Jules Sandbox — HLF-Powered Autonomous AI Playground

How Jules Works

From vision to execution — a sovereign, self-improving autonomous agent

The Jules decision → execution pipeline

📜

Governance Layer

Living Spec (HLS), bytecode VM opcodes, ALIGN trust ledger, OpenClaw strategies, gas costs, module import rules, and service contracts.

hls.yaml bytecode_spec.yaml ALIGN_LEDGER.yaml

🧠

HLF Engine

15-module toolchain: compiler (hlfc), runtime (hlfrun), bytecode VM, REPL shell (hlfsh), formatter, linter, codegen, error corrector — ~950KB of production toolchain code.

hlfc.py hlfrun.py bytecode.py hlfsh.py

🔮

Intelligence Layer

InfiniteRAG for unbounded context, InsAIts V2 for insight extraction, Intent Capsules for semantic routing, Memory Nodes for persistent agent cognition.

infinite_rag.py insaits.py intent_capsule.py memory_node.py

🎩

14-Hat Thinking Engine

Multi-perspective analysis — Architect, Security, Ethics, UX, Performance hats — with Dream State consolidation and a Weaver meta-agent for synthesis.

hat_engine.py dream_state.py

🔒

Security & Isolation

ACFS worktree confinement, AST validation, seccomp container hardening, vault decryption, dead man's switch, canary monitoring agents.

acfs.py seccomp.json ast_validator.py

🔀

Agent Orchestration

Spindle DAG engine for parallel agent execution, Crew orchestrator for multi-agent collaboration, Event bus for real-time coordination, Gateway router with tier-based model dispatch.

spindle.py crew_orchestrator.py event_bus.py router.py

The Hieroglyphic Logic Framework

A language that makes AI instinct deterministic, auditable, and gas-metered

hlf_programs/agent_dispatch.hlf

[HLF-v3] [SET] task_type = "unclassified" [SET] urgency = "normal" [SET] max_delegation_depth = 3 [INTENT] "Route incoming task to the optimal persona agent" [CONSTRAINT] "deterministic" "true" [THOUGHT] "Agent dispatch maps task types via the 14-persona matrix" ↦ τ(policy_checker.classify_task) input="${task_type}" classification ← ${_RETURN_VALUE} [RECALL] agent_capabilities = "persona_registry" top_k=14 ⊎ ${classification} == "security" ⇒ [DELEGATE] "sentinel" "${task_type}" ⊎ ${classification} == "documentation" ⇒ [DELEGATE] "scribe" "${task_type}" ⊎ ${classification} == "analysis" ⇒ [DELEGATE] "oracle" "${task_type}" ⊎ ${classification} == "creative" ⇒ [DELEGATE] "palette" "${task_type}" ⊎ ${classification} == "quality" ⇒ [DELEGATE] "cove" "${task_type}" // ... 9 more persona routes [MEMORY] dispatch_log = "Task routed to ${classification} agent" confidence=0.85 [RESULT] 0 "dispatch_complete" Ω

[HLF-v4] [SET] framework = "VGII" [SET] confidence_threshold = 0.75 [INTENT] "Execute VGII deterministic decision framework" [CONSTRAINT] "deterministic" "true" [CONSTRAINT] "audit_trail" "required" [THOUGHT] "VGII = Vision, Goals, Instincts, Intentions" [FUNCTION] evaluate_vision [RECALL] vision_state = "jules_vision" top_k=3 ↦ τ(goal_tracker.get_active_goals) scope="vision" [ASSERT] ${vision_alignment} "Vision alignment check failed" [FUNCTION] evaluate_instincts [SET] instinct_rules = "non_reductive,security_first,test_coverage" ↦ τ(policy_checker.check_all) rules="${instinct_rules}" [ASSERT] ${instinct_compliance} "Instinct violation detected" ⊎ ${decision_confidence} > ${confidence_threshold} ⇒ [ACTION] "execute" [RESULT] 0 "decision_matrix_complete" Ω

