Automation variables

Templates use {{variable}} placeholders. The run store is a flat string map: keys may contain dots (repo.full_name), but they are not nested objects.

How references work

Write the full catalog key inside {{…}}.
Use {{repo.full_name}}, {{workflow.name}}, {{steps.fix.summary}}.
Do not write {{repo}} or {{workflow}} alone — those keys are not published and become empty strings at run time.
Dots are part of the key name, not object path navigation. {{pr.head.ref}} looks up the single key "pr.head.ref".
Bare keys are valid when the catalog lists them ({{failed}}, {{head_branch}}). An event. prefix is optional for the same value ({{event.failed}} ≡ {{failed}}).
Step outputs require an id on the producing step, then {{steps.<id>.<field>}}. The latest agent / commands / check step also publishes {{agent.summary}}, {{commands.output}}, {{check.passed}}.
Unknown keys interpolate to empty. The editor and draft tools warn when a template references a key that is not available for the trigger or prior steps.

Event variables

Each trigger publishes fields from its catalog entry. Common examples:

Manual: message, inputs, plus each declared input as inputs.<name> and bare <name>
GitHub workflow run: conclusion, failed, workflow.name, run.url, head_branch, failed_jobs, repo.full_name
GitLab pipeline: pipeline.status, pipeline.failed, pipeline.ref, pipeline.url, project.path_with_namespace

Use get_automation_guide for the complete event field lists and check-step patterns.

Built-in automation fields

Variable	Meaning
`automation.name`	Workflow name
`automation.id`	Automation id
`event.type`	Trigger event id
`event.received_at`	When the event was received

Step outputs

After a step with id completes, later steps can read:

Pattern	Meaning
`steps.<id>.summary`	Agent assistant summary
`steps.<id>.outcome`	Agent outcome (`success` / `blocked` / `failed`)
`steps.<id>.thread_id`	Agent or commands thread id
`steps.<id>.task_marked_complete`	Whether the agent marked the task complete
`steps.<id>.exit_code`	Commands exit code
`steps.<id>.output`	Commands combined output
`steps.<id>.failed`	Commands failed (`true` / `false`)
`steps.<id>.last`	Last command string
`steps.<id>.passed`	Check expression result
`steps.<id>.expression`	Check expression text

The latest agent / commands / check values are also available as agent.*, commands.*, and check.* (for example {{agent.summary}}, {{commands.output}}, {{check.passed}}).

Prefer stable step ids so prompts stay readable.

Interpolation sites

Agent prompt, repository, branch
Commands strings, workingDirectory, repository, branch
Notify title, body, href
Check if expressions (operands resolve the same way)

Always pass repository and branch on agent / commands steps from event variables (for example {{repo.full_name}} and {{head_branch}}) so the step attaches the correct checkout.