Skip to main content
Every change to a running execution is a command. People, supervisors, the REST API, the MCP server, the CLI and the web UI all send the same commands through the same pipeline.

​
How commands work

  • All commands for one execution enter one queue and are applied in the order sent. The caller’s response arrives only after the command has been applied.
  • Commands for different executions are independent.
  • Every command for a known execution is recorded in the execution log as applied or rejected, with the actor: the API key name (key:<name>), a supervisor id, or the UI user. GET /api/control/audit lists who sent which command.
  • Before a command runs, reviewers can refuse it. A refusal returns 403 and is logged. See Supervisors.
  • The pipeline is built for people and supervisors, not as a high-volume API.

​
Pause and resume

Pausing stops a node, or the whole execution, from taking new messages. The message being processed finishes. All time limits freeze while paused, so a pause never causes a timeout.

​
Step

Step mode lets exactly n messages through a node, then pauses it again. Use it to watch a pipeline one message at a time.

​
Breakpoints

A breakpoint holds a node before the first message whose payload, a path inside the payload such as payload.order.id, or a metadata.key equals a given text. It is ordered and audited like any command.

​
Skip and unskip

Skipping a processor passes messages through it without calling its plugin, from the next message on. A message already inside the plugin finishes. Unskip resumes normal processing. Other node kinds reject skip.
A processor can also be skipped from the start of a queued execution, and unskipped later.

​
Stop

Three strengths of stopping an execution:
A single node can also be stopped, draining or immediately.

​
Restart, restart from a node, and retry a node

Recent finished executions stay restartable: the server keeps their graph, start payload and recorded node outputs (the newest 10 by default, with per-node caps). A node that exceeds the caps makes a restart that needs it refuse rather than replay partially, unless a spill directory is configured. After a server restart, use resume instead. A restart takes a slot like a new start, so it can be queued or refused with 429.

​
Signals

A signal hands a named answer to a waiting node, such as an approval. See Approvals and signals.

​
Change the queue

Operators can change the scheduler limits and a queued execution’s priority while the server runs. See Admission and limits.

​
Who can do what