Claude Agent Flow — Development skill for Claude Code

agent-flow

The agent flow pane as a plugin: `/flow` opens a live tree of the session's subagents and in-process teammates beside the transcript, and closes it again. Each row is one agent: status, type, name, description, elapsed time, what it is doing right now (a tool call and how long it has run, or a wait for the person's approval), its call count, and its tokens once it finished. A `[+]` on a row expands its details: model, prompt excerpt, recent tool calls, tokens. Agents waiting for approval are highlighted; a tool call over 30 seconds and a running agent silent for two minutes are marked too. A wait for approval stays shown until the approved tool call ends, since no engine event carries the person's answer.

The tree comes from engine events (`agent.spawn`, `tool.call`, `turn.complete`, the classic permission events) and is reconciled with `$.agent.list()` every two seconds while anything runs. Loops the engine never listed (a Workflow tool's agents, the engine's own forks) appear under a collapsed "unlisted loops" group so they never vanish silently.

Where the surface cannot draw a pane (a `-p` run, the VS Code extension as of 2.1.270) `/flow` prints the same tree as text; `/flow text` always prints it. When the surface seats the pane inline above the prompt (narrow terminals) it shows only the rows that need attention. The first spawn of a session opens the pane by itself on a terminal of 144 columns or more (110 when you kept it open before), unless you closed it.

Hooks modules are early access and load only where function hooks are enabled; see `mods/README.md`.

What it hooks

event	what the hook does
session.start	Binds the engine, registers /flow (stands down when another /flow is listed), reads the open preference, reconciles once.
ui.render of PromptHint	Reads the terminal's width for the auto-open decision.
ui.render of Pane	Draws the pane: header, root, tree, unlisted group, last event; th