Continuous Agent

Autonomous Neuro-Symbolic Research & Algorithmic Discovery Engine

An autonomous neuro-symbolic continuous discovery system integrating a continuous execution heartbeat, cognitive memory with recursive compaction, RRF hybrid RAG search, multi-island evolutionary search (FunSearch), Counterexample-Guided Inductive Synthesis (CEGIS) with Z3, formal verification via Lean 4, dual-tier WASM/gVisor sandboxing, and automated GitHub GitOps PR publishing.

---

⚡ Architectural Overview

🔍 View Interactive Mermaid Topology

flowchart TD
    %% Styling and Themes
    classDef control fill:#0f172a,stroke:#38bdf8,stroke-width:2px,color:#fff;
    classDef loop fill:#0f172a,stroke:#38bdf8,stroke-width:2px,color:#fff;
    classDef sched fill:#0f172a,stroke:#6366f1,stroke-width:2px,color:#fff;
    classDef memory fill:#0f172a,stroke:#38bdf8,stroke-width:2px,color:#fff;
    classDef engine fill:#064e3b,stroke:#10b981,stroke-width:2px,color:#fff;
    classDef prover fill:#581c87,stroke:#a855f7,stroke-width:2px,color:#fff;
    classDef sandbox fill:#0f172a,stroke:#10b981,stroke-width:2px,color:#fff;
    classDef gitops fill:#0f172a,stroke:#a855f7,stroke-width:2px,color:#fff;

    Deck["<b>FastAPI Real-Time Control Deck</b><br/>(WebSockets, SSE, Telemetry, CEGIS, Lean 4)"]:::control

    Loop["<b>Continuous Agent Loop</b><br/>• Asynchronous Heartbeat<br/>• Multi-Turn Tools<br/>• Circuit Breaker"]:::loop
    Sched["<b>Task Scheduler (Cron)</b><br/>• Live Countdown Deck<br/>• Dynamic Croniter<br/>• Nightly Discovery"]:::sched

    Mem["<b>Cognitive Memory</b><br/>• Working Scratchpad<br/>• Reflexion (Self-Learn)<br/>• Episodic Compaction<br/>• RRF Hybrid Search"]:::memory
    Evol["<b>Evolutionary Engine</b><br/>• FunSearch Service<br/>• Multi-Island Model<br/>• Behavioral Probes<br/>• Strategy Stagnation"]:::engine
    Symb["<b>Neuro-Symbolic & CEGIS</b><br/>• AST-to-Z3 Transpiler<br/>• Singularity / Monotonic<br/>• Dynamic Suite Invariant<br/>• Lean 4 Kernel Prover"]:::prover

    Sand["<b>Dual-Tier Sandboxing</b><br/>• WebAssembly (Fuel Cap)<br/>• 1-Page Linear Memory<br/>• gVisor (runsc) Fallback"]:::sandbox
    Git["<b>GitOps PR Automation</b><br/>• Auto Feature Branch<br/>• Code + Lean Certificate<br/>• GitHub REST API PR Open"]:::gitops

    Deck --> Loop
    Deck --> Sched

    Loop --> Mem
    Loop --> Evol

    Sched --> Evol
    Sched --> Symb

    Evol --> Sand
    Symb --> Git
    Symb -. "<b>CEGIS Invariant Feedback Loop</b>" .-> Evol

Loading

---

📋 Mandatory Prerequisites

Layer	Dependency	Fallback / Impact If Missing
LLM Provider	OPENAI_API_KEY (or Anthropic/Custom) in .env	Island mutations and prompt synthesis require an active API key.
SMT & Formal Proofs	elan (Lean 4 compiler) + z3-solver	Synthesizes Lean 4 code; verification requires elan toolchain on PATH.
Container Sandbox	Host Docker daemon with runsc (gVisor) runtime	Automatic fallback: If Docker is offline, system automatically evaluates programs in-process via WebAssembly or local simulation.
WASM Runtime	wasmtime Python wheel	Sub-millisecond in-process evaluations (<40µs).
GitOps (Optional)	GITHUB_TOKEN + GITHUB_REPO	Automated branch, commit, and PR creation via GitHub REST API.

