For AI agents: the complete documentation index is available at /llms.txt, the full documentation bundle is available at /llms-full.txt, and this page is available as Markdown at /changelog/2026-10-09.md.
A Task Chats module in the main sidebar: every task conversation you started — tasks assigned to Agents or Crews, quick actions, and conversations opened by an @ mention in a comment — now lives in one sidebar list. Each row shows the task (with its status and type glyphs) or the conversation title, the responsible Agent's avatar, and a second line with the current state, the newest message, and the activity time; clicking a row opens the task detail with that conversation. Conversations with fresh activity rise to the top, ones waiting on you carry a state marker, long-quiet ones fold into More at the end, hovering a row marks it done, and the module title collapses the whole list. Agent messages and waiting-for-reply / waiting-for-review hand-offs in task conversations no longer enter the Inbox — they surface here instead; desktop notifications and Slack reminders still go out (HF-13483).
A new-sync wizard with external accounts: the project sync card on platform integrations gains New sync, a five-step wizard — configure, map statuses, map users, import options, confirm — and every sync link opens a unified detail page with Sync settings / User mapping / Status mapping tabs; changes must be saved, and leaving with unsaved changes asks first. When an import or sync meets a user on the source platform it cannot match to a member, Sharkly now keeps them as an external account — a source identity that owns their tasks — instead of leaving work unassigned. External accounts are managed under People → External accounts, where you can filter and search them and link them to Organization members; linking backfills the assignee on historical tasks (tasks edited by hand are left alone), unlinking restores the external identity, and the members list gains a Linked accounts column. Untouched users in the mapping step now default to import as external account (HF-13070).
The chat input is plain text: the chat composer no longer formats what you type — Markdown syntax, URLs, and pasted code stay exactly as you wrote them, with no auto-linking and no formatting menu. @ mentions, / commands, $ skills, and attachments work exactly as before, and pasting text copied from a rendered page keeps the link URLs instead of silently dropping them. Agent replies still render rich text, code, and media unchanged.
Task conversations can stop auto-opening: Settings → Preferences gains a Task conversations group with two switches — auto-open after assigning and auto-open after using a quick action — both on by default; turning them off only stops the panel from opening by itself, the assignment or action still runs (HF-13528).
Run records show who triggered them: each run now shows Triggered by — the human who set it in motion; runs handed over between Agents trace back to the person who started the chain, external-channel runs show the external account, and truly unattributable runs honestly say System (HF-13658).
Failure messages name the Runtime that failed: when an Agent run fails, the message shows the actual Runtime that went wrong instead of a generic "Agent", making it easier to tell which layer broke (HF-13650).
The Agent status is visible in the chat panel too: the conversation panel header shows the same Agent status component as the task detail, opening straight to reply, view the run log, retry, or stop (HF-13653).
The Agents list inside a Space supports filter / search / pin: matching All Agents, with pins shared — pinning in one place pins in both (HF-13647).
Docs is available to everyone: the Docs library no longer sits behind an early-access switch (HF-13625).
Updates and quitting are more reliable: fixes the restart-after-update stuck on "Starting update…", the macOS Dock Quit not responding and leaving processes behind, and in-progress tasks not resuming after an automatic update (HF-13654, HF-13605, HF-13645).
Opening the Windows app no longer drags the whole machine down: fixes the local-service startup probe spawning large numbers of child processes and exhausting memory on Windows (HF-13627).
Long Windows paths no longer break developer tools: Agent workspace directories are shorter, staying under the Windows 260-character path limit (HF-13445).
Codex / Claude / ZCode CLIs could not start: fixes the three embedded adapters crashing at startup due to their packaging format, which kept tasks from ever starting (HF-13462).
Crew @ mentions no longer drop occasionally: fixes a Crew leader @ mentioning a member in a task conversation sometimes not firing and needing a manual follow-up message; when the other side is busy the run is queued and retried, and failures are reported clearly (HF-13667).
The task composer no longer claims the Agent is offline: the offline warning no longer appears while the Agent is actually online (HF-13344).
A shared computer's settings page recovers: fixes the computer settings page showing an already-shared local machine as "identity pending" and blocking access to its settings (HF-13632).
A failed managed-CLI check no longer reinstalls: when an installed CLI fails detection, the page reports "Detection failed" with a Re-check action instead of silently reinstalling, and error text is no longer garbled (HF-13626).
Imports and syncs follow the intranet proxy computer: fixes some new-sync, mapping, and history-import requests for intranet Jira instances not going through the configured proxy computer, which made loading or importing fail (HF-13395).
Unordered-list bullets no longer render as garbled glyphs (HF-13663).
Interface details: the Inbox search field no longer jumps when expanding and gains a one-click clear button (HF-13686); the task detail loading skeleton is simpler (HF-13648); @ candidate rows prioritize showing the Agent name (HF-13638) and the hover card no longer hides behind the list (HF-13685); the delete button hides when only one working directory remains (HF-13211).
