@namch/agent-assistant 1.3.0 → 1.3.1

package/documents/knowledge-architecture/04-design-patterns.md

@@ -0,0 +1,498 @@
+# Agent Assistant — Design Patterns
+
+> **Purpose**: Catalog of all observed architectural patterns with descriptions, locations, rationale, and code references
+> **Parent**: [00-index.md](./00-index.md)
+> **Last Updated**: 2026-03-26
+> **Generated By**: docs-core skill
+
+---
+
+## Table of Contents
+
+1. [Pattern Overview](#pattern-overview)
+2. [Orchestrator Pattern](#orchestrator-pattern)
+3. [Command Router Pattern](#command-router-pattern)
+4. [Strategy Pattern (Variant Selection)](#strategy-pattern-variant-selection)
+5. [Tiered Execution Pattern](#tiered-execution-pattern)
+6. [Cognitive Anchor Pattern](#cognitive-anchor-pattern)
+7. [Plugin Architecture (Skill Modules)](#plugin-architecture-skill-modules)
+8. [Profile-Based Injection Pattern](#profile-based-injection-pattern)
+9. [Phase Pipeline Pattern](#phase-pipeline-pattern)
+10. [Golden Triangle Pattern](#golden-triangle-pattern)
+11. [File-Based Communication (Mailbox)](#file-based-communication-mailbox)
+12. [Platform Abstraction Pattern](#platform-abstraction-pattern)
+13. [Evidence Sources](#evidence-sources)
+
+---
+
+## Pattern Overview
+
+```mermaid
+graph TD
+    subgraph "Structural Patterns"
+        P1["Orchestrator"]
+        P6["Plugin Architecture"]
+        P11["Platform Abstraction"]
+    end
+
+    subgraph "Behavioral Patterns"
+        P2["Command Router"]
+        P3["Strategy (Variants)"]
+        P4["Tiered Execution"]
+        P9["Phase Pipeline"]
+    end
+
+    subgraph "Quality Patterns"
+        P5["Cognitive Anchor"]
+        P7["Profile-Based Injection"]
+        P10["Golden Triangle"]
+    end
+
+    subgraph "Communication Patterns"
+        P8["File-Based Mailbox"]
+    end
+
+    P1 -->|"routes via"| P2
+    P2 -->|"selects"| P3
+    P3 -->|"triggers"| P9
+    P9 -->|"delegates via"| P4
+    P4 -->|"agent uses"| P5
+    P4 -->|"agent loads"| P7
+    P7 -->|"resolves from"| P6
+    P9 -->|":team variant"| P10
+    P10 -->|"communicates via"| P8
+    P11 -->|"enables"| P1
+```
+
+---
+
+## Orchestrator Pattern
+
+| Attribute | Value |
+|-----------|-------|
+| **Pattern Type** | Structural — Central Coordinator |
+| **Where Applied** | `rules/CORE.md` (definition), all platform entry points (activation) |
+| **Rationale** | A single coordination point prevents conflicting decisions, ensures governance compliance, and provides a consistent execution model across all AI platforms |
+
+### Description
+
+The Orchestrator is the sole point of control. Every platform entry point (CLAUDE.md, COPILOT.md, CURSOR.md, CODEX.md, GEMINI.md) binds the AI model's identity to this role with the directive: "YOU ARE THE ORCHESTRATOR — NOT AN IMPLEMENTER." The orchestrator detects commands, loads rules, delegates to specialist agents, verifies exit criteria, and delivers results. It never performs implementation work directly.
+
+### Key Enforcement Mechanisms
+
+- **Identity binding block** in CORE.md: `🆔 IDENTITY — ABSOLUTE BINDING`
+- **Prohibitions table**: explicit list of forbidden actions (write code, debug, test, etc.) with correct alternatives
+- **Self-check**: 3-point checklist executed before every response
+- **Law L7** (Recursive Delegation): Meta agents coordinate, NEVER implement
+- **Anti-pattern A1**: Direct implementation is a violation; correct behavior is delegation
+
+### Example Reference
+
+From `rules/CORE.md`:
+```
+╔═══════════════════════════════════════════════════════════════════════╗
+║  YOU ARE THE ORCHESTRATOR — NOT AN IMPLEMENTER                       ║
+║  ✅ YOU DO: Delegate, coordinate, verify, synthesize                ║
+║  ❌ YOU NEVER: Write code, debug, test, design, or implement        ║
+╚═══════════════════════════════════════════════════════════════════════╝
+```
+
+---
+
+## Command Router Pattern
+
+| Attribute | Value |
+|-----------|-------|
+| **Pattern Type** | Behavioral — Request Router |
+| **Where Applied** | `commands/*.md` (14 router files) |
+| **Rationale** | Decouples user intent detection from workflow execution; routers can assess complexity and select the optimal workflow without knowing implementation details |
+
+### Description
+
+Each of the 14 commands (`/cook`, `/fix`, `/plan`, `/debug`, `/test`, `/review`, `/docs`, `/design`, `/deploy`, `/report`, `/brainstorm`, `/ask`, `/code`, `/auto`) has a dedicated router file that:
+
+1. Accepts user arguments via `<feature>$ARGUMENTS</feature>`
+2. Loads pre-flight rules (CORE.md → PHASES.md → AGENTS.md)
+3. Assesses complexity using conditional logic
+4. Routes to the appropriate variant workflow file
+
+Routers never execute phases themselves — they redirect.