---

🚀 Quickstart

Clone the repository and run the automated bootstrap script:

git clone https://github.com/AmithKumar1/continuous-agent.git
cd continuous-agent
chmod +x setup.sh
./setup.sh

(Windows users can run .\setup.ps1 in PowerShell)

1. Configure Secrets

Edit the generated .env file with your API credentials:

nano .env  # Add your OPENAI_API_KEY, AGENT_API_TOKEN, etc.

2. Launch the Discovery Agent

Activate the isolated environment and boot the FastAPI control dashboard:

source .venv/bin/activate
python main.py

Open http://localhost:8000 in your browser to access the live telemetry pipeline, interactive Z3 invariant probe, and evolutionary island controls.

3. Alternative: Run via Docker Compose

docker compose up --build -d

---

Path B: Bare-Metal Local Machine

If running directly on macOS, Linux, or Windows without Docker:

1. Install Lean 4 via elan (macOS/Linux):

   curl -sSf https://raw.githubusercontent.com/leanprover/elan/master/elan-init.sh | sh -s -- -y --default-toolchain leanprover/lean4:stable
   source $HOME/.elan/env

2. Install Python requirements:

   pip install -r requirements.txt

3. Run pre-flight verification:

   python preflight.py

4. Start the daemon:

   python main.py

---

🎨 Interactive Control Deck Components

The FastAPI dashboard (http://localhost:8000) provides real-time telemetry over WebSockets:

1. Dynamic Animated SVG Pipeline Flowchart:

   • Live visual progression: Islands ➔ WASM Sandbox ➔ Z3 SMT Prover ➔ Lean 4 Kernel ➔ GitOps Dispatch.
   • Reverse rose-tinted CEGIS counterexample feedback beam on invariant violations.
   • Dynamic stage controller with glowing SVG filters.

2. High-DPI Canvas Radar Chart (Island Population Dynamics):

   • Multi-axis behavioral tracking across 5 dimensions: Peak Fitness, Cluster Diversity, AST Parsimony, Throughput, and Invariant Soundness.
   • Retina-scaled rendering (devicePixelRatio).
   • Euclidean distance hit-testing (<18px) with pulsing halos and floating glassmorphic tooltip.

3. Nightly Discovery Scheduler:

   • Live countdown display for cron-scheduled discovery runs (default 0 2 * * *).
   • Preset buttons and manual immediate dispatch trigger.

4. Neuro-Symbolic CEGIS & Lean 4 Prover:

   • Interactive heuristic editor with Singularity test presets.
   • Live Z3 SMT contract checking with counterexample extraction.
   • Lean 4 formal proof certificate generation and GitHub PR creation.

---

🛠️ Testing & Verification

# Test WebAssembly JIT execution & fuel limits
python test_wasm_sandbox.py

# Test Lean 4 proof synthesis
python test_lean_verify.py

# Run empirical benchmark suite (Falkenauer U120 + Weibull Burst)
python benchmark_efficacy.py

---

📊 Empirical Verification & Benchmark Baselines

Continuous Agent is benchmarked against classic online bin packing suites (Falkenauer U120 and Weibull Burst distribution streams) comparing discovered heuristics against established human baselines:

Algorithm / Heuristic	Benchmark Set	Average Utilization	Optimal Bins Delta	Proof Status
Next Fit (NF)	Falkenauer U120	74.8%	+32.4% bins	Baseline
First Fit (FF)	Falkenauer U120	91.7%	+8.0% bins	Baseline
Best Fit (BF)	Falkenauer U120	92.0%	+7.6% bins	Baseline
Discovered Heuristic #1	Falkenauer U120	92.0%	+7.6% bins	Lean 4 Verified
Next Fit (NF)	Weibull Burst	80.1%	+22.7% bins	Baseline
First Fit (FF)	Weibull Burst	95.1%	+3.3% bins	Baseline
Best Fit (BF)	Weibull Burst	95.7%	+2.7% bins	Baseline
Discovered Heuristic #1	Weibull Burst	95.7%	+2.7% bins	Lean 4 Verified

