Release 1.0.22 — Durable Workflow runtime, Console model enforcement, Cockpit operator tools

A large platform release centred on four themes: durable Workflow execution, model governance in Console, operator tooling in Cockpit, and admin-authored model ratings that show up across the product. Available to all Marketplace subscribers.

Highlights

• Workflows — Runs, schedules, and Wait nodes are now durable and survive service restarts. New File node Preview (download / create / extract / info), an artifact lifecycle (retention + background cleanup + Files download tab per run), and stable node references so expressions bind by node name, not internal id. Code nodes run untrusted code in an ephemeral, per-execution sandbox.
• Console — Enforced max-output-tokens and context window on the agent node and the playground/gateway. Provider diagnostics forwarded to developers so failed calls tell you what actually happened.
• Cockpit — Operator-fed Config Health page and platform-version surface. New license management and migration management UIs. SCIM push-provisioning so identity providers can push user changes automatically. Audit log readability fixes.
• Model selector — Admin-authored cost / speed / quality ratings and tags now flow into the model selector across Copilot and Cockpit.

What’s new

• Durable runtime. Scheduled runs, long-running Wait nodes, and in-flight jobs survive service restarts. Nothing gets silently lost when a pod cycles.
• File node. Preview Download, create, extract, and read info about files inside a workflow. Wired into the builder palette with per-operation config.
• Artifact lifecycle. Every run’s artifacts have a retention policy, a background sweep clears out stale ones, and each run has a Files download tab.
• Stable node references. Expressions and helpers reference nodes by their stable name instead of an internal id, so renaming or reordering a node no longer breaks the references pointing at it.
• Governed Code nodes. The Code node executes untrusted code inside a fresh, per-execution sandbox provisioned by Compute — no shared state between runs, hard limits enforced.

• Model max-output-tokens enforced. Playground, gateway, and the agent node now honor each model’s declared output limit.
• Context-window enforcement. Requests that exceed a model’s context window are rejected up front rather than failing deep in the stack.
• Provider diagnostics for developers. When a provider returns an error, the developer-facing response now carries the provider’s diagnostic detail while billing- and auth-sensitive fields stay redacted.

• Config Health page. Operator-fed dashboard showing the health of key subsystems, plus the platform’s exact version.
• License management UI. View and manage the platform license from the admin surface (gated).
• Migration management UI. Inspect and manage database migrations from Cockpit instead of the CLI.
• SCIM push-provisioning. The Keycloak createuser receiver now accepts SCIM push, so identity providers can push user create / update events into the platform automatically.
• Audit log readability. Filters no longer overlap and the layout has been tightened for scanning.

• Admin-authored ratings and tags. Admins can now tag models and rate them on cost, speed, and quality. Those ratings + tags feed the model selector everywhere it appears, so pickers reflect your organization’s judgment, not just the provider’s marketing.

• Assign a pipeline to a Model (LLM) — attach a guardrail pipeline directly to a model in Copilot, so every call through that model runs the pipeline without wiring per-agent.

Improvements

• Streaming large-file put and read — the platform now streams multi-GB files in and out of storage memory-flat, so a big upload no longer inflates service memory.
• Provider-agnostic storage backends — self-hosted and Marketplace customers can run the same platform image against different object-storage backends without code changes.

Fixes

• Fixed: storage-listing failures now surface as errors instead of silently returning an empty thread list. Empty state now really means empty; a failure now really shows the failure.

• Fixed: audit log filters no longer overlap; the layout has been laid out so long filter labels don’t collide.

• Removed: dead MVP access-policy tables and columns that were still shipping in earlier releases have been cleaned up. No behavior change; a small reduction in what your admin surfaces expose.

Rollout

All Marketplace subscribers were updated to 1.0.22 during the maintenance window on 3 August 2026. No customer action required.

Questions? Email support@alchemistudio.ai or reach us on LinkedIn .