Workflows

A workflow is a script by which several subagents do the work in phases: for example, two reviewers look at a file, each with its own task, and a third merges their findings into a result. The agent writes and edits the workflow — it is enough for you to describe in words what you need. The workflow runs in the background, you follow the phases, and the result arrives in the session.

When a workflow is more useful than an ordinary request:

• the work can be split into independent parts, and they are worth doing in parallel;
• you need a view from several angles — different reviewers with different tasks;
• the result is checked by code: tests, a command, a comparison — not only by the agent’s words;
• you will repeat the same work with other inputs.

How to ask

Describe in the chat who does what and what should come out:

Make a workflow: two reviewers look at src/cart.js, each with its own task (one — errors
in calculations, the other — readability), a third merges their findings into a short result. Run it.

The agent writes the script, checks it before running (result types, phases, commands) and runs it. While the run is in progress, the session shows a background work line, and the workflow’s agents ask for permission like an ordinary subagent — marked “from a subagent”.

When the run ends, the agent gets the result and tells you what was found. The workflow’s result (for example, a review summary) is saved as an artifact of the run.

The “Workflows” screen

• Saved — workflows you can run again. The card has the name, what it does, when to use it, and the arguments. “Run” asks for the argument values, “More details” — the description and arguments for editing.
• Runs — all runs: state, time, tokens, affected files and lines, where it was started from. “Run progress” opens the details:
  • the phases in order and the agents in each — with state, time and tokens;
  • artifacts — for example, “Review result” marked “Main result”, with a version;
  • “Open session” and “Fix in chat” — go back to the agent so it fixes the workflow.

Save and run again

If a workflow turned out well, say “save it”. The agent does not save on its own — it proposes and waits for your agreement; you confirm the save.

Save the workflow with arguments: the file path and two review focuses.

• Arguments — what changes from run to run: a file path, a branch, a review focus. An argument can have a default value and a “required” mark.
• Name — what the workflow is called on the screen. The file name is in Latin letters, the name is in your language; you can edit the name and description under “More details”.
• Saved workflows live with you, not in the repository: ~/.config/buff/workflows/.

You can run a saved workflow from the “Workflows” screen or by asking the agent: “run the file review for src/checkout.js”.

Worth knowing

• A workflow’s agents are subagents: they see the project but not the main session’s conversation. The workflow passes everything they need in the assignment.
• A run with an error in the script does not start: the agent gets a description of the error with a line number and fixes the workflow.
• A workflow’s usage is charged to the organisation’s plan, like any turn; in the workspace it is visible in the session’s records.

In the console

/workflows — saved workflows with a run action, and runs with their state and usage. You can ask the agent to make and save a workflow the same way as in the window.