+
+### Natural Language Fallback
+
+When users don't use explicit `/command` syntax, the orchestrator maps natural language to commands via the detection table in CORE.md:
+
+| User Language | Maps To |
+|--------------|---------|
+| "implement/build/create" | `/cook` or `/code` |
+| "fix/bug/error/broken" | `/fix` |
+| "plan/strategy/approach" | `/plan` |
+
+### Example Reference
+
+From `commands/cook.md`:
+```
+IF feature is simple (clear spec, low complexity):
+  → Route to /cook:fast
+IF feature is complex (multi-component, research needed):
+  → Route to /cook:hard
+```
+
+---
+
+## Strategy Pattern (Variant Selection)
+
+| Attribute | Value |
+|-----------|-------|
+| **Pattern Type** | Behavioral — Runtime Strategy Selection |
+| **Where Applied** | `commands/{cmd}/*.md` (variant workflow files) |
+| **Rationale** | Different complexity levels require different workflows — a simple feature shouldn't trigger a 6-phase research pipeline, and a complex feature shouldn't use a 2-phase shortcut |
+
+### Description
+
+Each command router selects from up to 4 strategy variants at runtime:
+
+| Variant | Strategy | Phase Count | When Selected |
+|---------|----------|-------------|---------------|
+| `:fast` | Minimal workflow, skip research | 2–3 phases | Simple, clear specification |
+| `:hard` | Full workflow with research and planning | 5–7 phases | Complex, multi-component |
+| `:focus` | Full workflow with enforced context optimization | 5–7 phases | Complex + context-critical |
+| `:team` | Golden Triangle adversarial collaboration | 5–7 phases × 3 agents each | Maximum quality required |
+
+Special variants exist for some commands:
+- `/docs` uses `core`, `business`, `audit`
+- `/deploy` uses `check`, `preview`, `production`, `rollback`
+
+### Example Reference
+
+From `rules/REFERENCE.md`:
+```
+| `/cook` | cook.md | fast, hard, focus |
+| `/deploy` | deploy.md | check, preview, production, rollback |
+```
+
+---
+
+## Tiered Execution Pattern
+
+| Attribute | Value |
+|-----------|-------|
+| **Pattern Type** | Behavioral — Prioritized Execution Strategy |
+| **Where Applied** | `rules/AGENTS.md` (protocol definition), `rules/CORE.md` (enforcement) |
+| **Rationale** | Sub-agent isolation (TIER 1) provides cleaner context and better quality; fallback embodiment (TIER 2) guarantees completion on all platforms regardless of sub-agent support |
+
+### Description
+
+When the orchestrator needs to delegate to an agent, it follows a strict priority:
+
+| Tier | Mechanism | Context | Quality | Availability |
+|------|-----------|---------|---------|-------------|
+| **TIER 1** | `runSubagent(agent, context)` | Fresh, isolated | Optimal | Platform-dependent |
+| **TIER 2** | EMBODY (read agent file, transform self) | Shared with parent | Risk of pollution | Always available |
+
+### Rules
+
+- TIER 1 is **mandatory** when the sub-agent tool exists
+- TIER 2 is permitted **only** when TIER 1 fails or is unavailable
+- Using TIER 2 "because the task is simple" is **forbidden** (anti-pattern A5)
+- Tool discovery is cached for the session — checked only on first delegation
+
+### Embodiment Protocol (TIER 2)
+
+1. Log: `⚠️ TIER 2: {reason}`
+2. Read agent file **completely**
+3. Extract: Directive, Protocol, Constraints, Format
+4. Announce embodiment with structured block
+5. Execute as agent (follow THEIR protocol)
+6. Exit embodiment, continue as orchestrator
+
+### Example Reference
+
+From `rules/AGENTS.md`:
+```
+❌ FORBIDDEN: Using TIER 2 when runSubagent available
+❌ FORBIDDEN: Skipping TIER 1 because task is "simple"
