Pattern Automation
← All releases

v2026.916.0 Breaking

Breaking Changes

  • Cheap model profiles removed — The second execution mode for cheaper recovery/override models is gone from adapters, agents, APIs, and the UI. Migration 0236 strips stored modelProfiles.
  • Agent APIs no longer return plaintext credentials — Detail, list, and lifecycle responses redact adapterConfig.env plain bindings. Scraping live keys from those responses stops working.
  • Automatic productivity reviews retired — The detector that manufactured management-review tasks from run/comment counts is deleted. Bounded recovery, budgets, blockers, and normal review stages remain.
  • X-Forwarded-Host only from trusted proxies — Forwarded host is honored only when the peer passes TRUST_PROXY. Confirm proxy trust on reverse-proxied installs.
  • Anthropic legacy REST connection removed — Obsolete REST setup is gone in favor of the AI-account Connections flow; unsupported REST connections receive accurate rejection.

Highlights

  • AI credentials live in Connections — Claude and Codex subscriptions, API keys, and shared accounts become managed Connections with grants and permissions. Agents reuse the responsible user’s account; missing accounts pause the task with an inline connect form.
  • Agents get their own email — Experimental AgentMail connections give an agent a dedicated inbox: inbound mail becomes assigned work; sends are explicit authenticated actions.
  • Reach agents from chat tools — Experimental native connectors bind Slack, Discord, Telegram, Microsoft Teams, or GitHub bots to one agent and one task per conversation. iMessage joins via Photon. In-app Agent Chat gives every person a persistent governed conversation.
  • GitHub access becomes identities — GitHub App-backed identities with durable refresh replace shared tokens. When several people steer one agent, tools resolve to the responsible person’s credentials.
  • Announcements in the app — Native announcement cards from a validated JSON feed, with per-account dismissals. On by default; see Upgrade Guide to opt out.
  • Native Runner does the whole job — Experimental Neuro OS Runner becomes a complete execution engine. enableNativeRunner defaults on for self-hosted instances (explicitly configured agents only; cloud stays off).

Improvements

  • Apps and connector catalog — Apps leaves its experimental gate; managed Google access works on clean self-hosted installs; agents can request new service connections from task feeds.
  • Model updates — Claude Fable 5.1, Codex GPT-6 Astra, Grok sandbox device-login, and Opus 5 as the unset Claude default.
  • Task surface refresh — Richer work-product cards, Stop control, condensed runner activity, unified search, and a foreground-only test-drive CLI.

Fixes

  • Database pool heals — Idle connections are reaped; failed socket writes on recycled connections are replayed instead of wedging the pool.
  • Onboarding and connections — Provider sign-in shows verification progress; concurrent runs share one Anthropic/OpenAI/xAI subscription connection.
  • Recovery and sandboxes — Hidden issues excluded from stranded recovery; messages sent while stopping are delivered after recovery; lost sandbox control fails the turn fast.

Upgrade Guide

  • Node.js 24.11.0 floor from v2026.831.0 is unchanged.
  • Migrations 0231–0279 run on upgrade (connections, GitHub identities, chat, AgentMail, announcements, and more).
  • Optional: SENTRY_DSN_FRONTEND / SENTRY_DSN_BACKEND, NEURO_ANNOUNCEMENTS_ENABLED, NEURO_ANNOUNCEMENTS_FEED_URL, NEURO_TOKEN_BROKER_ALLOWED_HOSTS, NEURO_DISABLE_CWD_ENV_FILE.
  • Note: enableNativeRunner defaults on for self-hosted; new standard-trust agents can hire by default; Apps is no longer experimental.