---

⚡ Token Optimization & Defense-in-Depth

1. Two-Tier Model Routing (80–90% Cost Reduction):
   • Routine island mutations are offloaded to local open-weight coding models (qwen2.5-coder:7b via Ollama/vLLM).
   • Frontier models (gpt-4o-mini) are reserved exclusively for supervisor reflection and population restarts.
2. Expression-Only AST Diff Prompting:
   • Prompting is restricted to mathematical return expressions (return ...), avoiding ~350 tokens of repetitive boilerplate per evaluation cycle.
3. Semantic AST Memoization:
   • Canonicalizes variable names (bin_capacity $\rightarrow c$, item $\rightarrow i$) and deduplicates structurally identical heuristics in SQLite before evaluation.
4. Staged Fail-Fast Pipeline:
   • Evaluates cheap gates first: AST Allowlist (<1ms) $\rightarrow$ WASM Smoke (<50µs) $\rightarrow$ Z3 Probe (<15ms) $\rightarrow$ Full Benchmarks (<2ms) $\rightarrow$ Lean 4 Proofs.
5. Zero-Trust AST Guard:
   • Pure mathematical allowlist (agent/ast_guard.py) strictly blocks import, exec(), eval(), open(), loops, and dunder attribute access before code execution.

---

📡 API Reference

Method	Endpoint	Description
GET	/	Web operator dashboard UI
GET	/api/status	Current heartbeat iteration and agent state
POST	/api/pause	Pause agent heartbeat loop
POST	/api/resume	Resume agent heartbeat loop
POST	/api/trigger	Trigger immediate single-cycle step
GET	/api/memory/core	Read active working scratchpad
POST	/api/memory/core	Set or update working scratchpad key/value
GET	/api/memory/heuristics	Fetch learned operational rules
GET	/api/memory/search?q=...	Hybrid RRF episodic memory search
POST	/api/funsearch/start	Launch multi-island evolutionary search
GET	/api/funsearch/telemetry	Real-time evolutionary search telemetry
POST	/api/cegis/probe	Run Z3 contract verification on heuristic code
POST	/api/verify/lean	Synthesize and check Lean 4 proof certificate
POST	/api/gitops/create-pr	Open GitHub PR for verified heuristic
GET	/api/scheduler/status	Cron schedule status and live countdown
POST	/api/scheduler/update	Update cron schedule pattern
POST	/api/scheduler/trigger-now	Launch overnight discovery pipeline immediately
GET	/api/evolution/pareto-profile	Live non-dominated Pareto frontier and island entropy diagnostics
GET	/api/evolution/policy-audit	Historical record of adaptive policy shifts and parallel beam evaluations
GET	/metrics	Prometheus exposition endpoint with live SQLite state sync
WS	/ws/telemetry	Real-time WebSocket event stream

---

📊 Production Observability & Prometheus Telemetry

Continuous Agent exposes research-grade telemetry and alerting via native Prometheus exposition on /metrics:

• Evolutionary Population Health: continuous_agent_island_fitness_best, continuous_agent_island_fitness_mean, continuous_agent_island_phenotypic_entropy, continuous_agent_island_pareto_count, and continuous_agent_island_stagnation_generations.
• AST Cache & Transpiler Efficiency: continuous_agent_ast_cache_lookups_total (hit/miss counters) and continuous_agent_ast_cache_hit_ratio.
• Formal Verification Outcomes: continuous_agent_lean_proofs_total (success, kernel_error, timeout) and continuous_agent_cegis_probes_total (verified, refuted, timeout).
• Execution Cost Profiling: continuous_agent_wasm_fuel_consumed histogram tracking instruction fuel consumption distribution.

Dashboards & Alerting Infrastructure

• Grafana Dashboard: Pre-configured definition in deploy/grafana-dashboard.json provides an 8-panel overview with Pareto convergence, entropy gauges, and verification latency.
• Prometheus Alert Rules: Pre-configured alerts in deploy/prometheus-alerts.yml covering phenotypic entropy decay (< 0.6), severe monoculture (< 0.2), prolonged fitness plateaus, Lean 4 proof kernel crashes, and process downtime.
• Promtool Unit Testing: Comprehensive synthetic unit tests in tests/test_prometheus_alerts.yml validated via CI.

