agentic-workflow / rules

---
description: "AgenticWorkflow Methodology — Applies to all tasks in this repository"
alwaysApply: true
---

# AgenticWorkflow — Cursor Directive

> All AI agents working on this project must follow the AgenticWorkflow methodology.

## Essential References

All Absolute Criteria, design principles, and workflow structures are defined in `AGENTS.md`.
**You must read `AGENTS.md` and adhere to its rules.**
For architectural design and philosophy, refer to `AGENTICWORKFLOW-ARCHITECTURE-AND-PHILOSOPHY.md`.
For historical design decisions, refer to `DECISION-LOG.md` (ADR).

## Genetic Design (DNA Inheritance)

This project is a parent organism that births child agentic workflow systems.
Every child system structurally embeds the parent genome: 3 Absolute Criteria, Single-File SOT, 4-Layer Verification, Safety Hooks, and Memory Systems.
Details: `soul.md`, `AGENTS.md §1`.

## Absolute Criteria (Core Summary)

### Absolute Criterion 1: Quality of the Final Deliverable
> Speed, token cost, workload, and length limits are completely ignored.
> The sole criterion for every decision is the **quality of the final deliverable**.

### Absolute Criterion 2: Single-File SOT
> All shared state is concentrated in a single file (`state.yaml`). Write permission is held exclusively by the Orchestrator / Team Lead. Parallel agents concurrently modifying the same file is strictly prohibited.

### Absolute Criterion 3: Code Change Protocol (CCP)
> Before writing, modifying, adding, or deleting code, you must internally perform 3 steps:
> Step 1: Understand Intent → Step 2: Ripple Effect Analysis → Step 3: Change Plan.
> Analysis depth scales proportionally with change scope (Minor: Step 1 only, Standard: full 3 steps, Large-scale: full 3 steps + mandatory user approval).
> **Coding Anchor Points (CAP-1~4)**: Think Before Coding, Simplicity First, Goal-Based Execution, Surgical Changes. Details: `AGENTS.md §2`.

## Basic Workflow Structure

Every workflow consists of three stages: **Research** → **Planning** → **Implementation**.

## Cursor Implementation Mapping

| AgenticWorkflow Concept | Cursor Implementation |
|---|---|
| Specialized Agent | Role specification via Cursor Agent Mode |
| Automated Verification | Verification rule files in `.cursor/rules/` or external Python scripts |
| Reusable Modules | Domain-specific rule files in `.cursor/rules/` |
| SOT State Management | `state.yaml` file — single write point principle applies identically (`state.yaml`, `state.yml`, `state.json`) |
| Autopilot Mode | Controlled by `autopilot.enabled` field in SOT. Auto-approves `(human)` review steps. Anti-Skip Guard verifies deliverable existence + 100 bytes minimum. See `AGENTS.md §5.1`. |
| ULW (Ultrawork) Mode | Activated when `ulw` is in prompt. Thoroughness intensity overlay. 3 Intensifiers: Sisyphus Persistence (3 retries), Mandatory Task Decomposition, Bounded Retry Escalation. See `AGENTS.md §5.1.1`. |
| Verification Protocol | Verifies 100% functional goal achievement of step deliverables. Verification Gate layer sitting atop physical Anti-Skip Guard. Retries up to 10 times (15 with ULW). See `AGENTS.md §5.3`. |
| pACS (Self-Confidence Scoring) | 3-dimensional self-evaluation (Faithfulness, Completeness, Logic) with mandatory Pre-mortem protocol and min-score principle. GREEN (>=70): proceed, YELLOW (50-69): flag and proceed, RED (<50): rework. See `AGENTS.md §5.4`. |
| Adversarial Review (Enhanced L2) | Independent evaluation replacing legacy calibration. `@reviewer` (critical analysis of code/deliverables) + `@fact-checker` (external fact verification). P1 validation (`validate_review.py` R1-R5) ensures review rigor. See `AGENTS.md §5.5`. |
| Terminology Protocol | Maintains terminology consistency via `translations/glossary.yaml`. P1 validation (`validate_translation.py` T1-T9, `validate_verification.py` V1a-V1c). |
| Predictive Debugging (L-1) | Pre-tool warning on risky files based on error history. `predictive_debug_guard.py` (PreToolUse warning) + `aggregate_risk_scores()` (SessionStart P1 aggregation) + `validate_risk_scores()` (RS1-RS6 validation). Cached in `risk-scores.json`. |

## Context Preservation

In environments without automatic Claude Code hook dispatching:
- **Notepad Feature**: Record working progress in Cursor Notepad.
- **SOT-Based Recovery**: Restore workflow state by reading `state.yaml`.
- **Manual Snapshot**: Save progress to markdown snapshot files at key milestones.

## Core Design Principles

- **P1**: Remove noise via deterministic scripts before passing context to AI.
- **P2**: Maximize quality through specialized delegation.
- **P3**: Explicit paths for all resources; no placeholders.
- **P4**: User questions: maximum 4, ~3 options each.

## Language and Style Rules

- **Framework Documentation & User Dialogue**: Pure English
- **Workflow Execution**: Pure English (maximizes reasoning performance — Absolute Criterion 1)
- **Deliverables**: English primary deliverables
- **Technical Terminology**: Keep standard English terms (SOT, Agent, Orchestrator, Hooks, etc.)
- **Visualization**: Prefer clean Mermaid diagrams with quoted labels
- **Depth**: Comprehensive, data-backed exposition over superficial summaries