Workflow Skill (v6)

Workflow Skill (v6)

Facilitator-rubric orchestration with hard quality gates.

Decision Engine — wicked-garden:propose-process

User project description
        │
        ▼
propose-process facilitator
  ├── Score 9 risk factors (0-3 each)
  ├── Read agents/**/*.md frontmatter → pick specialists
  ├── Pick phases from phases.json catalog
  ├── Set rigor_tier (minimal / standard / full)
  └── Emit process-plan.md + full task chain

All phase selection is judgment-driven by the facilitator, not rule-based.

Interaction Modes

Mode	How	Effect
normal	Default	Each phase gate requires explicit user approval before advancing
yolo	/wicked-garden:crew:yolo	Auto-advance through gates; gates still run, findings still logged

Rigor Tiers

Tier	Complexity	Gates	Reviewers
minimal	0-2	Advisory — gate findings logged but do not block	Single reviewer
standard	3-5	Enforced — REJECT blocks advancement	Single reviewer
full	6-7	Enforced — REJECT blocks; multi-reviewer	2+ reviewers

Security or compliance signals override to full regardless of complexity.

Phase Plan

clarify → design → [challenge?] → [test-strategy?] → build → test → review

Brackets = optional phases. The facilitator picks which phases apply based on factor readings. phases.json defines gate config per phase (min_gate_score, valid_skip_reasons, depends_on).

Phase Summary

Phase	Goal	Key deliverable
clarify	Define success criteria	Outcome statement + success criteria
design	Architect solution	Architecture docs + approach
challenge	Adversarial stress-test	Challenge findings + revised design
test-strategy	Define test approach pre-build	Test scenarios + risk assessment
build	Implement	Working implementation
test	Verify	Test results + convergence evidence
review	Multi-perspective validation	Review findings + sign-off

Gate Enforcement (v2.5.0+)

Gates are hard enforcement — not advisory — at standard and full rigor.

Verdict	Effect
APPROVE	Phase advances
CONDITIONAL	conditions-manifest.json written; verify before next phase
REJECT	Blocks advancement; mandatory rework

CONDITIONAL auto-resolution (AC-4.4): spec gap conditions → fixed inline. Intent-changing conditions → escalate to user or council.

Gate reviewer assignment happens at approve time, not at phase start. Banned reviewer values: just-finish-auto, fast-pass, auto-approve-*.

Build depends on design (depends_on: ["clarify", "design"]). To migrate legacy beta.3 projects, use /wicked-garden:adopt-legacy.

Specialist Discovery

Crew discovers specialists by reading agents/**/*.md frontmatter directly at runtime. No static enhances map. Fallback agents (facilitator, researcher, implementer, reviewer) handle phases when no specialist matches.

Convergence Tracking (build/test phases)

Artifact states: Designed → Built → Wired → Tested → Integrated → Verified. The convergence-verify gate flips REJECT → APPROVE only when every tracked artifact reaches Integrated. Stalls at threshold 3 sessions surface as findings.

Commands Reference

Command	Purpose
/wicked-garden:crew:start	Begin project — invokes propose-process
/wicked-garden:crew:status	View current phase and engaged specialists
/wicked-garden:crew:execute	Run current phase
/wicked-garden:crew:approve	Advance phase after gate
/wicked-garden:crew:just-finish	Autonomous completion (yolo-equivalent)
/wicked-garden:crew:gate	Run a specific quality gate
/wicked-garden:crew:evidence	Query evidence for a task
/wicked-garden:crew:yolo	Switch to auto-advance mode

Storage

Store	What
Native tasks	TaskCreate/TaskUpdate with validated metadata (see scripts/_event_schema.py)
wicked-garden:mem	Cross-session learning at project completion and gate failures
Local JSON	DomainStore fallback; always available

Refs (load on demand)

• Evidence Tracking — evidence tiers L1-L4, artifact naming
• Integration Details — mem, search, wicked-bus usage
• Scoring Rubric — HISTORICAL v5 only; see propose-process for v6
• Risk Dimension Signals — signal keywords per factor