execute

Workflow Execute

Start the dispatcher. For every task: worker builds → reviewer validates → PASS marks done, FAIL retries, max retries exceeded escalates to user.

Reviewer gate

After each worker output passes syntax verification:

1. If reviewers: is configured in workflow.yaml (or per-task reviewer: field in the task JSON), a reviewer agent is called with a domain-appropriate system prompt
2. Domain is auto-detected from task description: code-reviewer, editor, researcher, designer, data-scientist, or devops
3. Reviewer responds with PASS or FAIL on the first line, followed by specific feedback
4. On FAIL: feedback is appended to the retry prompt — worker tries again (up to max_retries)
5. After max_retries FAILs: task is escalated — user sees reviewer's final feedback and chooses: skip / retry with guidance / redesign
6. On PASS (or no reviewer configured): task is marked completed and committed
7. If reviewer LLM is unreachable: treated as PASS to avoid blocking the pipeline

Step 1: Check prerequisites

python3 -c "
from workflow_kit.runtime.state import load_state
from pathlib import Path
s = load_state(Path('.'))
print('phase:', s.phase)
" && \
ls workflow/tasks/pending/*.json 2>/dev/null | wc -l | xargs -I{} echo "pending tasks: {}"

• Phase not plan or execute: warn with correct next skill.
• 0 pending tasks: "Run /workflow-kit:plan first."

Step 2: Handle --stop

If user passes --stop or says "stop":

python -m workflow_kit stop

Done.

Step 3: Check if already running

cat workflow/dispatcher.pid 2>/dev/null | xargs -I{} kill -0 {} 2>/dev/null && echo "running" || echo "stopped"

If running: "Dispatcher already running. Use /workflow-kit:status."

Step 4: Advance state to 'execute'

from workflow_kit.runtime.state import load_state, save_state
from pathlib import Path
s = load_state(Path('.'))
if s.phase == "plan":
    s.transition_to("execute")
    save_state(s, Path('.'))
    print("Phase: plan → execute")

Step 5: Start dispatcher

python -m workflow_kit execute &
echo $! > workflow/dispatcher.pid

Step 6: Report

✓ Dispatcher started (PID N)

  Loop: worker builds → reviewer validates → PASS/FAIL
  Max retries: <from workflow.yaml>

  Monitor:
    /workflow-kit:status   — inline progress
    /workflow-kit:monitor  — live dashboard

  When all tasks complete:
    /workflow-kit:synthesize  — package and review deliverables

Reviewer escalation

When a task fails review max_retries times, the dispatcher pauses and prints:

⚠ Task <id> failed review N times
  Reviewer (<domain>): <feedback summary>

  A) Skip this task
  B) Retry with your guidance (describe what to fix)
  C) Redesign the task entirely

Respond inline — dispatcher waits before continuing.