How CodyMaster Works

Complete workflow guide with diagrams, use cases, and exception handling.

---

Core Architecture

---

The RARV Execution Cycle

The heart of CodyMaster is the enhanced RARV cycle — a self-correcting autonomous execution loop:

PRE-ACT ATTENTION — The Drift Preventer

Before every action, the agent asks itself 3 critical questions:

#	Question	If NO
Q1	Does my planned action serve the Active Goal?	Return to REASON
Q2	Am I solving the original problem, not a tangent?	Return to REASON
Q3	Have I seen this error pattern before in learnings?	Apply known prevention

This single check prevents the #1 AI failure mode: goal drift. Without it, AI agents frequently solve tangential problems instead of the actual task.

---

The 6-Gate Quality System

G3 (Blind Review): Reviewer only sees the diff — no task description, no implementation context. Forces genuine code review.

G4 (Anti-Sycophancy): If G3 says "everything is fine," a Devil's Advocate pass actively hunts for hidden issues.

---

The Judge Agent Protocol

---

Working Memory System

Protocol:

1. Session Start → Read CONTINUITY.md + check learnings
2. During Work → Update current task, add completed items
3. On Error → Capture what failed + why + how to prevent
4. Session End → Save state for next session

---

Progressive Disclosure (Token Savings)

Approach	Tokens Used	Efficiency
Load all 30 skills	120,000	0% saved
Progressive Disclosure	6,300	94.7% saved

---

Common Use Cases

1. Build a New Feature (Autonomous)

# Start autonomous execution
cm continuity init
/cm-start "Build user management with list, CRUD form, and role-based access"

2. Fix a Production Bug (Manual)

# Step 1: Investigate
@[/cm-debugging]   # Root cause analysis — don't guess, investigate

# Step 2: Fix with TDD
@[/cm-tdd]          # Write test that reproduces bug → fix → verify

# Step 3: Deploy safely
@[/cm-safe-deploy]  # 8-gate pipeline with rollback strategy

3: Setup New Project (Universal Agent Bootstrap)

# Verify identity first, then bootstrap
@[/cm-identity-guard]       # Ensure right GitHub/Cloudflare account
@[/cm-project-bootstrap]    # Full setup: design system, CI, staging, deploy
                              # + auto-generates configs for selected AI agent platforms

Phase 6.5 asks which platforms your team uses:

🌐 AGENT PLATFORM SETUP
[x] AGENTS.md (Open Standard)     — Always generated
[x] Claude Desktop / Claude Code   → CLAUDE.md
[x] Cursor                         → .cursor/rules/*.mdc
[ ] OpenClaw / MaxClaw              → IDENTITY.md, MEMORY.md, TOOLS.md, SHIELD.md
[ ] OpenFang                        → HAND.toml

One AGENTS.md → all platform configs derived automatically. Never edit derived files directly.

4. Mass Translation (i18n)

# Safe multi-language extraction
@[/cm-safe-i18n]  # Extract hardcoded strings → vi.json + en.json + th.json

5. CRO & Marketing Setup

# Full conversion tracking
@[/cm-ads-tracker]    # Meta Pixel + CAPI, TikTok, Google Ads, GTM
@[/cro-methodology]   # Funnel audit + A/B test design

---

Exception Handling

❌ What if tests fail continuously?

Rule: Max 3 retries per task. After 3 failures → BLOCKED + ESCALATE to user.

❌ What if the agent drifts from the goal?

The PRE-ACT ATTENTION check catches this:

1. Agent re-reads Active Goal from CONTINUITY.md
2. If planned action doesn't serve the goal → drift logged → return to REASON
3. This happens before every action, not just at the start

❌ What if working memory is lost?

# CONTINUITY.md gets corrupted or deleted
cm continuity reset    # Reset CONTINUITY.md, learnings.json preserved
cm continuity init     # Re-create from scratch if needed

Learnings survive resets. Architecture decisions survive resets. Only the active session state is cleared.

❌ What if deploy fails?

The dashboard tracks all deployments with rollback history. Use POST /api/deployments/:id/rollback to rollback via API.

❌ What if the wrong agent is assigned?

The Judge Agent detects stuck tasks and the Dynamic Agent Selection API suggests the best agent:

curl http://codymaster.localhost:6969/api/agents/suggest?skill=cm-tdd
# → { "domain": "engineering", "agents": ["claude-code", "cursor", "antigravity"] }

❌ What if quality gate is too strict?

Gates 1-2 (static analysis + tests) are mandatory. Gates 3-6 can be adjusted:

• G3 (Blind Review): Skip only if changes are < 10 lines
• G4 (Anti-Sycophancy): Auto-triggered, cannot skip
• G5 (Security): Skip only for internal tools
• G6 (i18n): Auto-skipped if project has no i18n

---

API Reference

Method	Endpoint	Purpose
GET	/api/projects	List projects
GET	/api/tasks	List tasks
POST	/api/tasks	Create task
PUT	/api/tasks/:id/move	Move task (kanban)
POST	/api/tasks/:id/dispatch	Dispatch to AI agent
GET	/api/judge	Evaluate all tasks
GET	/api/judge/:taskId	Evaluate single task
GET	/api/agents/suggest?skill=X	Suggest best agents
GET	/api/continuity	All projects' memory
POST	/api/continuity/:id	Update memory state
GET	/api/learnings/:id	Learnings list
POST	/api/learnings/:id	Add learning
GET	/api/decisions/:id	Decisions list
GET	/api/activities	Activity history
GET	/api/deployments	Deploy history
POST	/api/deployments	Record deployment
GET	/api/changelog	Version changelog

---

Golden Rules

1. 🔒 Identity First — cm-identity-guard before push/deploy
2. 📐 Design Before Code — cm-planning always first
3. 🧪 Test Before Code — RED → GREEN → REFACTOR
4. 🛡️ PRE-ACT ATTENTION — check goal alignment every action
5. 📊 Evidence Over Claims — only trust terminal output
6. 🚀 Deploy via Gates — all 6 gates must pass
7. 🧠 Read Memory First — CONTINUITY.md at session start
8. 📚 Capture Learnings — every failure becomes wisdom
9. 🌐 One Source of Truth — AGENTS.md is the master, platform configs are derived