LLM action
Use @clankpipe/action-llm to run a Pi prompt directly or in a terminal.
[[sources.actions]]
uses = "@clankpipe/action-llm"
[sources.actions.with]
prompt = "Implement {{ item.reference_id }}: {{ item.title }}"
mode = "foreground"
[sources.actions.with.terminal]
kind = "herdr"
container = "tab"
harness = "pi"
harness_args = ["--approve"]
cwd = "/home/me/dev/project"Set exactly one of prompt or prompt_file. The prompt file path is read after
ClankPipe renders action inputs.
Use worktree to create or reuse a clean Git worktree before the agent starts.
The action does not switch an existing worktree to another branch.
[sources.actions.with.worktree]
repo = "/home/me/dev/project"
root = "/home/me/dev/project.trees/{{ item.id }}"
branch = "clankpipe/{{ item.id }}"Set terminal to zellij, herdr, tmux, or generic. Set harness to the
agent harness executable and harness_args to its arguments. The rendered
prompt is the final harness argument. Pi starts a new saved session by default.
Configure --session-id or --no-session explicitly when you need different behavior. Direct launches wait for Pi in
foreground mode. Direct background launches return running immediately
and ClankPipe records the final result when the child exits.
Terminal launchers return running after they accept the command. Herdr
worktree launches create a new tab in the worktree workspace. Herdr, Zellij,
and tmux do not expose agent completion to this Action. ClankPipe
marks such executions stale (disconnected) during recovery unless a later
Action execution records completion.