Release Notes: uxopian-ai v2026.1.0
Release Date: November 2026 Version: 2026.1.0
🚀 2026.1.0 is the first minor release after general availability. Plans run on a new engine built to scale across replicas, the review queue grows the tools a review team asks for every day, and one gateway can now serve several clients. And this time, the upgrade is in place: your data stays where it is.
⚠️ REST integrators of plans and review tasks: read the Upgrade notes. Plan executions became plan runs, under a new path, and the review export now has one line per task.
👀 At a glance
| What's new | In one sentence | |
|---|---|---|
| 🧩 | New plan engine | Runs spread over replicas, sub-plans become child runs, and a failed list element no longer sinks the node. |
| 🛑 | Plan limits | Token budgets, run validity, unique runs and an output type, set per plan. |
| ✅ | A review queue for teams | Date filters, archive, delete, Undo all, shareable links, and one CSV line per document. |
| 🌐 | Per-client routes | One gateway, several clients: routes match on the host and carry their own FlowerDocs and ARender. |
| 🔑 | Login redirect | A browser that is not logged in goes to the login page instead of a bare 401. |
| 📂 | Shared folder for ARender | ARender reads the review documents from a mount instead of receiving an upload. |
| ⬆️ | In-place upgrade | Same OpenSearch, same data: new fields are added at startup. |
🌟 Highlights
🧩 A plan engine that scales
Plan runs are now shared between the replicas of a deployment: any replica can run any node, up to 1024 at once each, and signals travel through Hazelcast. You can even split replicas into brokers and workers. Sub-plans and plans called as tools become child runs that you can open and stop like any other. A fan-out whose elements partly fail now ends Partial with what succeeded, and runs 32 elements at a time instead of 8. See Agentic plans.
🛑 Plans that know their limits
A plan can now cap its tokens per day or per month, set how long a run may last (24 hours by default), refuse a second run with the same inputs while one is active, and declare the type of its output. Each node is bounded in time, output size and tokens. When a run starts, the agents, prompts and providers it uses are frozen: editing them never changes a run already started.
🖼️ A plan page made for building
The flow takes the whole page. Add agents, tools and sub-plans from a floating toolbar, edit a node in a floating inspector (agent, prompt and code included), set the conditions on the links, and save everything at once. See Managing plans.
✅ A review queue for teams
The queue gets a Dates panel (created or reviewed, to the minute), archive and delete, a Being reviewed status, accent-insensitive search, tags combined with or, and Created and Last modified columns. In a task, Undo all puts every decision back, an empty value shows an Empty badge, and leaving the task hands it back at once. Filters and the open task live in the address, so a link opens the same view. The CSV now has one line per document, with the extracted value, the final value and the status of each field. See Human review.
🌐 One gateway, several clients
A gateway route can now match on the host and carry attributes, signed with the identity. uxopian-ai reads FLOWERDOCS_ENDPOINT and ARENDER_ENDPOINT from them, so each client reaches its own FlowerDocs and ARender through one uxopian-ai. See Per-client routes.
✨ New features
🧩 Plans
- Child runs for sub-plans and plans called as tools, with cycles refused before the run starts.
- Partial fan-out, an All or nothing switch, and an
ON_PARTIALcondition. - Token budget, run validity, unique runs, plan output type, node limits.
- Unbound tool arguments take the value of the same name.
- Broker and worker roles (
agentic.role).
✅ Human review
- Dates panel, archive, delete (with the document, on request), Being reviewed status.
- Undo all (
Shift+U), clear or fill a value, cleaned pasted text, field and table row captions. - Shareable links to a filtered queue or a task.
- One CSV line per task, with three columns per field.
🌐 Gateway
- Routes matched on the host, route attributes signed into the identity.
- Redirect to the login page for a browser that is not logged in.
📂 ARender
- The rendition broker can read the review documents from a shared, read-only folder.
→ Let ARender read the documents from a shared folder
🛠️ For helper authors
@HelperMethod(label = …)gives a helper call a readable label in the prompt editor.
🏗️ Platform & improvements
- 🔑 Your own encryption key, rotated without loss:
APP_SECURITY_SECRET_KEYandAPP_SECURITY_PREVIOUS_SECRET_KEY, also in the Helm chart. - 🗄️ OpenSearch tuned: up to 100 connections, replicas that follow the size of the cluster (
0-1), shards and translog settable per index. - ⬆️ New fields added at startup to the existing indices: no more manual mapping after an upgrade.
- 💬 Quick Prompt polish: an input that grows with the text, dictation that stops on send, a retry that keeps the question, tooltips everywhere, readable helper labels in the prompt editor.
- ♿ A more accessible admin console: a Skip to content link, titles that wrap at 200 % text, clearer announcements for screen readers.
- 🔗 LangChain4j 1.23, with the newer MCP handshake.
- 🔒 Leaner, safer images: base image 1.0.14, no APT, and the vulnerable Base64 intrinsic of Temurin 25 turned off (CVE-2026-114938).
🐛 Bug fixes
- 🧬 NuExtract positions are back: NuMind changed its localization API, and 2026.0.0 returned values without positions.
- 📄 Review tasks with an unread value open instead of failing, and a value shows its page only when it was located.
- ✍️ Monaco editors on the plan page take Space and Backspace again.
- 🔐 Fast2 users land in the tenant of their token (also in gateway 2026.0.1).
- 💬 Quick Prompt: answers stream through the FlowerDocs plugin proxy, and the microphone hides when the page forbids it.
🔄 Upgrade notes
From v2026.0.0
The upgrade is in place: point the stack at the 2026.1.0 images, on the same OpenSearch. The details are in Upgrading from 2026.0.0.
🛠️ Installation
- Plan runs (action required). Finish or stop the running plans first. Runs of 2026.0.0 (
plan_executionsindex) are neither resumed nor shown; plans themselves are kept. - Settings.
agent.execution.*properties are ignored: useagentic.*. A customAPP_SECURITY_SECRET_KEYmust be Base64 of 16, 24 or 32 bytes, or the service does not start. - Replicas. They must form one Hazelcast cluster and run in the same time zone.
- Gateway. Upgrade it to 2026.1.0 too: uxopian-ai works behind 2026.0.x, but route attributes, host matching and the login redirect need the new gateway.
🔌 Integrations
- Plan REST API (action required for REST integrators).
/api/v1/admin/plan-executionsis now/api/v1/admin/plan-runs:POST /plan-runswithinputstarts a run, the signal body field issignal,CANCELLEDisSTOPPED, andEXPIREDandPARTIALare new. → REST API reference - Review task API (action required if you read it). The CSV export has one line per task. Tags combine with or,
orderBy=touchedis nowmodified,status=PENDINGno longer includes the tasks someone holds (RESERVED), and a list withoutarchivedincludes the archived tasks. - Fan-out. A failed element no longer fails the node: set All or nothing where it must. The default parallelism goes from 8 to 32.
- Quick Prompt priority. Higher numbers now show first: invert the priorities you set.
- Fast2 behind a multi-tenant Fast2. Users land in the tenant of their token: data stored under the old configured tenant stays there.
Ready to start? Check out the Quick Start or the full Installation Guide.