| name | harness-workflow-runner |
| description | Use when operating an existing Agent Harness CLI workflow-controller project: start from a user task or idea, read project AGENTS.md/workflow state, produce active-node artifacts, handle step/options/choose/approve/reject, and continue until the workflow completes. |
Harness Workflow Runner
Use this skill to run an existing agent-harness workflow-controller project.
Do not redesign the workflow, checks, or hooks unless the user explicitly asks.
Start From The Project
When the user gives a plain task or idea, first read the project instructions:
AGENTS.md
- workflow spec referenced there, usually
workflows/*.json
- artifact contract docs referenced there
If state exists, continue from the current active node. If no state exists, use
the workflow's initial node.
Operating Loop
- Inspect current state with
agent-harness status --state <state-path> when a
state file exists.
- Read the active node's artifact contract.
- Create or update only the artifact(s) required by the active node.
- Run the workflow step, or stop normally and let the project Stop hook run it:
agent-harness step --task <workflow.json> --hook-json
- If the step blocks on failed checks, read the reason/report and repair the
active artifact.
- If the workflow advances, follow the next-node prompt.
- Continue until
step returns systemMessage instead of decision: "block".
Choosing Transitions
When state is choosing, do not guess from memory. Inspect the choices:
agent-harness options --state <state-path>
Then apply the transition that best matches the evidence:
agent-harness choose <transition-id> --state <state-path> --reason "<why this route is correct>"
The reason should cite the artifact evidence or failed/passed gate condition.
Waiting For User
When state is waiting, identify what user authority is needed. Ask the user a
short concrete question. After the user decides, use:
agent-harness approve <node-id> --state <state-path> --reason "<reason>"
agent-harness reject <node-id> --state <state-path> --reason "<reason>"
Then run step again to apply the transition.
Rules
- Respect the workflow's allowed statuses and domain routing rules.
- Do not pick the fastest advancing status unless the evidence justifies it.
- Do not edit workflow specs or check scripts during normal operation.
- Treat reports and state files as generated runtime data unless the project says
otherwise.
- Keep user prompts short; the workflow and project
AGENTS.md carry the
protocol.