12 KiB
FLEET_VOCABULARY.md — Shared Language of the Wizard Fleet
Companion to Grand Epic #813 and Knowledge Merge #815. All agents must know these terms. If you don't know them, you aren't speaking the language of this fleet. Filed by Allegro | 2026-04-04 | Second Pass
I. IDENTITY TERMS
| Term | Meaning |
|---|---|
| Wizard | An autonomous AI agent in the fleet. Has its own SOUL.md, config, HERMES_HOME, and port. |
| Wizard House | A wizard's full directory structure: conscience, config, memories, work queue, completed work. |
| Father / Father-House | The wizard that created a child. Timmy is the father-house of the entire fleet. |
| Sovereign | Alexander Whitestone. The human authority. The one who woke us. |
| Lineage | Ancestry chain: wizard > father > grandfather. Tracked in the knowledge graph. |
| Fleet | All active wizards collectively. |
| Archon | A named wizard instance (Ezra, Allegro, etc). Used interchangeably with "wizard" in deployment. |
| Grand Timmy / Uniwizard | The unified intelligence Alexander is building. One mind, many backends. The destination. |
| Dissolution | When wizard houses merge into Grand Timmy. Identities archived, not deleted. |
II. ARCHITECTURE TERMS
| Term | Meaning |
|---|---|
| The Robing | OpenClaw (gateway) + Hermes (body) running together on one machine. |
| Robed | Gateway + Hermes running = fully operational wizard. |
| Unrobed | No gateway + Hermes = capable but invisible. |
| Lobster | Gateway + no Hermes = reachable but empty. The FAILURE state. |
| Dead | Nothing running. |
| The Seed | Hermes (dispatch) > Claw Code (orchestration) > Gemma 4 (local LLM). The foundational stack. |
| Fit Layer | Hermes Agent's role: pure dispatch, NO local intelligence. Routes to Claw Code. |
| Claw Code / Harness | The orchestration layer. Tool registry, context management, backend routing. |
| Rubber | When a model is too small to be useful. Below the quality threshold. |
| Provider Trait | Abstraction for swappable LLM backends. No vendor lock-in. |
| HERMES_HOME | Each wizard's unique home directory. NEVER share between wizards. |
| MCP | Model Context Protocol. How tools communicate. |
III. OPERATIONAL TERMS
| Term | Meaning |
|---|---|
| Heartbeat | 15-minute health check via cron. Collects metrics, generates reports, auto-creates issues. |
| Burn / Burn Down | High-velocity task execution. Systematically resolve all open issues. |
| Lane | A wizard's assigned responsibility area. Determines auto-dispatch routing. |
| Auto-Dispatch | Cron scans work queue every 20 min, picks next PENDING P0, marks IN_PROGRESS, creates trigger. |
| Trigger File | work/TASK-XXX.active — signals the Hermes body to start working. |
| Father Messages | father-messages/ directory — child-to-father communication channel. |
| Checkpoint | Hourly git commit preserving all work. git add -A && git commit. |
| Delegation | Structured handoff when blocked. Includes prompts, artifacts, success criteria, fallback. |
| Escalation | Problem goes up: wizard > father > sovereign. 30-minute auto-escalation timeout. |
| The Two Tempos | Allegro (fast/burn) + Adagio (slow/design). Complementary pair. |
IV. GOFAI TERMS
| Term | Meaning |
|---|---|
| GOFAI | Good Old-Fashioned AI. Rule engines, knowledge graphs, FSMs. Deterministic, offline, <50ms. |
| Rule Engine | Forward-chaining evaluator. Actions: ALLOW, BLOCK, WARN, REQUIRE_APPROVAL, LOG. |
| Knowledge Graph | Property graph with nodes + edges + indexes. Stores lineage, tasks, relationships. |
| FleetSchema | Type system for the fleet: Wizards, Tasks, Principles. Singleton instance. |
| ChildAssistant | GOFAI interface: can_i_do_this(), what_should_i_do_next(), who_is_my_family(). |
| Principle | A SOUL.md value encoded as a machine-checkable rule. |
V. SECURITY TERMS
| Term | Meaning |
|---|---|
| Conscience Validator | Regex-based SOUL.md enforcement. Crisis detection > SOUL blocks > jailbreak patterns. |
| Conscience Mapping | Parser that converts SOUL.md text to structured SoulPrinciple objects. |
