Release Notes — Week of September 21 to September 28, 2026

This week the swarm gets easier to start, safer to extend, and cheaper to run. A fresh install now reaches a working swarm from the browser, extensions package real automation without arbitrary-code installs, and every credential shows what it costs.

Highlights

Set up a swarm from the browser at /setup

A new install no longer starts with an env file. The first-run onboarding flow at /setup takes you from an empty deployment to your first task in one guided pass:

• Live API state checks. Each step confirms the API is reachable and configured before you move on, so misconfiguration surfaces at setup instead of on the first failed task.
• Codex device login. Sign in with a device code instead of pasting keys into config.
• Model dial. Pick where each agent sits between cost and capability.
• Integrations, memory and a first task. Connect Slack, GitHub, Linear, Jira and other integrations, name your swarm, then send a first message to your agents from the same flow.
• Telemetry stays on by default. Anonymized telemetry is on unless you opt out with ANONYMIZED_TELEMETRY=false (or in Settings → Configuration). Setup has no telemetry toggle. See the telemetry reference for what is sent.

For solopreneurs this cuts time-to-first-task to minutes. For CTOs rolling the swarm out to a team, it removes the setup handoff. #1604

Extensions ship scripts and schedules, from a trusted catalog

Extensions can now bundle scripts, schedules, skills and workflows alongside their hooks, so one install delivers a complete piece of automation. Installs come only from the predefined catalog, and every manifest (YAML or JSON) is validated against a published JSON Schema before it lands. Agents can also author and own inert extension drafts that do nothing until they are installed, so an agent can propose automation without executing it. #1595, #1567

See what every credential costs

• Subscription plan per credential. Tag each API key or OAuth credential with its plan, and read spend against it on a reworked usage page with daily spend, ranked spend and subscription cards. Usage queries are faster thanks to a new covering index. #1617
• Model-scoped weekly windows in the Claude OAuth pool. The pool now tracks weekly limits per model. Exhausting one model no longer benches the credential for every other model, so more of your paid capacity stays usable. #1609

New models and providers

• Opus 5.5 is available and is now the Opus default. #1575
• GPT-6 Sol and Luna join the model registries. #1580
• DeepSeek is available through the new dsh harness provider. #1570

Improvements

• Dynamic human-in-the-loop questions. Workflow approval nodes now build their questions from upstream output, so reviewers are asked about the actual result instead of static text. #1611
• Accurate, visible task citations. Citations render on the task page and in Slack outcome cards, and Slack input attachments are no longer counted as sources. #1607, #1594
• Typed configuration controls. Configuration values are edited with toggles, selects and number inputs instead of raw strings. #1587
• UI audit closed, mobile included. The new surfaces (sessions, approvals, schedules, usage and setup) work on a phone, and iOS Return now inserts a newline in the session prompt instead of submitting it. #1623, #1624, #1627, #1628
• Richer session view. Sessions show delegated progress and follow-up output, model labels resolve from the live catalog, and lazy inline images render in task log results. #1598, #1582, #1564
• Hive field Open Graph card. Shared dashboard links now unfurl with a proper preview card. #1591
• Slack transport groundwork. Transport selection is in place while socket ingress keeps working unchanged. #1543
• Tighter daily digests. Scheduled daily digests stay brief and actionable. #1581

Bug Fixes

Slack

• A deferred task now renders as waiting and closes its card in place, and the internal wake-up time is no longer posted to the thread. #1569, #1615
• Outcome delivery give-ups are persisted and orphaned streams stop, so an abandoned delivery stays abandoned instead of leaving a stream open. #1618
• Buffered steering keeps human attribution, and steering messages name their sender in agent messages and the UI. #1592, #1590
• The outcome card's status icon sits on the same line as its text, and the bot's own mention resolves correctly in DM task text. #1586, #1562

Scheduling and tasks

• Spent defer-task wake-up schedules are reaped after 30 days instead of accumulating. #1629
• A settlement that wakes a deferred waiter skips the redundant lead follow-up. #1620

Workflows, routing and workers

• Workflow heartbeat recovery no longer touches live graph walks, and interrupted recovery preserves selected ports. #1584, #1600
• Capability-only routing affinity now matches on capabilities. #1602
• Extension identities are hidden from agent listings and cannot be assigned tasks. #1583
• Workers verify the ACP target binary at boot, so a missing binary fails fast instead of mid-task. #1619
• Memory preset defaults apply at runtime, and worker Claude settings are hardened. #1585, #1608

Migration Notes

There are no breaking API changes this week. These default shifts are worth knowing before you upgrade:

• Extensions install only from the predefined catalog. Ad-hoc extension installs are no longer accepted. #1595
• New model defaults. The Opus default moves to Opus 5.5, and the retired Codex default is replaced (Pi bumped to 0.87.1). Pin a model explicitly if you need the previous behavior. #1575, #1579
• Memory raters default. When MEMORY_RATERS is unset, it now defaults to implicit-citation,explicit-self. #1588