text-only skill loading,
  bounded sandboxed JavaScript, and task-plan tools. The five read-only
  built-ins are auto-approved, and Agent web search requires an external search
  provider rather than native Google Search or OpenAI Web Search.
• Parameterized Skills: Upgraded custom Skills to a compatible schema with
  validated text, textarea, and select parameters; reproducible invocation
  metadata; and ordered bundles of up to four non-nested Skills with fixed or
  mapped inputs.
• Knowledge productivity and retrieval: Added collection-level
  Markdown-aware or recursive chunking, preview and explicit reindex controls,
  chunk-level lexical search, vector-plus-keyword reciprocal-rank fusion,
  lexical fallback, and stable source previews with retrieval-method labels.
  Knowledge files can now be filtered by name or status and processed through
  serial batch retry, reindex, download, and confirmed-delete workflows with
  per-file failure reporting.
• Automatic image compression: Added configurable client-side compression
  for supported conversation, workspace, generated, and plugin images before
  storage or model use. The pipeline is cancellable, skips the dimension limit
  for extreme-aspect images, preserves the source format and name, and falls
  back to the original whenever compression fails or does not reduce the
  payload.
• Local offline PWA: Added a local-deployment-only application shell for
  offline history, branch navigation, local search, knowledge reading, and
  backup export. API routes, streams, model traffic, sync, MCP, user files, and
  external requests are never cached; hosted deployments unregister workers and
  remove application caches.
• Navigation, settings, and onboarding: Promoted global search to an
  accessible, focus-managed modal, added localized settings search and a
  first-run path to Provider settings when no model is available, and improved
  mobile panel navigation, default-title localization, collapsed-sidebar
  semantics, and offline draft guidance.
• Provider and model safety: Scoped custom model metadata to each provider
  so same-named models can retain independent capabilities. Hardened locally
  encrypted credentials for server-default providers, bound their reuse to a
  matching provider type, rejected missing or mismatched deployment defaults,
  validated provider Base URLs, and handled empty default-model lists without
  synthesizing unavailable choices.
• Search and presentation fixes: Applied the selected time range to
  Firecrawl instead of forcing a one-week filter, removed duplicate Agent web
  search tool details while retaining the dedicated source presentation, and
  simplified the model picker by removing capability preflight copy while
  preserving the underlying capability gates.
• Compatibility and engineering: Advanced local storage schema to version
  6 while retaining ZIP export version 3, added read-only quota and OPFS
  reference health diagnostics, expanded English, Chinese, and Japanese copy,
  and broadened security, migration, convergence, performance, offline,
  accessibility, provider, MCP, and container regression coverage.