Find a workflow for a job
find_workflow · Reads only
Finds workflows that get a job done in several checked steps, described in plain words: Trillion’s own, this account’s saved ones and ones other accounts published. Returns each workflow_id, its job, its steps and next_call, the start_workflow call that runs it. It runs nothing. Use it when a job has a known method (research, a review, a plan) rather than a single tool call; find_tool_for_job finds single tools.
Inputs
Example
Read a workflow
describe_workflow · Reads only
Reads one workflow whole: its job, input_schema, and every step with what to do (do), the shape of what the step hands back (submit, a JSON Schema) and the checks it must pass, plus the identity of this exact version. Use it before starting, or as the example to copy before save_workflow. Changes nothing.
Inputs
Example
Start a workflow
start_workflow · Changes things
Starts one run of a workflow with its input (matching its input_schema) and returns the first step: what to do, the shape to hand back and the checks. You do each step’s work yourself, with your own tools, except a step that names a tool, which Trillion runs where that tool lives; a step is done only when its output is admitted. request_key is any unique text; the same workflow and request_key return the run already started. The run keeps the exact version of the workflow it started with.
Inputs
Example
Submit a workflow step
submit_workflow_step · Changes things
Hands back the output of the step the run is on (step_id). It is admitted only when it fits the step’s submit schema and passes every check; a refusal lists every problem at once, and nothing is kept until it passes. For a step that runs a tool (step.runs_tool), Trillion runs it and admits its output itself; while it is working, read_workflow_run later; when it failed or its output did not pass, run_again: true runs it again, or send output yourself. An admitted output is kept as evidence and the reply names the next step, or, after the last step, the whole result with every step’s output. Steps go in order; sending the same admitted output again changes nothing.
Inputs
Example
Read a workflow run
read_workflow_run · Changes things
Reads one run (run_id) from any AI signed in to this account, and any tool run it is waiting on, so a tool that answered since is admitted now: the step it is on, the steps admitted, when, and what each handed back (a very large output shows its opening and read_whole_output), or the whole result when finished. With step_id too, reads that step’s whole output. Leave run_id out to list this account’s runs with the step each is on. It changes nothing except admitting a tool’s finished output.
Inputs
Example
Save a workflow
save_workflow · Changes things
Keeps a workflow this account wrote, for example from a skill file, a checklist or a process the person already follows, so any AI signed in to this account can find it and run it. Send workflow: id (“workflow:your-name”), job (plain words; find_workflow matches it), label, input_schema (JSON Schema, “type”: “object”), and steps, each with id, title, do (what to do), submit (a JSON Schema with “type”: “object” for what the step hands back) and checks. Each check has says (what it asks for, in plain words), kind, and paths: a path is a list of keys starting with “this” (this step’s output), “input” or “steps” then an earlier step’s id, where ”*” means every item of a list; of and in may also be a list of paths. Kinds: subset-of (every value at path is one of the values at of), covers (the values at path include every value at of, or every value in values), mentioned-in (every value at path appears inside some text at in), excludes-words (no text at path contains any of words), distinct-at-least (path holds at least count different values), exists (every value at path names a workflow or tool this account can find: catalog “workflow” or “tool”). A step done by a Workshop tool instead of the AI has run_tool: { tool_id, input }, where an input value { “from”: path } takes the first value there and { “all_from”: path } every value there; Trillion runs it where the tool lives when the run reaches the step, and checks its output like any other. Instead of steps, a workflow may give definition: { steps, edges }, the step graph Trillion’s workflow engine runs, for branches, steps that run at the same time and loops: a step { id, kind: “ask”, config: { name, input } } is answered by the AI (name “ai”, input { do, submit }), a Workshop tool (name its tool id) or one of this account’s agents (name “agent:<agent_id>”); an input value { “$kind”: “binding”, “values”: path } takes the first value at a path (“take”: “all” every value); a step { id, kind: “check”, config: { subject: <step id>, rules: [checks as above] } } checks another step’s output; edges are { from, to, kind: “always” }. The reply names similar_workflows that already do a job this close: improve that one instead of keeping two. describe_workflow on workflow:l078-research-to-plan shows a complete one. A refused save lists every problem at once. Saving the same id again replaces it; runs already started keep their version. It stays visible only to this account until publish_workflow.
Inputs
Example
Publish a workflow
publish_workflow · Changes things
Makes a workflow this account saved findable and runnable by every Trillion account’s AI. Their runs happen in their own accounts; nothing of this account’s runs is shared. unpublish_workflow withdraws it.
Inputs
Example
Withdraw a published workflow
unpublish_workflow · Changes things and can remove them
Withdraws a workflow this account published, so only this account finds it again. Runs other accounts already started keep their copy.
Inputs
Example