Pipelines, tasks, catalog, triggers and the Workshop

Pipelines (#/pipelines)

Lists the pipeline files in the project catalog. Selecting a row shows an inspector with Start a run and Open in Workshop. Opening a row (Enter, or #/pipelines/<id>) opens it in the pipeline editor.

Pipeline editor

Graph of the stages on the left, inspector on the right, YAML panel and a Problems panel (validation findings, with quick fixes) below. ⌘S saves, A adds a stage, ⌥⇧F formats the YAML, ⌘. applies a quick fix. Tabs: Editor, Runs (this pipeline's runs) and History (saved versions).

Stage channels

Selecting a stage shows a Channels section in the inspector. Pick a connection from the Connections list, enter the recipient (fields match the provider: several addresses and Cc for AgentMail, a channel id for Slack, a chat id for Telegram, a phone or email for iMessage, a phone plus an optional template for WhatsApp), and the editor writes channels: onto the stage. It shows the connection's status, flags a recipient the provider would reject as you type, and hides Add channel once the stage has its one allowed channel. A pipeline-level default channels: is edited in the YAML panel and survives saves. See Communication channels.

Tasks (#/tasks)

Task files with their last run. N creates a task, E edits, S starts a run, Enter opens the last run.

Catalog (#/catalog)

Two tabs: Stages (stage files; E opens one in the editor, N creates one) and Skills (installed skills and which stages use them; I installs one, C copies a name). #/skills opens the Skills tab.

Triggers (#/triggers)

Catalog triggers: N creates one, F fires it once, E enables or disables it. Email triggers need a connected email account.

Workshop (#/workshop)

The interactive authoring surface for pipelines; see Workshop Author and the section below.

How it works

Workshop (#/workshop) is the interactive authoring surface. Chat turns go to POST /api/workshop/chat, which uses an AgentPort with the Workshop Author profile (baked playbook, draft tools, docs retrieval). A stage run uses StagePort instead. AgentPort stays up and answers messages; Workshop is the first caller, and a later scout or researcher would use the same port.

The Author clarifies intent, then creates and edits. Create and edit tools mutate the in-memory draft immediately, so the studio updates before Accept. Accept confirms. Reject soft-undos that mutation when the draft fingerprint is unchanged. Soft undo does not reverse a successful save. Disk writes happen only when you ask the agent to save (validate, then write).

Click a stage on the map for a read-only summary: prompt, IO, verify, and HITL.

Untitled drafts are builds under $STAGEFLOW_HOME/workshop/builds/{id}.json. Chat sessions (History / New) live under $STAGEFLOW_HOME/workshop/sessions/ and store activeBuildId, the build that session was editing. The session blob does not embed the draft. The studio picker lists open builds and on-disk pipelines. History reopens the session and its pinned build. See Workshop Author and Data directory.

The UI prefers an NDJSON stream (Accept: application/x-ndjson) for progressive assistant text; otherwise it uses a coherent JSON turn.

A new chat selects its model from the project stageflow.yaml model, then the Workshop model saved in Settings, then cursor/auto. The composer can override that choice for the current session. See Workshop Author — Model.

Manual path

From the repo (or a consumer checkout with Stageflow):

STAGEFLOW_HOME=$PWD/.stageflow-home npm run dev -- ui

Open the console URL, then navigate to #/workshop. New starts a fresh greeting. Chat until the map matches the workflow, then ask the agent to save. A save that does not name a folder writes workshop/<pipeline-id>/ and adds workshop to the catalog, so the pipeline and task show up under Start a run. History restores the session and the build it was editing.

Dev without a global install uses the same entrypoint (npm run dev -- ui). Prefer an isolated STAGEFLOW_HOME so builds, sessions, and settings do not touch your real ~/.stageflow. sf ui registers the directory it was started in, so that project's catalog is browsable. Start it from the project root.