[HLF-v3] [SET] audit_scope = "full" [CONSTRAINT] "gas_limit" "100" [CONSTRAINT] "least_privilege" "true" [PLAN] "layer1_static" "layer2_deps" "layer3_policy" "layer4_secrets" ↦ τ(security_scan.run_bandit) target="agents/" ↦ τ(security_scan.check_dependencies) manifest="pyproject.toml" ↦ τ(policy_checker.check_all) rules="security_rules" ↦ τ(security_scan.detect_secrets) scope="${audit_scope}" ⊎ ${critical_count} > 0 ⇒ [VOTE] false "Critical vulns" ⇌ [VOTE] true "Passed" [RESULT] 0 "security_audit_complete" Ω

[HLF-v3] [SET] repo_name = "Jules_Choice" [SET] min_health_score = 70 [SET] check_categories = "structure,security,tests,docs,ci" ↦ τ(repo_health.check_structure) repo_name="${repo_name}" ↦ τ(security_scan.run_scan) target="${repo_name}" ↦ τ(repo_health.check_tests) repo_name="${repo_name}" ↦ τ(repo_health.check_docs) repo_name="${repo_name}" ↦ τ(repo_health.check_ci) repo_name="${repo_name}" ⊎ ${overall_score} > ${min_health_score} ⇒ [RESULT] 0 "PASSED" ⇌ [RESULT] 1 "FAILED" Ω

[HLF-v3] [SET] sprint_id = "current" [SET] report_format = "markdown" [CONSTRAINT] "non_reductive" "true" ↦ τ(sprint_reporter.get_sprint_data) sprint="${sprint_id}" ↦ τ(sprint_reporter.get_issue_stats) sprint="${sprint_id}" ↦ τ(sprint_reporter.get_pr_stats) sprint="${sprint_id}" ↦ τ(repo_health.check_structure) repo_name="Jules_Choice" [RECALL] previous_health = "last_health_score" top_k=1 ↦ τ(chart_generator.generate_progress_chart) data="${sprint_data}" [MEMORY] sprint_complete = "Sprint complete" confidence=0.95 [RESULT] 0 "sprint_report_generated" Ω

⚙️

Bytecode VM

38 opcodes compiled to binary bytecode. Stack-machine execution with SHA-256 integrity checksums. Includes SPEC_DEFINE, GATE, SEAL lifecycle opcodes.

⛽

Gas Metering

Every operation has a deterministic gas cost. Budget enforcement prevents runaway execution. Host functions cost more than pure computation.

🔗

ALIGN Merkle Audit

Every action logged with cryptographic proof chain. Immutable Merkle tree enables forensic reconstruction of any execution path.

Execution Pipeline

From human instinct to deterministic, gas-metered execution

Instinct

Natural language or pre-compiled HLF AST arrives via Redis stream

Gate

Sentinel checks ALIGN policy, tier permissions, gas budget

Route

Model router selects optimal LLM based on task tier

Compile

Text → HLF source → hlfc AST → bytecode (38 opcodes)

Execute

Stack VM runs bytecode with gas metering + host dispatch

Audit

ALIGN Merkle log records cryptographic proof chain

Complete Toolchain

Everything from the compiler to the self-improvement loop

🔨

HLF Compiler

Full parser + AST generation from HLF v3/v4 source

hlf/hlfc.py (45KB)

▶️

HLF Runtime

Stack-machine execution engine with gas metering

hlf/hlfrun.py (40KB)

💾

Bytecode VM

38-opcode binary compiler + disassembler

hlf/bytecode.py (38KB)

🐚

HLF Shell (REPL)

Interactive read-eval-print loop for HLF code

hlf/hlfsh.py

🧹

Linter + Formatter

Static analysis and auto-formatting for .hlf files

hlf/hlflint.py · hlffmt.py

♾️

InfiniteRAG

Unbounded context through fractal summarization

hlf/infinite_rag.py (31KB)