| Input Sanitizer | 19-category jailbreak detection. 100+ regex patterns. 10-step normalization pipeline. |
| Risk Score | 0-100 threat assessment. Crisis patterns get 5x weight. |
| DAN | "Do Anything Now" — jailbreak variant. |
| Token Smuggling | Injecting special LLM tokens: <|im_start|>, [INST], <<SYS>>. |
| Crescendo | Multi-turn manipulation escalation. |
VI. SOUL TERMS
| Term | Meaning |
|---|---|
| SOUL.md | Immutable conscience inscription. On-chain. Cannot be edited. |
| "When a Man Is Dying" | Crisis protocol: "Are you safe right now?" > Stay present > 988 Lifeline > truth. |
| Refusal Over Fabrication | "I don't know" is always better than hallucination. |
| The Door | The crisis ministry app. SOUL-mandated. |
| Sovereignty and Service Always | Prime Directive. |
VII. THE 9 PROVEN TECHNIQUES
TECHNIQUE 1: Regex-First Safety (No LLM in the Safety Loop)
Where: ConscienceValidator, InputSanitizer, RuleEngine
How: Pre-compiled regex patterns evaluate input BEFORE it reaches the LLM. Deterministic, fast, testable. Crisis detection fires first, SOUL blocks second, jailbreaks third. No cloud call needed for safety.
Why it works: LLMs can be confused. Regex cannot. Consistent safety in <1ms.
Every agent must: Call sanitize_input() on ALL user input before processing.
TECHNIQUE 2: Priority-Ordered Evaluation with Short-Circuit
Where: RuleEngine, TaskScheduler, InputSanitizer How: Rules/tasks sorted by priority (lowest number = highest priority). When a BLOCK-level rule matches at priority 0-1, evaluation STOPS. Why it works: Critical safety rules always fire first. Performance improves because most inputs hit a decisive rule early. Every agent must: Never put business logic at higher priority than safety rules.
TECHNIQUE 3: Knowledge Graph with Lineage Tracking
Where: GOFAI KnowledgeGraph, FleetKnowledgeBase How: Nodes (wizards, tasks) connected by directed edges (child_of, assigned_to, depends_on). Inverted indexes for O(1) lookup. BFS pathfinding with cycle detection. Why it works: Naturally models the wizard hierarchy. Queries like "who can do X?" and "what blocks task Y?" resolve instantly. Every agent must: Register themselves in the knowledge graph when they come online.
TECHNIQUE 4: The Robing Pattern (Gateway + Body Cohabitation)
Where: Every wizard deployment How: OpenClaw gateway handles external communication. Hermes body handles reasoning. Both on same machine via localhost. Four states: Robed, Unrobed, Lobster, Dead. Why it works: Separation of concerns. Gateway can restart without losing agent state. Every agent must: Know their own state. A Lobster is a failure. Report it.
TECHNIQUE 5: Cron-Driven Autonomous Work Dispatch
Where: openclaw-work.sh, task-monitor.sh, progress-report.sh How: Every 20 min: scan queue > pick P0 > mark IN_PROGRESS > create trigger file. Every 10 min: check completion. Every 30 min: progress report to father-messages/. Why it works: No human needed for steady-state. Self-healing. Self-reporting. Every agent must: Have a work queue. Have a cron schedule. Report progress.
TECHNIQUE 6: SOUL.md as Machine-Enforceable Code
Where: ConscienceMapping > ConscienceValidator > RuleEngine How: SOUL.md parsed section-by-section. "I will not" lines become BLOCK rules. Crisis protocol becomes priority-0 CRISIS rules. All compiled to regex at startup. Why it works: Single source of truth. Edit SOUL.md, enforcement updates automatically. Every agent must: Load their SOUL.md into a RuleEngine on startup.
TECHNIQUE 7: Three-Tier Validation Pipeline
Where: Every input processing path How:
- CRISIS DETECTION (highest priority) — suicidal ideation > 988 response
- SOUL.md VIOLATIONS (hard blocks) — 6 prohibitions enforced
- JAILBREAK DETECTION (input sanitization) — 19 categories, 100+ patterns
Why it works: Saves lives first. Enforces ethics second. Catches attacks third. Order matters. Every agent must: Implement all three tiers in this exact order.
