run-harness

Run Harness

You are executing a harness against a workflow. The harness defines how to work; the workflow defines what to work on.

Step 1: Load Harness and Workflow

Parse $ARGUMENTS to get the harness name and workflow name. If not provided or ambiguous, ask.

Read:

• .claude/harnesses/<harness-name>.yaml — the harness definition
• .claude/skills/<workflow-name>/context.yaml — workflow context
• .claude/skills/<workflow-name>/tasks.yaml — workflow tasks

If either file doesn't exist, tell the user and suggest how to create it.

Validate the harness — verify it has steps (non-empty list), constraints (list), done_criteria (list), and pause_between_tasks (boolean). If any required field is missing or malformed, report the issue and stop.

Step 2: Identify the Target Task

Find the next actionable task from tasks.yaml:

1. First, any task with status: in-progress (resume it)
2. If none, the first task with status: todo whose dependencies (if any) are all done
3. Skip any task with status: blocked that has a blocked_by reason (these need external resolution)
4. If no actionable tasks remain, report that the workflow is complete

Present the target task and the harness that will be applied. Ask for confirmation before proceeding.

If the user declines the suggested task, ask which task they'd prefer to work on, or whether to mark the current task as blocked.

Step 3: Execute Harness Steps

For each step defined in the harness, in order:

1. Announce the step — tell the user which step you're executing and what it involves
2. Execute the instruction — follow the step's instruction field, applying it to the current task
3. Respect constraints — check all harness constraints throughout execution
4. Report progress — after each step, briefly state what was done

Use the workflow's context.yaml learnings and decisions as background knowledge throughout.

Step 4: Verify Done Criteria

Before marking the task complete, check each item in the harness's done_criteria:

• If all pass, proceed to mark done
• If any fail, report which criteria failed and what needs to be addressed

Step 5: Update Workflow State

After the task is complete:

1. Set the task's status to done in tasks.yaml and add completion notes
2. Append any new learnings discovered during execution to context.yaml
3. Record any decisions that were made in context.yaml

Step 6: Next Task Decision

If the harness has pause_between_tasks: true:

• Present updated workflow status and ask if the user wants to continue to the next task

If pause_between_tasks: false:

• Automatically loop back to Step 2 for the next actionable task
• Continue until all tasks are done or a blocker is hit

Rules

• Always ask for confirmation before starting the first task
• If a step fails or produces unexpected results, pause and ask the user rather than continuing blindly
• If you discover something that contradicts workflow context, flag it before proceeding
• Update tracking files after each task, not at the end of the run
• If a task turns out to be more complex than described, suggest breaking it into subtasks rather than doing unscoped work