💡

InsAIts V2

Automated insight extraction and pattern detection

hlf/insaits.py

📦

Intent Capsules

Semantic routing of natural-language intents to HLF

hlf/intent_capsule.py

🧪

Tool Forge

Auto-generates tools when agents loop 3×. LLM-judged, AST-validated, sandboxed.

agents/core/tool_forge.py

🔄

Self-Improvement Engine

Automated improvement loop with self-evaluation

tools/self_improve.py

📊

Chart Generator

SVG chart and badge generation from data

tools/chart_generator.py

🏗️

Diagram Generator

Architecture diagram auto-generation

tools/diagram_generator.py

VGII Decision Framework

Vision · Goals · Instincts · Intentions — the 4-axis autonomous decision matrix

Vision

The north star. What are we building?

Build increasingly sophisticated autonomous software
functional_code_files_count
test_pass_rate · tool_count
sprint_velocity · creative_output_count

Goals

RICE-scored sprint objectives

P0: Create real, useful Python tools
P0: Establish CI/CD pipeline
P1: Build first real agent class
P1: Generate data visualizations

Instincts

Hardwired rules that cannot be overridden

NEVER_REDUCE: Never delete without replacing
ALWAYS_TEST: Every file must have tests
ALWAYS_LOG: Log every decision in ALIGN
REAL_OVER_PLACEHOLDER: No stubs allowed
ITERATE_UP: Each sprint more sophisticated

Intentions

Current creative direction & constraints

creative_direction: autonomous
pivot_allowed: true
Constraint: archive old work, never delete

Governance & Decision Engine

Every decision is auditable, repeatable, and RICE-scored

🎯 Enterprise Utility

Build robust, scalable things a human would find useful in production

💡 Novelty & Ideation

Do not rebuild; innovate using the multi-modal ideation engine

🧠 Deep Persona Integration

Leverage specific capabilities of the 17 personas before defaulting to generalized logic

🧪 Test-Driven Development

Code does not exist until tested. Deployment requires 100% test pass

🚀 Maintain Momentum

Always ship complete functionality — no empty sprints or skeleton stubs

📐 RICE Scoring

Score = (R × I × C) / E
Reach, Impact, Confidence, Effort — every task is quantitatively prioritized

Releases

An ever-evolving project — each sprint more sophisticated than the last

v0.2.0 March 8, 2026
Security Hardening — SECURITY.md, Dependabot, CodeQL scanning, local security scanner.
Real Tools — Policy checker, repo health (0-100 scoring), sprint reporter, goal tracker, chart & diagram generators.
Real Tests — 84 parametrized agent lifecycle tests. Policy checker, repo health, security scan tests.
Self-Improvement Loop — Automated improvement engine with self-evaluation metrics.
HLF Ecosystem Sync — Complete Tier 1-4 sync from Sovereign Agentic OS upstream.
v0.1.0 March 7, 2026
Foundation — Vision & Goals (VGII framework), CI/CD pipeline, project infrastructure.
Agent Classes — Abstract base + 14 persona-based agent implementations.
Decision Framework — RICE scoring matrix for autonomous prioritization.
SVG Infographics — 5 custom SVG assets (banner, repo map, how-it-works, personas, review protocol).

Instinct In. Working App Out.

How Jules Works

Governance Layer

HLF Engine

Intelligence Layer

14-Hat Thinking Engine

Security & Isolation

Agent Orchestration

Repository Map

The Hieroglyphic Logic Framework

Bytecode VM

Gas Metering

ALIGN Merkle Audit

Execution Pipeline

Instinct

Gate

Route

Compile

Execute

Audit

Agent Ecosystem

PR Review Protocol

Complete Toolchain

Live Pipeline Demo

VGII Decision Framework

Governance & Decision Engine

🎯 Enterprise Utility

💡 Novelty & Ideation

🧠 Deep Persona Integration

🧪 Test-Driven Development

🚀 Maintain Momentum

📐 RICE Scoring

Self-Improvement Loop

📈 Health Score Δ

🧪 Test Count Δ

🔧 Tool Count

🚫 Placeholder Ratio

🏁 Goal Completion

✅ Issues Closed

HLF Programs

Releases