Repository navigation
@agent-native/core@0.155.0
Minor Changes
- 89f194f: Support durable, scoped GitHub sources and replayable sync for Builder design-system imports.
Patch Changes
-
89f194f: Action routes now fall back to the caller's stored active organization when a
cookie session resolves no org, matching what the adapter/A2A path already did.
An empty org silently narrowed every scoped read to rows with a nullorg_id,
so a user could stop seeing their own org-scoped dashboards and resources. An
explicit Personal selection still resolves to no org, and a transient database
failure propagates instead of reading as "this user has no org". -
89f194f: Add
motiontoBrandKitTokenTypeso durations, easings, and transitions have
a real category. Extractors drop tokens they cannot classify, so the missing
bucket meant no imported design system ever carried its motion. -
89f194f: Open agent chat image attachments in a full-size lightbox when their thumbnails are clicked.
-
89f194f: Keep the chat-first New chat action icon-only when the desktop rail is collapsed.
-
89f194f: Simplify share controls with compact copy-link rows and always-visible access details.
-
89f194f: Ensure durable resource instructions remain in lean agent prompts and fail loudly when AGENTS.md cannot be read.
-
89f194f: Hide
x-cloakcontent in the extension iframe shell until Alpine boots.
Extension content is a body snippet, so it cannot define the rule itself: an
x-cloakoverlay painted over the whole extension until the deferred Alpine
CDN script resolved, and permanently when it failed to. -
89f194f: Keep feedback and other sibling overlays open when launched from the Agent panel overflow menu.
-
89f194f: Repair and validate native SQLite bindings against the Node runtime used by development, builds, and production starts.
-
89f194f: Convert bare Slack user IDs in outbound agent responses into native mentions.
-
89f194f: Add folder-backed agent packs with safe Claude/Cowork-style import, agent-owned
references and skills, and a shared Factory Agents surface for managing simple
agents alongside mounted agentic apps. -
89f194f: Grayscale non-selected workspace app icons so the active app is easier to identify in the chat-first rail.
-
89f194f: Add an optional
submitContextto the guided-questions payload, appended to the
context of whichever message the card sends. A question card's answer opens a
continuation turn that inherits nothing from the turn that posed it, so context
the follow-up work depends on had no way to survive the hop. -
89f194f: Detect active organization-scoped feature-flag rollouts for anonymous Desktop discovery while keeping authenticated authorization scoped to the user's email and organization.
-
89f194f: Preserve signed Builder callback state when mounted route events normalize away the raw query string.
-
89f194f: Provision cross-app SSO state and authorization-code tables during release migrations so production serverless requests never perform schema DDL.
-
89f194f: Keep approval continuations out of visible chat history and keep approval controls usable at narrow widths.
-
89f194f: Continue chat turns when a background tool completes before the assistant sends its final response.
-
89f194f: Add a safe inline Markdown renderer for compact user-authored text surfaces.
-
89f194f: Repair existing workspace app skill copies during scaffold updates by linking them to the shared workspace skill surface.
-
89f194f: Chat no longer renders the same assistant turn twice — the long-standing report
of a final message streaming in two places at once and tool outputs appearing
more than once. Four independent causes, all of which let one run be folded into
UI state more than once:- SSE resume cursors were kept in a single browser-wide slot, and
updateActiveRunSeqtook no run identity, so it wrote the caller's sequence
into whichever run happened to occupy the slot. With chats streaming in
parallel (agent teams, multiple tabs) runs evicted each other, and
resolveReconnectAfterSeqthen returned 0 — replaying an entire run on top of
history that already contained it. Cursors are now stored per{threadId, runId}, identity is required to advance one, and a cursor outlives its run
losing focus so a later reconnect resumes instead of replaying. - The adapter's stream and the reconnect reader could both fold one run at once.
Ownership was a React ref re-checked by a 1s poll that is skipped while the tab
is hidden, and the refs were per-component-instance while several chat
instances mount against one run. Ownership now lives in a module-scoped
registry claimed and checked synchronously, and the adapter preempts the
reconnect fallback when it takes over. - The reconnect overlay was deliberately kept mounted beside the live message
list for up to 2500ms after handoff, leaving two independent folds of the same
turn on screen with only content-similarity heuristics hiding the second. The
overlay now renders only while no runtime owns the turn. - The server fold pushed a tool card for every
tool_start, including the
replays that journal and zombie-ledger recovery emit for calls that already
ran. The live client coalesced those onto the original card, so a duplicate
tool output appeared only after a reload. A replayedtool_startnow folds
onto its existing card.
- SSE resume cursors were kept in a single browser-wide slot, and
-
89f194f: Prevent assistant panel crashes when dense chat replays synchronously update React.
-
89f194f: Keep nested exception debugging context out of Amplitude event properties while preserving it for internal error tracking.
-
Updated dependencies [89f194f]
- @agent-native/toolkit@0.14.3