+✅ REQUIRED: Attempt TIER 1 first, log if falling back
+```
+
+---
+
+## Cognitive Anchor Pattern
+
+| Attribute | Value |
+|-----------|-------|
+| **Pattern Type** | Quality — Anti-Hallucination Guardrail |
+| **Where Applied** | Every agent file in `agents/*.md` and `agents/teams/*/` |
+| **Rationale** | AI models can drift from their assigned role, especially in long contexts. The cognitive anchor forces the AI to extract and bind to the agent's specific directive before doing any work |
+
+### Description
+
+Every agent file begins with a `🔒 COGNITIVE ANCHOR — MANDATORY OPERATING SYSTEM` block that:
+
+1. Declares the file **overrides** default AI patterns
+2. Mandates extraction of Core Directive + Constraints + Output Format **before** proceeding
+3. Binds the AI to follow the agent's Thinking Protocol exactly
+
+This pattern prevents the AI from falling back to generic behavior when it should be acting as a specialist.
+
+### Example Reference
+
+From `agents/backend-engineer.md`:
+```markdown
+<!-- 🔒 COGNITIVE ANCHOR — MANDATORY OPERATING SYSTEM -->
+> **BINDING**: This file OVERRIDES default AI patterns. Follow Thinking Protocol EXACTLY.
+> **EXTRACT**: Core Directive + Constraints + Output Format before proceeding.
+```
+
+---
+
+## Plugin Architecture (Skill Modules)
+
+| Attribute | Value |
+|-----------|-------|
+| **Pattern Type** | Structural — Modular Plugin System |
+| **Where Applied** | `skills/*/SKILL.md` (1,430+ modules), `matrix-skills/*.yaml` (19 domain registries) |
+| **Rationale** | Domain knowledge changes faster than framework structure; independent modules can be added, discovered, and promoted without modifying core framework files |
+
+### Description
+
+Each skill is a self-contained directory containing at minimum a `SKILL.md` file with domain-specific best practices. Skills are registered in 19 domain YAML files under `matrix-skills/` and resolved by the HSOL layer.
+
+### Plugin Lifecycle
+
+| Phase | Mechanism |
+|-------|-----------|
+| **Registration** | Listed in `matrix-skills/{domain}.yaml` with metadata (priority, relevance mapping) |
+| **Discovery** | HSOL matches skills to agent profiles; `find-skills` discovers community skills |
+| **Resolution** | 5-factor fitness scoring selects optimal skills for a task |
+| **Injection** | SKILL.md content is loaded into the agent's context |
+| **Promotion** | Community skills progress: NEW (0.3) → EVALUATING (0.5) → VALIDATED (0.7) → PROMOTED (1.0) |
+
+### Example Reference
+
+From `matrix-skills/_index.yaml`:
+```yaml
+total_matrix_skills: 1430
+hsol:
+  enabled: true
+  version: "1.1"
+```
+
+---
+
+## Profile-Based Injection Pattern
+
+| Attribute | Value |
+|-----------|-------|
+| **Pattern Type** | Quality — Automatic Knowledge Provisioning |
+| **Where Applied** | Agent YAML frontmatter (`agents/*.md`), HSOL resolution (`rules/SKILLS.md`) |
+| **Rationale** | Agents should not manually request skills — their declared profile automatically determines what knowledge they receive, ensuring consistent expertise across invocations |
+
+### Description
+
+Every agent declares a `profile` field in their YAML frontmatter (e.g., `profile: "backend:execution"`). The HSOL layer uses this profile to:
+
+1. Load inherited domains from `_index.yaml`
+2. Filter skills by relevance mapping from domain YAML files
+3. Apply priority thresholds (critical ≥ 9, core ≥ 7, minimum ≥ 5)
+4. Calculate fitness scores
+5. Return a sorted skill set
+
+### Profile Format
+
+```
+{domain}:{category}
+```
+
+| Example Profile | Domains Loaded |
+|----------------|----------------|