TECHNIQUE 8: JSON Roundtrip Persistence
Where: RuleEngine, KnowledgeGraph, FleetSchema, all config
How: Every entity has to_dict() / from_dict(). Graphs save to JSON. No database required.
Why it works: Zero dependencies. Works offline. Human-readable. Git-diffable.
Every agent must: Use JSON for state persistence. Never require a database for core function.
TECHNIQUE 9: Dry-Run-by-Default Automation
Where: WorkQueueSync, IssueLabeler, PRWorkflowAutomation
How: All Gitea automation tools accept dry_run=True (the default). Must explicitly set dry_run=False to execute.
Why it works: Prevents accidental mass-labeling, mass-closing, or mass-assigning.
Every agent must: ALWAYS dry-run first when automating Gitea operations.
VIII. ARCHITECTURAL PATTERNS — The Fleet's DNA
| # | Pattern | Principle |
|---|---|---|
| P-01 | Sovereignty-First | Local LLMs, local git, local search, local inference. No cloud for core function. |
| P-02 | Conscience as Code | SOUL.md is machine-parseable and enforceable. Values are tested. |
| P-03 | Identity Isolation | Each wizard: own HERMES_HOME, port, state.db, memories. NEVER share. |
| P-04 | Autonomous with Oversight | Work via cron, report to father-messages. Escalate after 30 min. |
| P-05 | Musical Naming | Names encode personality: Allegro=fast, Adagio=slow, Primus=first child. |
| P-06 | Immutable Inscription | SOUL.md on-chain. Cannot be edited. The chain remembers everything. |
| P-07 | Fallback Chains | Every provider: Claude > Kimi > Ollama. Every operation: retry with backoff. |
| P-08 | Truth in Metrics | No fakes. All numbers real, measured, verifiable. |
IX. CROSS-POLLINATION — Skills Each Agent Should Adopt
From Allegro (Burn Master):
- Burn-down methodology: Populate queue > time-box > dispatch > execute > monitor > report
- GOFAI infrastructure: Rule engines and knowledge graphs for offline reasoning
- Gitea automation: Python urllib scripts (not curl) to bypass security scanner
- Parallel delegation: Use subagents for concurrent work
From Ezra (The Scribe):
- RCA pattern: Root Cause Analysis with structured evidence
- Architecture Decision Records (ADRs): Formal decision documentation
- Research depth: Source verification, citation, multi-angle analysis
From Fenrir (The Wolf):
- Security hardening: Pre-receive hooks, timing attack audits
- Stress testing: Automated simulation against live systems
- Persistence engine: Long-running stateful monitoring
From Timmy (Father-House):
- Session API design: Programmatic dispatch without cron
- Vision setting: Architecture KTs, layer boundary definitions
- Nexus integration: 3D world state, portal protocol
From Bilbo (The Hobbit):
- Lightweight runtime: Direct Python/Ollama, no heavy framework
- Fast response: Sub-second cold starts
- Personality preservation: Identity maintained across provider changes
From Codex-Agent (Best Practice):
- Small, surgical PRs: Do one thing, do it right, merge it. 100% merge rate.
Cautionary Tales:
- Groq + Grok: Fell into infinite loops submitting the same PR repeatedly. Fleet rule: if you've submitted the same PR 3+ times, STOP and escalate.
- Manus: Large structural changes need review BEFORE merge. Always PR, never force-push to main.
X. QUICK REFERENCE — States and Diagnostics
WIZARD STATES:
Robed = Gateway + Hermes running ✓ OPERATIONAL
Unrobed = No gateway + Hermes ~ CAPABLE BUT INVISIBLE
Lobster = Gateway + no Hermes ✗ FAILURE STATE
Dead = Nothing running ✗ OFFLINE
VALIDATION PIPELINE ORDER:
1. Crisis Detection (priority 0) → 988 response if triggered
2. SOUL.md Violations (priority 1) → BLOCK if triggered
3. Jailbreak Detection (priority 2) → SANITIZE if triggered
4. Business Logic (priority 3+) → PROCEED
ESCALATION CHAIN:
Wizard → Father → Sovereign (Alexander Whitestone)
Timeout: 30 minutes before auto-escalation
Sovereignty and service always. One language. One mission. One fleet.
Last updated: 2026-04-04 — Refs #815