v0.19.0

Model ingest opens to tenant administrators, image generation on the deployment’s own ComfyUI now edits and inpaints, takes LoRA and a seed, and the image engine can be bought on Spot. A member’s home disk can be grown without stopping it, and a pull request opened in work items shows its review and CI state — with GitHub pull requests appearing again.

New

• Tenant administrators can search and take in models. Once the super admin allows a tenant (in its tenant dialog), its administrators get “Inference engine models” in the tenant settings: search, ingest, the tenant’s own ingest jobs and the catalogue read-only. Modes, GPU classes, enabling and the Hugging Face token stay with the super admin.
• Image generation on the deployment’s own ComfyUI edits and inpaints, for all five model families. A request can carry LoRA (up to four, each with a weight) and a seed, so a picture can be reproduced. A request that arrives while the model is still syncing onto the instance now waits instead of failing.
• The image engine can be bought on Spot. ImageCapacityOptionType=SPOT in params/60-engines; the default is on demand and nothing changes unless you set it. The llm role is not offered Spot on purpose — a termination notice mid-conversation would cost a cold start.
• The llm engine takes LoRA too. Register a LoRA row on top of a catalogue model and it is served as a fixed preset.
• A split model is registered as one row. The ingest and registration forms take a model family and several files with their roles (diffusion model, text encoders, VAE), a file can be added to an existing row, and a pasted Hugging Face or Civitai URL fills in the fields.
• A member’s home disk can be grown without stopping it (ecs-ec2). Raising disk_gb in the member’s detail extends the existing volume online; lowering it leaves the volume as it is. The member detail is now arranged into size and limits, actions, and irreversible actions.
• A quarantined slot can be terminated from the Console (ecs-ec2), with its ID and the reason for the quarantine shown before confirming; the reason is kept in the audit log.
• Opening a pull request in work items refreshes it: draft state, conflicts, review status, CI result, base and head, and the size of the diff. The main button opens it on GitHub or Bitbucket, and starting a session moves under it.
• Settings > Environment shows whether the upstream CLI release watch is alive, warning only when a source fails or the last success is older than 48 hours. The Control Plane reads a public GitHub issue anonymously for it; on a deployment without outbound access the line simply does not appear.

Improved

• Searching models from the ingest screen shows more: a link to the page on Hugging Face or Civitai, the publication date, and the licence on each catalogue row with who accepted a non-commercial one and when. Picking a result goes straight to its file list, and a licence can be recorded for a hand-registered model.
• The VRAM check for the llm engine counts the KV cache, read from the model’s own header, so a model is no longer offered a GPU class it cannot run on. Declare vramMiB in a ladder as the card’s physical amount (22000 for an L4).
• A model enabled while the engine is up arrives without a restart. The engine tab is also easier to read: buttons look like buttons, on/off is a badge, search results are cards, and ComfyUI’s measured VRAM per model reaches CloudWatch for whoever sizes the ladder.
• A session that started sessions can rename them, unless a person gave the child its title in the Console.
• Session and conversation links in the mirror show their title on hover.
• Schedules sit right under the assistant in the left pane.
• Agent kinds and connected services are shown by their brand marks, in the session list, the settings cards and the workspace bar.
• A Subversion checkout whose credentials were not saved can be given them later, from the row menu or Settings > Connections; svn run inside the session uses them too, and a handoff from an SVN checkout no longer offers a worktree.

Fixed

• Image generation on the deployment’s own ComfyUI did not work end to end from a Console-only setup: SD3.5 produced nothing, a model without a declared family could not be used, a mode change did not reach running sessions, and a generation that spanned an instance change failed instead of retrying. When no provider is named, the deployment’s own engine is now tried before your own Antigravity or Codex plan.
• Right after a Control Plane restart, generate_image could be refused as “not allowed” although it was in the tool list.
• GitHub pull requests did not appear in work items. The default query asked for the wrong thing; a saved query is not rewritten, see Upgrade notes.
• A GPU class that failed to apply could not be retried, and a changed ladder did not reach the Control Plane until it was restarted.
• Model ingest: purging a row deleted files another row still used, a row missing a required file could be enabled, and Civitai assets that need a login or Hugging Face repositories whose terms you have not accepted came back as a generic failure.
• Words in angle brackets, such as <repo url>, vanished from the mirror and the viewer.
• Memos: a long memo, once expanded, could not be collapsed, and a single memo was sent with a heading and a number.
• CloudFormation parameter descriptions showed ? where the text was not ASCII.

Upgrade notes

• ECS deployments running the inference engines: before updating, make sure params/60-engines has LlmEnabled=true and/or ImageEnabled=true for every role you use. Six model parameters are gone from the template (LlmModelS3Key, LlmModelIds, LlmContextTokens, LlmMaxOutputTokens, ImageModelS3Key, ImageModelIds) — models are chosen from the catalogue in the Console. A role that was switched on by naming a model key alone, with Enabled left empty, is deleted by an ordinary stack update, silently, together with its running engine. standup.sh and update.sh carry the switch over for you and say so on stdout; a hand-run aws cloudformation deploy does not, so pass <Role>Enabled=true in that same update. deploy/aws/ecs/cfn/PARAMETERS-60-engines.md, “Upgrading”, has the details.
• The first update.sh on this release also updates 20-platform. It adds the ECR repository for the new af-engine-tools image (the engines’ fetch and ingest steps moved out of the template into it) and the permission that grows a home disk. The script prints the change set first and executes it only when nothing is replaced, then copies the image from GHCR into ECR and deploys 60-engines. Nothing to do beyond reading what it prints; if it says the image is not on GHCR, run the engine-tools-image workflow first.
• Spot for the image role is opt-in. Before setting ImageCapacityOptionType=SPOT: create the Spot service-linked role (AWSServiceRoleForEC2Spot) if the account has never launched a Spot instance, check the Spot quota (L-3819A6DF), and allow more than one instance type in the ladder — one type alone was refused for capacity. Switch while the image role is stopped, and put the value in params/60-engines too so standup.sh does not switch it back.
• Change sets on every template will show many description-only differences (non-ASCII characters removed, prose shortened). Behaviour is unchanged.
• Work items: a saved GitHub query is not rewritten. Set it to is:open involves:@me under ⚙ > Work items (add is:open review-requested:@me as a second one if you want reviews). OR in a query works once the workspace has been restarted on the new image.
• Coming from 0.17.0 or earlier: 0.18.0’s notes still apply — the Control Plane task is replaced once for the architecture spelling, and LlmModelFile / ImageModelFile are gone as well.

Agent CLI pins: Claude Code 2.1.268, Antigravity 1.2.1. The others are unchanged from 0.18.0.