| name | go |
| description | Main entry point: end-to-end workflow for building features. Use /superharness:go 'requirement' to start. Orchestrates brainstorm → plan → TDD implementation → review → QA. |
Superharness Workflow
End-to-end workflow engine: brainstorm → plan → worktree → TDD implementation → dual review → QA.
Note on Recovery
Unfinished task detection is handled by the session-start hook, not by this skill. When a session starts, the hook checks .superharness/tasks/.current-task and injects task status into the session context. The AI will automatically ask the user whether to continue the unfinished task before any skill is invoked.
If the user chooses to continue an unfinished task, the AI resumes from the current phase without invoking /superharness:go. If the user chooses to start fresh, they invoke /superharness:go "new requirement" which proceeds with the normal flow below.
Normal Flow
Step 1: Brainstorm
Invoke superharness:brainstorm skill to clarify requirements and produce a design spec.
Write trace event:
echo '{"ts":"'$(date -u +%Y-%m-%dT%H:%M:%SZ)'","phase":"brainstorm","event":"start","detail":"用户输入: {requirement}"}' >> .superharness/tasks/{task}/trace.jsonl
Step 2: Plan
Invoke superharness:writing-plans skill to create implementation plan with task breakdown.
The plan creates:
.superharness/tasks/{MM}-{DD}-{name}/ directory
task.json with status, phase, sprint progress
prd.md with requirements
contract.md with Sprint Contract / Done Definition
.current-task pointer
Step 3: Worktree
Invoke superharness:using-git-worktrees skill to create an isolated development environment.
Step 4: Implement
Invoke superharness:subagent-driven-development skill to execute the plan:
- One fresh subagent per task
- TDD (Red-Green-Refactor) for each task
- Dual review: spec compliance → code quality
- Update task.json phase as work progresses
Step 5: Complete
Invoke superharness:finishing-a-development-branch skill:
- Verify all tests pass
- Generate trace-summary.md from trace.jsonl
- Present merge/PR/keep/discard options
- Clean up worktree if merging
Step 6: QA (Optional)
If the user wants QA evaluation:
- Run
/superharness:sh-qa to trigger external QA services
- If issues found: run
/superharness:sh-fix to address them
Task State Management
Throughout the workflow, keep task.json updated:
{
"name": "task-name",
"title": "Human-readable title",
"status": "planning | in_progress | review | completed",
"phase": "brainstorm | plan | implement | check | complete",
"worktree_path": "/path/to/worktree",
"sprint": {
"current": 1,
"total": 5
},
"created_at": "ISO date",
"updated_at": "ISO date"
}
Update .current-task to point to the active task directory (relative path from project root):
echo ".superharness/tasks/{MM}-{DD}-{name}" > .superharness/tasks/.current-task
Trace Logging
At each major transition, append to trace.jsonl:
echo '{"ts":"'$(date -u +%Y-%m-%dT%H:%M:%SZ)'","phase":"{phase}","event":"{event}","detail":"{detail}"}' >> .superharness/tasks/{task}/trace.jsonl
Key events to log:
brainstorm:start, brainstorm:spec_confirmed
plan:tasks_created, plan:worktree_created
implement:task_start, implement:tdd_red, implement:tdd_green, implement:task_complete
check:spec_review, check:code_review
qa:qa_start, qa:qa_result
fix:fix_start, fix:fix_complete
Red Flags
- Never skip brainstorm phase, even for "simple" tasks
- Never start implementation without a plan
- Never work on main/master without creating a worktree first
- Never skip the dual review (spec + quality)
- Never declare completion without running tests