Workflows
Chain coding agents, model prompts, commands, browser checks, Git review, MCP tools and approvals — pass results along, run steps in parallel, and start them by hand, on a schedule or on every new commit.
A workflow is a repeatable sequence: a shared brief, then up to sixteen steps that run from left to right. Build one in Workflow mode (Ctrl+Alt+4). The left column lists your workflows; right-click one to Duplicate or Delete it — deleting also removes its run records.
The brief
Every workflow starts with a brief:
| Field | Meaning |
|---|---|
| Name | Shown in the list and in run records |
| Goal | What the workflow should achieve or improve — required. Dictate it with the microphone |
| Expected result | How you recognize success |
| Project | The folder the steps work in |
Agent and model steps receive the brief automatically, so you don’t have to repeat context in every step. The header shows Unsaved or All changes saved; Save stores the workflow.
Recipes
An empty workflow offers Start with a recipe. A recipe asks a few questions and builds the steps for you; you can edit everything afterwards.
| Recipe | Asks for | Steps it creates |
|---|---|---|
| Build a feature | What it should do, who needs it, how you’ll know it works | Sharpen the idea into a short PRD → implement → review and test (read-only) → approval |
| Fix a bug | What’s broken, your test command | Find the cause (read-only) → fix → test command → Git review |
| Check a release | The address of the running app, your build command | Build and browser check in parallel → Git review → go/no-go report |
| Update docs | Which docs | Compare docs with code → update → approval |
Steps
Select a step to configure it in the inspector on the right. Each step card shows its kind, a short summary, its status during a run and the key its result is saved under.
| Kind | What it does |
|---|---|
| Coding agent | Claude Code or Codex works on an instruction in the project, with an optional model. Allow edits or keep it read-only |
| Model prompt | A prompt to Ollama or your own endpoint — for summaries, classification, drafts. Configure the APIs in Settings → Agents & accounts |
| Test command | Runs a shell command in the project; it must succeed |
| Browser check | Fetches a URL and returns its title and a text preview |
| Git review | Checks the diff in the project and summarizes the changes |
| MCP tool | Calls a tool on one of your managed connections |
| Approval | Pauses the run until you approve or reject |
A step that is still missing its instruction is marked Instruction missing; ready steps say Ready.
Passing results along
Give a step an output key, and later steps can use its result as {{key}} in their instruction, command or arguments — for example Fix the bug. Cause: {{cause}}. Ask for JSON and turn on JSON check to make sure a step really returned valid JSON before the next one uses it.
Conditions
A step can run only if an earlier output contains, does not contain, equals or is not empty compared to a value — useful for “only open a draft if the tests passed”. Steps whose condition isn’t met are marked skipped.
Parallel groups
Steps with the same group name run at the same time. Approvals and agents that may edit always run on their own, so two agents never edit the same project at once.
Timeouts
Agent steps may run 30 minutes, model steps 5 minutes, everything else 3 minutes by default. Override the limit per step.
Triggers
| Trigger | Runs |
|---|---|
| Manual | When you click Run workflow |
| On schedule | Every N minutes — from 5 minutes to 7 days — while PaneScale is open. Automatic runs begin after you save |
| On new commit | Whenever a new commit appears in the project |
Runs
Run workflow starts a run. While it runs, every step card shows its status — waiting, running (with what the agent is doing), done, failed, skipped, needs approval, approved, rejected — and the editor is locked. Stop ends the run after the current step.
When the run reaches an approval step, a card at the top says Needs your approval with the step’s note, and Reject or Approve and continue. The same approval appears in the workflows panel next to the grid.
Afterwards, each run has a complete record:
- a step log with inputs, outputs, duration and tokens,
- Git evidence — what changed in the project during agent steps,
- a failure analysis when a step fails,
- an exportable receipt of the whole run.
Retry a failed run at any time. Workflows can also run from the workflows panel beside the terminal grid, without opening the studio.
Agent steps count against your Token Guardian budgets like every other run.