Workflow Builder

Workflow Builder Skill

Use this skill when the user wants to compose multiple agents into a repeatable workflow such as draft → review → judge, judge-and-retry, parallel reviews followed by a merge, or any other bounded coordination of agents.

This skill authors and operates workflows through the OpenColab CLI. It is not the workflow engine itself. The CLI/runtime owns parsing, execution, persistence, and human gate handling.

Source of truth

• Workflow definitions live at: projects/<project_id>/workflows/<workflow_id>/workflow.xml
• Workflow runs live at: projects/<project_id>/workflows/<workflow_id>/runs/<run_id>/
• Each run folder contains RUN.md (with YAML frontmatter), state.json, status.json, events.jsonl, inputs/initial.md, and steps/<step_id>/.

Never edit state.json, status.json, or events.jsonl by hand. They are written by the runtime.

Allowed step types

• agent — runs one project agent with a generated prompt. Required: id, type="agent", agent, <prompt>. Optional: <output name="..." />.
• decision — runs one project agent and expects a strict <workflow-decision action="..." next="..." reason="..." /> line in the response. Required: id, type="decision", agent, <prompt>, <choices>.
• human_gate — pauses the run for human input. Required: id, type="human_gate", <prompt>. Optional: allow="approve,stop,retry,edit,branch".
• merge — concatenates prior outputs into a single named value. Required: id, type="merge", at least one <input name="..." />, <output name="..." />.
• terminate — explicitly ends the run with a status of success, failed, or stopped.

Loops must be bounded

Every <loop> requires at least one of:

• maxIterations="<n>"
• maxSteps="<n>"
• maxRuntimeMinutes="<n>"

Unbounded loops are rejected during validation.

Prompt references

Inside any <prompt> you can use template variables:

• ${input.<name>} — values supplied when starting the run.
• ${<output_name>} — named outputs from prior steps (<output name="..." />).
• ${<step_id>.output> — fallback reference for any step's raw output text.

Author a workflow

1. Decide which project agents will play each role (drafter, reviewer, judge).
2. Pick a template:
   • blank for an empty scaffold
   • review-loop for draft → review → judge with a bounded loop
   • judge-and-retry for draft → judge → retry-or-stop
3. Run:

   opencolab workflow create --workflow-id <id> --from review-loop
   

4. Edit the generated projects/<project_id>/workflows/<id>/workflow.xml to reference real project agents, set the loop bound, and tighten prompts.
5. Validate before running:

   opencolab workflow validate --workflow-id <id>
   

Run a workflow

opencolab workflow run --workflow-id <id> --input "<task text>"
# or
opencolab workflow run --workflow-id <id> --input-file inputs.json --wait true

--input <text> is shorthand for { "task": "<text>" }. For richer inputs, pass --input-file <path> or --input-json '{"foo":"bar"}'.

--wait true blocks the CLI until the run reaches a terminal or paused state and streams events as they happen.

Observe a run

• Latest status snapshot:

  opencolab workflow status --run-id <runId>
  

• Tail events:

  opencolab workflow logs --run-id <runId> --follow
  

• List all runs:

  opencolab workflow runs
  

Stop, resume, approve

• opencolab workflow stop --run-id <runId> — cooperative stop.
• opencolab workflow resume --run-id <runId> — resume a paused run.
• opencolab workflow approve --run-id <runId> --decision continue|stop|retry|branch:<step>|edit — record a human gate decision. Use --values-json '{"task":"..."}' with --decision edit to update inputs before continuing.

Decision contract for decision steps

Decision agents must end their response with a single line that looks like:

<workflow-decision action="continue" next="draft" reason="The answer still misses two constraints." />

• action: continue, stop, branch, needs_human, or fail.
• next: target step id when action="continue" or action="branch".
• reason: short justification; surfaced in events.jsonl and Studio.

You may use <choice name="..." /> blocks under <choices> to give the agent a fixed menu. The agent can pick a choice by name via:

<workflow-decision action="continue" choice="continue" />

If the decision block is missing or invalid, the run pauses for human review by default (or fails if onInvalid="fail" was declared on the step).

Memory isolation

Workflow steps do not pollute normal agent chat memory. The runtime uses a workflow-scoped session identity for every agent call and surfaces results through the workflow's run folder, not the agent's memory/Session/.

Things to avoid

• Do not author workflows outside the project workflows directory.
• Do not invent step types beyond the supported list — the validator rejects unknown tags and unsupported step type attributes.
• Do not skip loop bounds. Bounded loops are mandatory.
• Do not pretend this skill executes workflows. Always shell out to the CLI.