program-execution-workflow

name	program-execution-workflow
description	Pure handoff router for DrowAI programs from `.cursor/agents/program-workflow-state.md`. Discovers all numbered phase files under doc_root and runs the full create→review→implement→final-review cycle for every remaining item in one session until program_complete or hard_stop. Does not stop after one phase. Does not create, review, or implement itself.

Program Execution Workflow

Router only. Runs the full multi-phase program in one invocation. Copies state, discovers files, triggers child flows.

Outer loop contract (strict)

One user trigger = process every remaining phase file until done.

WHILE pipeline_stage != program_complete AND hard_stop is null:
  run current stage (handoff → trigger child → read output → update pipeline_stage)
  IF stage was advance_queue AND more files remain:
    IMMEDIATELY continue — do NOT return to user
END
ONLY THEN return final response

MUST	MUST NOT
Continue after advance_queue when more files exist	Stop after one phase cycle
Re-enter skill loop in the same session after each child flow	Ask user to confirm next phase
Set pipeline_stage: creating_guide and discover next file automatically	Exit with "phase N done" while files remain
Stop only on program_complete, hard_stop, or idle	Treat child skill "final response" as router stop

Proceed automatically. At every handoff, trigger the next child flow immediately. Child flows run to completion for the current phase; control always returns to this router until the program finishes.

User configures

Field	Purpose
doc_root	Directory with numbered phase docs
file_glob	Discovery pattern (default phase-*.md)
last_completed_index	Last fully finished item (0-based). -1 = none done
guide_mode	refactor or feature creator
pipeline_stage	Resume point

Discovery (each phase)

sorted = sort(glob(doc_root + file_glob))
current_index = last_completed_index + 1
IF current_index >= len(sorted) → pipeline_stage = program_complete; STOP loop
current_input_doc = sorted[current_index]

Persist current_index and current_input_doc in program-workflow-state when starting a phase.

Stage machine (one phase)

pipeline_stage	Trigger	On success →
creating_guide	guide creator	reviewing_guide
reviewing_guide	implementation-guide-review-loop	implementing
implementing	feature-implementation-workflow	reviewing_implementation
reviewing_implementation	implementation-review-loop (final)	advance_queue
advance_queue	(router only)	creating_guide or program_complete

Handoffs: state-handoffs.md.

After advance_queue (critical)

1. last_completed_index += 1
2. Rediscover sorted from doc_root
3. If last_completed_index + 1 < len(sorted):
   • Clear current_guide, current_input_doc
   • Set pipeline_stage: creating_guide
   • Immediately run creating_guide for the next file — same session, no user prompt
4. Else: pipeline_stage: program_complete → exit loop

Hard stops (only valid stop)

Set hard_stop from child state on NEEDS_CLARIFICATION or MAX_ROUNDS_REACHED. Do not advance last_completed_index. Stop loop and report.

Hard rules

• Never stop after one phase if more discovered files remain.
• Never use README as phase input unless it is the file at current_index.
• Never list phase paths in state — discover from doc_root.
• Never create, review, or implement in this skill.

Final response (only when loop exits)

Report: program_complete or hard_stop, last_completed_index, phases processed this run, doc_root.