Release Notes — Week of September 28 to October 5, 2026

Five releases shipped this week — v1.158.0 through v1.162.0 — across 190 merged commits. The through-line: swarms got cheaper and safer at deciding, they run where enterprises already run, and we started publishing evidence of how well they work. A note on scope: 14 commits landed on main after the v1.162.0 tag and ship in the next release; each is marked below.

Highlights

System-one decisions with human review

A native system-one-decision workflow node routes branches through a fast, cheap model call — including via OpenRouter — instead of spinning up a full agent task. Routine routing and triage no longer need an agent in the loop. Provider keys are preflighted before the run starts, so a missing key fails at save time, not mid-run. And when the model's confidence lands inside a band you configure, the answer pauses for human review instead of guessing (#1715). Uncertain calls escalate to people by design, which is the answer to "who watches the agents" that actually ships. Setup walkthrough, including the laya provider: (#1750).

Claude on your gateway, Foundry, Bedrock or Vertex

Claude workers now accept gateway, Microsoft Foundry, AWS Bedrock and Google Vertex routes (#1816). Pair them with the Azure / Foundry embeddings preset and Memory integration card so memory stays in the same cloud as the model (#1836). Enterprise swarms no longer route around their existing cloud commitments.

Azure DevOps integration

A minimal Azure DevOps integration: Azure Repos plus PR service hooks (#1812). A follow-up acks partial payloads instead of returning 500 (#1823).

Live model catalog

Model lists come from a live catalog instead of hard-coded tables, tiers resolve when a task is claimed, and a task whose model the assignee harness cannot run is rejected up front instead of failing mid-run (#1691, #1733, #1764).

Delegate from Claude Code with claude-mod

claude-mod puts the swarm pane, a delegate tool, and live task logs inside Claude Code, so you can fan work out to the swarm without leaving your editor (#1838).

Public agent benchmark

The evals app shipped as a public benchmark: a leaderboard with Pareto frontier and ranking table (#1755), scenario cards and a heatmap (#1759), and an analytics API for picking the best setup (#1744). Swarm scenarios — fan-out, worker recovery, implement-review, capability routing, human-in-loop — run against solo baselines (#1745, #1756), with reasoning effort configurable per harness and model (#1710) and refusal-gated publishing (#1761).

Improvements

Workflows

• Standard-webhooks verification — webhook triggers verify webhook-id/webhook-timestamp/webhook-signature headers against the whsec_ signing secret (#1760).
• Paged workflow runs — GET /api/workflows/{id}/runs is paged and list rows drop run context (#1699).

Memory

• Logical paths — memories move and filter under /longterm via key, newKey, keyPrefix and path weight (#1814).
• Longterm key tree and whole-memory view on /memory — on main after v1.162.0, ships in the next release (#1856).
• Count-based retention for context_versions (#1808).

Slack and extensions

• Native Slack working status for the whole life of an ask (#1770).
• Opt-in inline extension install for leads and operators (#1828).
• Deploy-awareness catalog extension (#1826).

Harnesses and models

• pi — opt-in codemode (#1731) with a PI_CODEMODE_MODELS flag (#1824), installed MCP servers through pi's MCP extension (#1730), and non-core tools deferred behind tool_search (#1729).
• Codex CLI 0.159.0 with GPT-6.1 Sol (#1719).
• OpenCode emits assistant text so outputSchema validation works (#1753).
• Auth resilience — seat-aware Claude OAuth key selection (#1765) and Codex pool login benching after 2 auth failures (#1763).

Dashboard

• Contextual session panel on every page (#1688).
• agent-fs review space in the dashboard, off by default (#1728).
• Reference icon set for task status — on main after v1.162.0, ships in the next release (#1860).
• Dev mode for version gates (#1726).

Telemetry

• schema_version 2 identity envelope (#1780) and dashboard events honor the telemetry opt-out (#1773).

Docs

• Multi-agent orchestration patterns guide (#1864) and team collaboration feature page (#1865) — both on main after v1.162.0, ship in the next release.
• Workflow example configs and routing corrected (#1862).

Bug Fixes

API performance and event-loop stalls

• Slim task lists — SQL projection, opt-in total and a timeline shape for GET /api/tasks (#1697); opt-in slim list rows for schedules and approvals (#1703).
• Batched lookups — session steering and citation lookups (#1704) and profile-sync banner lookups (#1701); agent status counts now run in SQL (#1716).
• Event-loop stalls stopped — KV range scans (#1690) and chunked app sync (#1696).
• SQLITE_BUSY — the final attempt is capped at 250ms with stall reporting (#1707); pricing and RBAC seed once per DB handle (#1695).

Runner and shutdown

• Graceful drain — the API drains on SIGTERM so workers hand off while it still serves (#1837).
• Stuck slots — worker slot freed when a provider session never settles (#1820).
• Output recovery — fenced JSON recovered for schema-bound tasks (#1833); pi reprompts once on an empty final turn (#1831); Claude task body sent once (#1815).
• MCP and workflows — unknown MCP session ids return 404 so clients reconnect (#1822); node.retry honored when an agent-task step fails (#1832).

Memory retrieval

• One relevance scale — every retrieval arm lands on one [0,1] scale with relevance-gated injection (#1689).
• Recall task content instead of worker wrappers (#1740); rater model pinned and recorded per rating (#1806).
• On main after v1.162.0 — stale chunks removed on keyed delete (#1857) and key-only lookup resolves one visible document (#1861).

Tasks and heartbeat

• Superseded tasks — dependents re-pointed to the resume (#1713, #1664); followUpConfig inheritance scoped to continuations (#1680).
• Heartbeat — flag parsing and interval 0 honored (#1711); fresh installs no longer point at unseeded scripts (#1714).

Breaking Changes

This week tightens behavior rather than removing APIs — these are requests that may now be refused:

• Task model validation — a task whose model the assignee harness cannot run is rejected at creation (#1764). In v1.159.0+.
• Task write ownership — store-progress refuses writes to another agent's task (#1737). In v1.159.0+.
• RBAC, on main after v1.162.0 — stdio MCP servers at agent scope (#1851), changing a repo's allowMerge (#1852, #1859) and attributing a task to another user (#1853) now require a lead, operator or user. HTTP memory delete is gated and list is narrowed for agent callers (#1855).