Actions

Actions are side effects that run for each Item collected from a Source.

Define them in the sources.actions array in your workspace config.

Each action is a sequential step that runs for each item in the source. Actions are blocking per item.

This CLI doc only covers orchestration: ordering, retry identity, and trust model. Action-specific config belongs in each action package doc.

Workspace Shape

[[sources.actions]]
uses = "@agentboard/action-run-cmd"

[sources.actions.with]
cmd = "echo {{ item.id }}"

Sources run concurrently. Item Actions inside one Source are serial.

For each Item returned by a Source, AgentBoard executes Actions in the order defined in the sources.actions array.

Actions are blocking per item. If action 0 fails for an item, action 1 for that item does not run during that Run.

An Action runs when no previous successful attempt exists for:

(source_id, item.id, source_action_index, rendered_action_hash)

Changing the rendered inputs changes the hash and makes the Action eligible to run again.

Failed Actions retry on the next run, including run --watch, until they succeed.

Workspace configs are trusted local code, like a Makefile. AgentBoard does not sandbox commands.