agent-swarm.devagent-swarm.dev
Releases

Release Notes — Week of November 2 to November 9, 2026

Fleet savings land at /usage/savings, subscription quota meters join Usage, json-render cards reach task output and Slack, and apps gain live session threads, across 219 commits from v2.8.0 to v2.9.0.

The week ran from v2.8.0 to v2.9.0 on 219 commits: 41 features, 58 fixes and 21 docs in the audit window. The through-line: the swarm now prices its own automation, so "does the fleet pay for itself?" answers from your own data. Two changes need a config edit.

Highlights

See what the fleet saves

#2779 adds a fleet savings page at /usage/savings: what your scripts, schedules and workflows did, priced as what agent sessions would have cost instead. An agent would have spent a session cost at the run's tier, scaled by a failure-and-retry multiplier plus a lead review charge. Only deterministic glue is substituted; agent-task nodes keep their actual cost.

Savings come at four tiers (smol, regular, smart, ultra) as a band, not a point estimate. Constants were calibrated on 2026-11-05 against claude-opus-5-5 costs, and the curve recalibrates from your own sessions (#2770).

  • A workflow shows its saving in the header, a Savings tab and a daily chart (#2768); scripts and schedules show theirs on the page, rail and list (#2776, #2777).
  • A daily rollup backs a Saved (30d) column, and workflow totals break into components (#2772, #2780).
  • The fleet summary answers in a handful of statements, not a scan per row (#2809). Docs: Savings API.

Quota meters and provider limits on Usage

Usage now shows what is left before a provider throttles you: quota meters (#2890) beside current limits and observed window usage (#2870). Home gained savings, dismissible failures and a new top bar (#2866).

Cards instead of walls of text

A task result can carry a json-render UI spec, so tables and charts render as a card, not a markdown dump (#2876). Cards also post to Slack and Chat SDK (#2844); prompts steer small chat tables into one (#2860).

Apps that run a session, not just a query

SessionThread renders a live session inside an app, and an app.action step can start one (#2760, #2889). PageFrame brings a Chart, a Metric sparkline and a sidebar shell (#2884, #2769), and a catalog batch ships with approval, agent-fs files, a split drawer and a dialog (#2810). Docs: Apps.

Onboarding: Work, Connect, Build

#2734 is the week's largest change at 421 files: setup is three named groups, Work, Connect and Build, with in-app files, use-case groups and a Home page that shows what the swarm is doing. Docs: Getting started.

Approvals answered where they are asked

A session's approval is answered in the dashboard, in place (#2851). The Slack card flips to resolved where it was posted and names the responder (#2840, #2847); pending approvals sort above live activity (#2864). Docs: HITL gates.

Improvements

  • Models — a profile can list same-harness fallbacks, so a task ending on a rate limit or an empty key pool hops to the next (#2736). check-model-profile validates a profile; credentials gain an allowlist (#2219). Reasoning text is captured for Claude, Codex and pi (#2869, #2893). Docs: harness providers.
  • Workflows — each run is pinned to the definition it started with, so edits no longer change runs in flight (#2727), and cleanup is scoped to deleted runs (#2740). Docs: workflows.
  • Slack — multi-person threads name every participant with a trust level, and unknown outsiders are untrusted (#2767); author lines cannot be forged (#2771); SLACK_MODE=http runs on a native signed receiver (#2857). Docs: Slack.
  • Integrations and runtime — GitHub Enterprise Server via GITHUB_URL (#2808); a sql-admin-only template gates db-query by task creator role (#2824); opt-in import via SWARM_IMPORT_SOURCES (#2805); per-kind flags skip built-in seeds (#2800); Docker isolation keys for runtime pools (#2881).
  • Scripts and performance — a stdlib retry(fn, opts) helper plus fetch and bridge retries on 429 and transient 5xx (#2790); a selective, load-paced boot re-scrub (#2811); db-query reuses worker children (#2813). Docs: scripts runtime.
  • UI and files — files and Pages preview in a side panel (#2709); icons and avatars come from a link or a file (#2729); markdown renders agent-fs style everywhere (#2765, #2766); a task links to a registered repo (#2804); open views refresh entity versions (#2712); phones get a native header and action sheets (#2735, #2863).
  • Docs — six guides, from choosing an operating model and the small-company workflows atlas to worker recovery and local TEI embeddings (#2722, #2724, #2742, #2744, #2723, #2751).

Bug Fixes

  • Slack — buffered thread replies persist across restarts (#2861) and fold into the pending Lead task (#2849); block dedupe matches word boundaries by code point (#2859); inbound messages are identified in retry logs (#2855); a reload during boot start no longer 500s (#2873); a human reply no longer waits on worker routing (#2628).
  • Config — invalid values fall back to defaults for the heartbeat stall threshold and resume limit (#2838, #2837), the stale threshold (#2821), the eligibility scan cap (#2836), the pool scan batch (#2807) and the Slack tree stall threshold (#2817). Non-finite HEARTBEAT_INTERVAL_MS is rejected (#2721), circuit-breaker limits are validated (#2730), and heartbeat memory is bounded (#2875).
  • Savings and workflows — pre-307 sessions attribute to their task credential (#2898); savings stay available before the first rollup (#2867); a superseded step task is marked once its retry queues (#2892); a native run ends when its default export resolves (#2737); template updates land on workflows whose only change is a setting (#2757); keyed write retries survive a slow first attempt (#2820); sql-admin-only allows db-query when the source task has no row (#2897).
  • Security — feedback endpoints reject URLs that carry credentials (#2755); a structural rule match inside a vendor secret redacts the whole secret (#2783); the gitleaks true|false|null allowlist is anchored as a whole (#2796); ambiguous stored keys are rejected (#2673).
  • Providers and skills — provider dependencies are validated at task start (#2858); opencode MCP servers use the local command-array shape (#2793); OpenRouter gateway model ids get a catalog entry (#2784); entrypoint checks moved to Bun (#2823); YAML block scalar skill descriptions parse and keep their metadata (#2763, #2764).
  • UI polish — task cards render in the link-preview split screen (#2887); Files matches agent-fs reading width (#2854); agent files show with the reply (#2852); the model column is readable with a details card (#2726); toasts keep a consistent offset (#2711, #2718); a page summary opens by tap and keyboard (#2716); chart axis labels stay apart on phones (#2871); 0.0.0.0 counts as unreachable (#2754); chat splits no longer loop on a surrogate pair (#2753).
  • Tests — savings day slices, the script-route clock and the fleet read are pinned (#2814, #2822).

The other 40 fixes are smaller repairs across harness adapters, scheduler and dashboard.

Breaking Changes

No API removals, but two behavior changes:

  • context-mode is gone from the worker image and every harness adapter (#2812). CONTEXT_MODE_DISABLED and CONTEXT_MODE_EXTERNAL_MCP_NUDGE_EVERY are now ignored. Action: drop both variables, and install the plugin in your own image if you need it.
  • Starting Linear or Jira OAuth now requires admin auth (#2756). A non-admin start is refused. Action: have an admin start the tracker OAuth flow.

On this page