Task chains
Start a task after other tasks finish, feed it their outputs, and get a periodic check-in on long tasks.
On this page
A task can wait for other tasks, take values from their outputs, and post a periodic digest while it runs.
hexel harness task create "Audit the api repo" --agent auditor
# Task tsk_1 queued ...
hexel harness task create "Summarise the audit" --agent writer \
--after tsk_1 --input-from report=tsk_1.outputs.textFields
| Field | Description |
|---|---|
after | 1 to 20 task ids this task starts after. Each must be a task you can see, in the same workspace and environment, and not already failed, cancelled or out of budget (409 dependency_failed). |
inputs_from | 1 to 32 inputs taken from those tasks' outputs: {"<name>": "<task_id>.outputs.<key>[.<key>...]"}. Each task named is waited on as if it were in after (at most 20 tasks in all). |
checkin | {"every_s": n}: a task.digest event every n seconds (3600 to 2592000) while the task is not finished. |
A task's outputs are {text, result}: <task_id>.outputs.text is the agent's last message, and <task_id>.outputs.result.<field> a field of its structured result (a key of digits indexes a list). An input name is an identifier of up to 64 characters that is not also a key of inputs.
The values inputs_from names are read in your runtime when the task starts; they never pass through the harness's control plane.
How a chained task runs
- It is created
waitingwithwaiting_reason: dependency, and its stream getstask.dependency{waiting_on}. - Within seconds of the last of its dependencies succeeding, it is
queuedand starts. - If one of them fails, is cancelled, runs out of budget or is deleted, it ends
failedwithend_reason: dependency_failed, and so does every task after it in turn.
While it waits, it cannot be paused (409 waiting_on_dependencies; cancel it instead), and its retry or resume is refused (409 dependency_unmet) until every task it starts after has succeeded. A chained task counts toward your organization's running tasks from the moment it is created.
GET /harness/tasks/{id} shows a task's chain: {after, inputs_from, checkin}.
Check-ins
A check-in is a task.digest event: {state, runs, steps, cost: {usd, tokens}, waiting_reason?, since}. runs counts the task's runs (sub-agents included), steps its model calls, and since is when the previous digest (or the task) was written. Send them to a webhook to hear from a long task without following it.
hexel harness task create "Watch the release" --agent watcher --checkin-every 6hinputs_from is not available in a zero-retention environment (409 inputs_from_unavailable): the outputs it names are erased.