All pages

Docs · Right-sidebar surfaces

Workflows (Flows)

Availability: macOS · cloud workspace (runbooks_*). run_runbook itself is macOS-only.

The Flows tab in the right sidebar is where saved automation lives. It collects three kinds of reusable automation in one place, each with run logs, tags, and import/export:

  • Runbooks — saved agent sessions you can replay
  • Browser actions — recorded browser sequences replayed without the LLM
  • Tab-monitor rules — pair a trigger (cron, file change, idle, git commit, or an output line match) with an action

A search field at the top (Search runbooks, actions, rules) and tag filters narrow the list. The overflow menu offers Import workflow… (a .trmnly-workflow file picker) and Reveal storage folder.

Runbooks

A runbook is a saved sequence of agent steps. Capture one with Save current session (⌘⇧S) at the top of the panel, from any agent session menu → Save as Runbook…, or start from a starter loop with From Template (build/test/fix, debug, infra bootstrap, and more).

Each row shows the name, step count, a parameters badge when the runbook takes input, last-run metadata, and tags. The Play button runs it; the row menu has Edit…, Reveal source session, Edit tags…, Export…, and Delete….

Step types a runbook can sequence:

Step Does
Prompt Send a prompt to the agent
Browser action Replay a saved browser sequence
Runbook Run another runbook (cycles are guarded)
Shell command Run a command
Wait Pause for N seconds
Repeat Run nested steps a fixed number of times
Loop until Repeat nested steps until the agent says a marker phrase or a file exists, up to an iteration cap

Parameters. Prompts can embed placeholders that are filled at run time:

  • Built-ins {cwd}, {project}, and {date} resolve from the active window and project. {branch}, {file}, and {selection} are recognized but currently expand to empty.
  • {?name:"Label":"default"} (or {?name:"Label"}) prompts you for input via a parameter sheet before the runbook starts.

Browser actions

Disk-based browser sequences saved under agent/browser-actions. Record one by asking the agent to start a recording ("record a saved action called Login"), then stop it. The Play button does a direct replay — no LLM round-trip; the row menu also offers Ask agent to replay…, Reveal in Finder, Edit tags…, Export…, Convert to runbook (a one-step runbook that replays the action), and Delete…. See Browser for the recording tools.

Tab-monitor rules

Rules that pair a trigger — a cron interval, file change, tab idle, git commit, or a terminal-output line match — with an action. This is the same Tab Monitor feature that used to be a standalone tab; it now lives here with run logs, tags, and project scoping. Each row has an enable toggle and a menu with Edit in full editor…, Test fire, Edit tags…, Export…, and Delete. Create new rules via Open editor… (a single entry point, so you never end up with empty dangling rules).

See Tab Monitor for the rule anatomy, actions, and pattern examples.

Scope and persistence

Tab-monitor rules can be scoped to a project (rows show a scoped marker). Runbooks and browser actions pick up project context at run time for parameter expansion. Every run — runbook or rule — records a one-line log entry with outcome, duration, and timestamp.

Scripts can start a runbook with trmnly://workflow/run/<id> or replay a saved browser action with trmnly://workflow/run-browser-action/<name>.

From the agent

list_runbooks, save_runbook, update_runbook (edit a runbook in place by id or name; omitted fields stay as they are), delete_runbook and run_runbook manage runbooks. run_workflow_steps runs an ad-hoc list of steps (prompt, browser action, runbook, shell, wait) without saving it, and returns at once with an owner_key; list_workflow_runs shows how it went. list_workflow_tags and set_workflow_tags read and replace the tags on a runbook, browser action or monitor rule — the same tags the filter chips use.