nodecode-workflows

Workflows: agent-authored workflow graphs run in the background. A Nodecode rewrite of jacobhausler/hermes-workflows by jacobhausler.

Get it

Tell the agent "add the workflows cell", or run nodecode add nc://workflows.

Set it up

The workflows section is optional and has defaults: max_parallel (3), tick_seconds (2), node_timeout_seconds (900) and output_chars (4000). See config.example.jsonc. Nothing else is needed: the cell runs nodes as Nodecode turns in sessions of its own, so the model in use is the model they take.

Use it

The model writes a graph in JSON and hands it to the workflow tool:

{"name": "check",
 "nodes": [{"id": "inspect", "type": "agent", "goal": "Inspect the target and answer with one JSON object."},
           {"id": "report", "type": "agent", "goal": "Write the report.", "after": ["inspect"]}]}

and ends its turn. One worker thread runs the graph: each agent node is one turn in a session of its own, after edges order the nodes and hand every child its direct ancestors' committed output under ## Inputs, and the session that started the run is woken with one line when the run settles.

Action	What it does
`run`	validate a graph and start it; returns at once
`validate`	lint a graph, every problem in one answer, no writes
`status` / `list`	one run node by node / every run in the home
`show`	what one node committed
`stop`	nothing new starts; what committed stays
`answer`	answer a gate and let the run go on

A graph is data, so a node can also be:

{"type": "echo", "echo": "text"} — a constant, no turn spent;
{"type": "gate", "question": "Publish?", "options": ["yes", "no"]} — the run holds there until workflow{action:"answer"} or /workflow answer;
{"fanout": {"items": ["a", "b"]}} or {"fanout": {"items_from": "plan.items"}} — one lane an item, id.1, id.2, and {"from": "plan"} hands a later node the whole array;
{"inputs": ["plan.items"]} — a dotted path from any ancestor;
{"schema": {...}} — the reply contract, written into the prompt.

Every output is written under workflows/<run-id>/ in the Nodecode home, with the graph and each node's state in run.json: a node that committed is never run again, so a run interrupted by a restart picks up where it stopped.

From the keyboard: /workflow lists the runs, /workflow status ID, /workflow show ID NODE, /workflow run PATH, /workflow stop ID, /workflow answer ID NODE TEXT. From Lisp through eval: (workflows:start GRAPH), (workflows:status "id"), (workflows:show "id" "node"), (workflows:runs), (workflows:stop "id"), (workflows:answer "id" "node" "text"), (workflows:validate GRAPH). (help :workflow) is the manual page for the model.