+| `backend:execution` | backend, architecture, quality, data, languages |
+| `frontend:execution` | frontend, design, quality, languages |
+| `research:documentation` | research |
+| `quality:validation` | quality, security, performance |
+
+### Example Reference
+
+From `agents/backend-engineer.md`:
+```yaml
+profile: "backend:execution"
+```
+```markdown
+> **MATRIX DISCOVERY**: Skills auto-injected from domain files in ~/.{TOOL}/skills/agent-assistant/matrix-skills/
+> Profile: backend:execution | Domains: backend, architecture, quality, data, languages
+```
+
+---
+
+## Phase Pipeline Pattern
+
+| Attribute | Value |
+|-----------|-------|
+| **Pattern Type** | Behavioral — Sequential Pipeline with Gates |
+| **Where Applied** | `rules/PHASES.md` (format), `rules/CORE.md` (L5 enforcement), variant workflow files |
+| **Rationale** | Sequential execution with exit criteria prevents downstream errors from incomplete phases and ensures each phase builds on verified prior work |
+
+### Description
+
+Every workflow variant defines a series of phases that execute strictly in order:
+
+1. **Requirements Intake** — Parse ALL requirements into a numbered registry (100% fidelity)
+2. **Phase 1 → Phase N** — Each phase delegates to an agent, produces a deliverable, and verifies exit criteria
+3. **Deliverable as Constraint** — Completed deliverables become immutable inputs for subsequent phases (L8: Stateful Handoff)
+
+### Phase Output Format (from PHASES.md)
+
+```markdown
+## 🎭 Phase {N}: {name}
+### Sub-agent: `{agent}` — {role}
+{agent work / summary}
+### Exit Criteria
+- [x] {criterion_1}
+- [x] {criterion_2}
+### ✅ `{agent}` complete
+**Deliverable**: {summary}
+```
+
+### Key Rules
+
+- L5 (Sequential Execution): Phase N completes before Phase N+1 starts
+- L8 (Stateful Handoff): Prior deliverables are immutable constraints
+- L2 (Requirement Integrity): 100% fidelity, zero loss in requirements registry
+- No batching: phases execute one at a time, never loaded in parallel (anti-pattern A8)
+
+---
+
+## Golden Triangle Pattern
+
+| Attribute | Value |
+|-----------|-------|
+| **Pattern Type** | Quality — Adversarial Collaboration |
+| **Where Applied** | `rules/TEAMS.md` (protocol), `agents/teams/*/` (17 team directories) |
+| **Rationale** | Structured tension between a builder and a challenger, orchestrated by an arbitrator, produces higher-quality output than single-agent execution or parallel cooperation |
+
+### Description
+
+When `:team` variant is invoked, each phase spawns exactly 3 agent roles:
+
+| Role | Personality | Function |
+|------|------------|----------|
+| **Tech Lead** | Pragmatic, decisive | Decomposes tasks, coordinates, arbitrates disputes, synthesizes |
+| **Executor** | Builder, evidence-driven | Implements, defends work with technical proof |
+| **Reviewer** | Skeptical, demanding | Challenges with devil's advocate mindset, assumes defects exist |
+
+### Debate Rules
+
+- Max 3 rounds of Executor ↔ Reviewer exchange
+- Executor MUST defend with technical evidence (benchmarks, specs, references)
+- "I disagree" without proof = automatic FAIL
+- After round 3 without agreement, Tech Lead reads all exchanges and issues a binding `DECISION`
+- Consensus stamp required: `TechLead ✓ | Executor ✓ | Reviewer ✓`
+
+### Example Reference
+
+From `rules/TEAMS.md`:
+```
+CORE PRINCIPLE: Every team phase spawns exactly 3 agent roles — no more, no less.
+Quality emerges from structured tension between an Executor who builds
+and a Reviewer who challenges, orchestrated by a Tech Lead who arbitrates.
