Approve agent merges in your Inbox, auto-fix red builds, and turn issues into Tasks

Some outcomes are not a number: "ship feature X across three repositories" or "migrate the docs site". A delivery Goal covers them.

1. Open Goals and create a new Goal. Set Goal kind to Delivery, and the metric source and target fields disappear.
2. Enter the Definition of Done, one criterion per line. At least one is required, and every line you enter starts approved.
3. Activate the Goal. The execution loop hands iterations to agents as usual.

The Goal completes as achieved when every approved criterion is done or waived. Criteria an agent proposes do not count until you approve them, so an agent cannot finish a Goal by declaring it finished. A deadline still applies, and a passed deadline ends the Goal as missed.

A brand-new Goal also no longer stalls for want of an agent: if none is pinned and none has worked on it yet, it picks from the eligible agents in the Goal's own scope.

Limits: the kind is fixed once the Goal exists, and the checklist can be edited but never emptied. Existing Goals stay metric Goals.

Cost and tokens. When a run executes on one of your machines, Ever Works now records its token counts and, when the CLI prices the run, its cost, so budgets, Goal spend and cost views include it. The figure is the CLI's own estimate, billed to the account that machine's CLI is signed in to, never to platform credits. Codex reports tokens but no price.

Daily cost ceilings. On Settings → Fleet, set a Daily cost ceiling for the whole fleet, or open a machine's details to set one for that machine. Days are counted in UTC. Crossing a ceiling drains the affected machines and files one Inbox notice. Drained machines stay disabled until you re-enable them, so raise the ceiling first if the spend was expected. A machine ceiling on a CLI that reports no price drains that machine on its first finished run.

Panic controls. Drain all nodes disables every enrolled machine and returns their jobs to the queue without cancelling anything. Cancel in-flight work is a separate step that aborts running jobs and their agent runs, optionally including queued ones. A banner on the page tells you when the platform-wide stop is on; that stop is set by the platform operator.

Concurrent iterations. A Goal still runs one iteration at a time by default. On the Goal page, open Adjust limits and set Concurrent iterations (1 to 10) to let the loop start several at once, each as its own Task and agent run. Every iteration passes the same checks as a single run, including budgets, credits and concurrency limits, so a higher number asks for more parallelism without adding capacity.

Raise it only for Goals whose iterations are independent. Nothing keeps their workspaces apart, so iterations editing the same repository on the same branch will overwrite each other.

Template steps across repositories. A step in a workflow template can now name its own Work and extra repositories, so one template can say "write the spec in the platform, update the docs site, publish the announcement". Create the Task tree as before from Tasks, new Task, From template. Each step's Work and repositories are checked against you when the template is saved and again when it is used, and a step you can no longer reach stops the whole tree instead of filing work in the wrong place.

Limits: creating and editing workflow templates is API-only for now.

A machine that had stopped itself used to look online and idle while refusing every job. Now:

• Worker state in the machine's details. On Settings → Fleet, open a machine to see Idle, Working, Paused, Quarantined or Throttled, with the reason and since when. A machine that has never reported shows Unknown, not idle, and its job history shows each run's real outcome.
• Inbox notices. You get one notice when a machine goes offline, another if it stays offline (30 minutes by default), and one when it quarantines itself. Each resets when the machine recovers.
• Routing to machines that can take the job. Only machines that match the Agent's pinned machine and the capabilities the job needs count as free. If none can take it, your execution routing rule applies: by default the work runs in the cloud and you are told why, or, if you chose to wait for your own machines, it stays queued and starts once a matching machine is back online.
• No job waits forever. An agent job no machine picks up within 24 hours fails with a Fleet run never started notice.

Limit: offline notices are checked while a dashboard page is open or when work is dispatched, so with nothing open you see them on your next visit.

Sleep-safe runs. Desk machines sleep. Before, a sleeping machine's job could be handed to another machine while the first model run was still going, and one of the two runs on the same branch was wasted. Now the node notices when it wakes, stops any run whose claim lapsed while it slept, pushes nothing and reports the failure, and the job carries on elsewhere.

Disk floor. A machine refuses new work while its workspace volume has less than 2 GiB free, or when free space cannot be measured, and shows as Throttled. A job refused for disk goes back for another machine instead of failing. Change the floor with --min-free-disk, or switch it off with --no-disk-floor.

Workspace cleanup. With --work, the node removes Task workspaces older than 14 days (--workspace-max-age) only when it can prove they are safe: not in use, no uncommitted changes, fully pushed, and the branch merged or gone. ever-works-node doctor shows what it would do, and ever-works-node gc --dry-run previews a sweep. A machine's details show Disk & workspaces: above or below the floor, workspaces kept and the last cleanup.

Action needed: update the node app on every machine. Node builds from before this release cannot keep or finish jobs.

Env files on your machines. Store a repository's .env files under Settings → Repositories. When a run on one of your machines needs them, the node fetches them only while it holds that job, writes them readable by the owner only, keeps them out of Git, and deletes them before the run's first Git command and again when the run ends. The job records which files are needed, never their contents, and the values are scrubbed from what the node reports back. If a file can no longer be resolved, the run fails instead of starting with half an environment.

Env grants (API only). You can also let runs read specific variables that are already set on the machine, by exact name. Agent-driven code can then read those values, so grant as few as possible.

Repository-declared setup and checks. A repository can list tasks.setup and tasks.checks commands in .works/works.yml, including checks for another repository mounted in the same Task. Setup runs before the agent, and a failed setup is reported as a setup problem, not as a failing test. Nothing runs unless each command exactly matches the Work's allow-list, which is set through the API for now.