Templates

Action inputs are MiniJinja templates.

[sources.actions.with]
cmd = "echo {{ item.reference_id }}: {{ item.title }}"

Context

Templates can read:

• workspace
• source
• item
• action
• actions (preceding named Actions only)

Example values:

{{ workspace.id }}
{{ workspace.path }}
{{ source.id }}
{{ source.source.uses }}
{{ source.actions[0].uses }}
{{ item.id }}
{{ item.reference_id }}
{{ item.title }}
{{ item.status }}
{{ item.url }}
{{ action.uses }}
{{ action.index }}
{{ actions.issue_worktree.inputs.root }}

source is the complete configured Source. Adapter settings stay nested under source.source, and configured Actions stay under source.actions.

An Action may declare a Source-scoped id. Later Actions can read that named Action's final rendered inputs through actions.<id>.inputs. Unnamed Actions are absent, and missing or forward references fail rendering for that Item.

[[sources.actions]]
id = "issue_worktree"
uses = "@agentboard/action-worktree"
[sources.actions.with]
repo = "~/Projects/MyProject"
root = "$WORKTREE_ROOT/{{ item.id | slugify }}"
branch = "{{ item.id | slugify }}"

[[sources.actions]]
uses = "@agentboard/action-run-cmd"
[sources.actions.with]
cwd = "{{ actions.issue_worktree.inputs.root }}"
cmd = "pwd"

Stored-success skips and dry runs still render each Action in order, so later Actions receive freshly rendered named inputs. Action IDs do not change retry identity.

Use item.reference_id for provider-facing names and messages. item.id is the stable identity used by the Store and Action retry checks.

item.raw contains the original Source payload:

{{ item.raw.jira.fields.priority.name }}

Filters

AgentBoard registers slugify for conservative path and branch names:

{{ item.title | slugify }}

Example:

"Fix Login!" -> "fix-login"

Rendering and expansion order

AgentBoard processes Action inputs in three stages:

1. MiniJinja renders every input.
2. AgentBoard expands leading ~/, $VAR, and ${VAR} only in path inputs:
   • @agentboard/action-run-cmd: cwd
   • @agentboard/action-worktree: repo, root
3. @agentboard/action-run-cmd passes cmd and healthcheck to sh -c. The shell expands command variables after changing to cwd.

[[sources.actions]]
uses = "@agentboard/action-worktree"
[sources.actions.with]
repo = "~/Projects/MyProject"
root = "$WORKTREE_ROOT/{{ item.id | slugify }}"
branch = "{{ item.id | slugify }}"

[[sources.actions]]
uses = "@agentboard/action-run-cmd"
[sources.actions.with]
cwd = "$WORKTREE_ROOT/{{ item.id | slugify }}"
cmd = '''printf '%s: %s\n' "{{ item.reference_id }}" "$PWD"'''

Here AgentBoard expands $WORKTREE_ROOT in path fields, MiniJinja renders item fields, and the spawned shell resolves $PWD from its configured cwd.

Action hash

AgentBoard hashes the final inputs after MiniJinja and AgentBoard-time path expansion. Those exact strings are passed to the Action, so shell variables in cmd remain literal in both the hash and the command given to sh -c. The hash is part of retry identity:

(source_id, item.id, source_action_index, rendered_action_hash)

If a template renders different inputs, AgentBoard treats it as new work.

Examples

Create one worktree per item:

[sources.actions.with]
repo = "~/Projects/MyProject"
root = "~/Projects/MyProject.worktrees/{{ item.id | slugify }}"
branch = "{{ item.id | slugify }}"

Open a terminal tab:

[sources.actions.with]
cmd = "zellij action new-tab --name {{ item.id | slugify }}"
cwd = "~/Projects/MyProject"