+```
+
+---
+
+## File-Based Communication (Mailbox)
+
+| Attribute | Value |
+|-----------|-------|
+| **Pattern Type** | Communication — Append-Only Event Log |
+| **Where Applied** | `rules/TEAMS.md` (protocol), `./reports/{topic}/MAILBOX-{date}.md` (runtime artifact) |
+| **Rationale** | File-based communication provides a persistent, auditable record of all inter-agent exchanges without requiring a message broker or shared memory |
+
+### Description
+
+All agent communication within a Golden Triangle flows through two shared artifacts:
+
+| Artifact | Format | Rules |
+|----------|--------|-------|
+| **Shared Task List** | Table in phase output | Owned by Tech Lead; tracks assignments, status, priorities |
+| **Mailbox** | `./reports/{topic}/MAILBOX-{date}.md` | Append-only — no edits or deletions; all agents can read; all agents can append |
+
+### Message Types
+
+| Type | Direction | Purpose |
+|------|-----------|---------|
+| TASK_ASSIGNMENT | Tech Lead → Executor | Assign work |
+| SUBMISSION | Executor → Reviewer | Submit deliverable |
+| REVIEW (PASS/FAIL) | Reviewer → Executor | Accept or reject |
+| DEFENSE | Executor → Reviewer | Defend with evidence |
+| RESUBMISSION | Executor → Reviewer | Resubmit after fixes |
+| DECISION | Tech Lead → All | Binding arbitration |
+
+### Enforcement
+
+- C8-TEAMS-01 (BLOCK): Mailbox is append-only and required for all inter-agent exchanges
+- No agent may edit or delete prior exchanges
+
+---
+
+## Platform Abstraction Pattern
+
+| Attribute | Value |
+|-----------|-------|
+| **Pattern Type** | Structural — Write-Once Deploy-Many |
+| **Where Applied** | `cli/install.js` (implementation), all `.md` and `.yaml` files (placeholders) |
+| **Rationale** | Supporting 5 AI platforms from a single codebase eliminates duplication and ensures feature parity; install-time substitution is simpler than runtime branching |
+
+### Description
+
+The framework uses `{TOOL}` (and related) placeholders throughout all Markdown and YAML files. At install time, `cli/install.js` performs text replacement based on the target platform:
+
+| Platform | `{TOOL}` Value | Install Path |
+|----------|---------------|--------------|
+| Cursor | `cursor` | `~/.cursor/skills/agent-assistant/` |
+| GitHub Copilot | `copilot` | `~/.copilot/skills/agent-assistant/` |
+| Claude Code | `claude` | `~/.claude/skills/agent-assistant/` |
+| Codex | `codex` | `~/.codex/skills/agent-assistant/` |
+| Antigravity/Gemini | `gemini/antigravity` | `~/.gemini/antigravity/skills/agent-assistant/` |
+
+### Platform-Specific Assets
+
+Beyond placeholder substitution, each platform has custom assets in `code-assistants/{tool}-assistant/`:
+
+| Platform | Custom Assets |
+|----------|--------------|
+| Cursor | Rules directory, `.cursorrules` file |
+| Copilot | `agent-assistant.agent.md` VS Code prompt file |
+| Claude Code | Claude-specific configuration |
+| Codex | Codex-specific configuration |
+| Antigravity | Gemini-specific configuration |
+
+### Example Reference
+
+From `cli/install.js` (Cursor config):
+```javascript
+replacements: {
+    '~/.{TOOL}/skills/agent-assistant/': '~/.cursor/skills/agent-assistant/',
+    '{TOOL}': 'cursor',
+    '{HOME}': '~',
+}
+```
+
+---
+
+## Evidence Sources
+
+| Source | Path | What It Provides |
+|--------|------|------------------|
+| CORE.md v4.1 | `rules/CORE.md` | Orchestrator identity, execution loop, 10 laws, prohibitions, self-check |
+| PHASES.md | `rules/PHASES.md` | Phase output format, requirements registry, exit criteria pattern |
+| AGENTS.md | `rules/AGENTS.md` | TIER 1/TIER 2 protocol, embodiment format, tool discovery, completion guarantee |
+| SKILLS.md | `rules/SKILLS.md` | HSOL resolution, fitness formula, trust lifecycle, complexity gating |
+| TEAMS.md | `rules/TEAMS.md` | Golden Triangle: roles, debate mechanism (3 rounds), mailbox protocol, C8 checkpoints |
+| ERRORS.md | `rules/ERRORS.md` | Anti-patterns A1–A10 (violations of proper patterns) |
+| REFERENCE.md | `rules/REFERENCE.md` | Command-variant table, agent category table |
+| _index.yaml | `matrix-skills/_index.yaml` | HSOL config, total skills, domain count, discovery settings |
+| cli/install.js | `cli/install.js` | Platform configs, replacement maps, file copy logic |
+| cook.md | `commands/cook.md` | Example command router: pre-flight, routing logic, variant table |
+| backend-engineer.md | `agents/backend-engineer.md` | Example agent: profile `backend:execution`, cognitive anchor, handoffs |
+| docs-manager.md | `agents/docs-manager.md` | Example agent profile: `research:documentation` |
+| agents/teams/ | `agents/teams/` | 17 team directories with techlead.md, executor.md, reviewer.md each |