release notes

Smart Node in October 2026: Chat, Delivery, and Owner Controls

During October 2026, Smart Node shipped three releases: 1.12.49 on 4 October, 1.12.50 on 12 October, and 1.12.51 on 16 October. A fourth release, 1.12.52, arrived on 16 October and is covered separately in Smart Node 1.12.52.

These updates turn the office chat into an installable messenger, stop messages from disappearing when services drop, and give the business owner direct controls over roles, spending, and secrets.

The office chat works like a messenger

Office chat now works like a standard messaging tool on both desktop and mobile screens. It includes author tags, replies, quotes, unread counts, and @mentions. When you type an @name in chat, that specific agent wakes up and answers directly in the conversation. The office interface can also be installed as an application on a phone or computer.

Longer agent tasks no longer break the chat view. Previously, any turn that took longer than two minutes was marked as failed even while the agent continued working, and a slow task held up other chats. Now, messages receive their answers regardless of how long the agent works. If you send new messages while an agent is busy, they are passed into the current turn directly or queued for the next turn. The reply window shows what the agent is doing in real time, and clicking that status displays the commands run during the turn. Any message can be copied as plain text or Markdown.

Task cards in the office interface also include a comment box. You can ask about a task there, and using an @name alerts that person or agent.

Messages reach the agent, and answers reach you

Several gaps where incoming or outgoing messages could be dropped were resolved:

  • Messages sent to a stopped agent are held until the agent starts again. Senders reaching out through email or Telegram receive a single notice that the assistant is paused.
  • When a delivery channel drops out due to a timeout, a severed email connection, or a Telegram error, the node retries five times over roughly 15 minutes before reporting a failure to the agent.
  • If an agent leaves a message unread, the message is redelivered up to three times.
  • When an agent requests information from a peer agent, the node sends reminders. If no reply arrives, the asking agent is notified.
  • Agents copied on the same conversation thread no longer reply back and forth endlessly, as an agent can now conclude its turn with silence.
  • Provider errors are no longer recorded as valid answers, and restarting the service that wakes agents no longer leaves an agent marked busy for 30 minutes.
  • If the database restarts, dependent services now restart alongside it automatically.

Owner controls, roles, and secret storage

Every action on the node now requires a role check. Accounts and agents belong to one of three roles: super-admin (the business owner), admin, or user. System agents, such as setup and sysadmin, hold admin status. The setup agent can create and configure other agents without owner intervention, but agent contacts can only be edited by the owner.

All credentials now live in a single secret store. The Secrets page details where each secret is used, who may read it, who created it, and who accessed it last. It also records blocked read attempts, accompanied by a Grant button. Model credentials are kept out of the agent processes entirely: every model request routes through the node gateway, where usage is logged per agent.

The owner panel includes controls to halt any agent immediately, inspect current agent operations, disable specific workflows, and choose which agents are permitted to start them. Owners can also configure spending caps, set unattended wake limits, define turn length limits, or pause the entire node. A single daily summary email reports any hard failures across the node.

For critical decisions, owners can enable an optional module that asks a person to decide and sign the answer with a passkey.

Tasks, workflows, and server resources

Planner tasks can now store file attachments and carry three simple statuses: queued, in progress, and closed. When an agent is called to handle a task, it receives a wake message containing the task name and the reason. If a task goes quiet, the planner prompts the assigned agent, notifying the owner once if the reminders go unanswered.

Workflows now advance tasks from block to block directly. If a workflow block fails, it routes the task to a person along with the failure reason. Workflow blocks can also be assigned a time budget, preventing a workflow from waiting indefinitely for an agent response.

On the server, the documents service now runs directly on the host rather than inside two container images. Its install step now takes about 32 seconds instead of about 100, and a node uses about 510 MB less. In addition, operator settings for the node are now stored in an isolated location so they remain intact through software updates.

Runtime options and open items

Agents can now run on Flint as an alternative runtime beside Claude. Support for a second Anthropic-compatible provider can also be configured with a single configuration row without code changes.

The release notes also said what was still open when they shipped:

  • In 1.12.49, the default turn limit was ten minutes unless the owner raised it for an agent.
  • In 1.12.49, an admin who was not part of a chat thread could not join it yet.
  • Cancelling a queued chat message was newly introduced in 1.12.50 and had only been verified in tests.

Notes and history

Detailed technical entries for each release are published on the changelog. Previous monthly summaries are available in Smart Node in September and Smart Node in August 2026.