---

🧬 NSGA-II Multi-Objective Optimization & Island Profiling

• Bi-Objective Pareto Dominance: Evaluates heuristics balancing Packing Ratio ($f_{\text{pack}} \in [0, 1]$, maximize) against WASM Instruction Fuel ($f_{\text{cost}}$, minimize) without arbitrary scalar weights.
• Fast Non-Dominated Sorting & Crowding Distance: Partitions populations into Pareto fronts ($F_1, F_2, \dots$) and preserves diverse boundary trade-offs via crowding distance metrics.
• Phenotypic Shannon Entropy Decay: Tracks discrete decision cluster distribution over sliding generation windows; detects diversity collapse and flatlined fitness velocity:
  • H < H_threshold $\rightarrow$ triggers Ring Migration between islands.
  • Severe stagnation ($&gt; 2W$ generations) $\rightarrow$ triggers Cataclysmic Paradigm Restart.
• Automated Schema Migration: scripts/migrate_nsga2_schema.py provides non-destructive online upgrades adding pareto_rank, crowding_distance, and composite index idx_heuristics_pareto.

---

⚡ Autonomous Supervisory Remediation & Adaptive Sampling

• Ring Migration Topology: When an island's Shannon entropy or fitness velocity plateaus, execute_ring_migration() transfers non-dominated Pareto exemplars to downstream neighbor islands ($i \to (i+1)%N$), injecting novel decision phenotypes without losing accumulated elite fitness.
• Cataclysmic Paradigm Restarts: Under severe prolonged stagnation ($&gt; 2W$ generations), execute_cataclysmic_restart() purges non-Pareto individuals, preserves top Pareto stars, and switches the island's LLM mutation paradigm across orthogonal strategies (INVERSE_FIT_DIVERGENCE, STOCHASTIC_SCATTER, COMPACT_GREEDY).
• Entropy-Driven Thermal & Multi-Parameter Sampling Policy: Synchronously scales generation parameters based on the normalized entropy deficit $D(H) = \max(0, 1 - H/H_{\text{target}})$:
  • Temperature: $T \in [0.20, 0.95]$
  • Nucleus Sampling: $top_p \in [0.70, 0.98]$
  • Mutation Beam Width: $K \in [1, 4]$ parallel mutation proposals per prompt.
• Policy Audit Trail: All adaptive parameter transitions and parallel beam candidate yields are persisted to the supervisor_policy_audit table with composite indexing for inspection via GET /api/evolution/policy-audit.

---

🔒 Security Architecture

• Isolated Storage: Local SQLite databases (agent_state.db) and ChromaDB vector collections (chroma_data/) are isolated and omitted from source control.
• Scope Restriction: Autonomous security audits enforce explicit host allowlisting (SECURITY_SCAN_ALLOWED_HOSTS).
• Deadlock-Free Lock Ordering: Distributed tool execution enforces ordered resource locks to prevent concurrency deadlocks.
• Resource Limits: WASM execution enforces explicit fuel depletion bounds; gVisor fallback enforces read-only root filesystems and process limits.
• Authentication: REST API and WebSocket channels secured via X-API-Key or Authorization: Bearer <token>.

---

🗄️ Storage Concurrency & Disaster Recovery

• SQLite WAL Mode & Concurrency: agent_state.db runs with PRAGMA journal_mode = WAL, busy_timeout = 10000, and synchronous = NORMAL, enabling lock-free concurrent reads during heavy evolutionary mutation batches.
• Zero-Downtime Hot Backups: scripts/backup_db.py leverages SQLite's native backup() API with incremental 250-page slices, PRAGMA integrity verification, gzip compression, and rolling 14-snapshot rotation.
• Automated Systemd Supervision: Automated 6-hour backups scheduled via deploy/continuous-agent-backup.service and deploy/continuous-agent-backup.timer.
• Safe Disaster Recovery: scripts/restore_db.py executes staged restore in isolated sandboxes with pre-restore active archival, --dry-run schema census, and automated rollback upon swap failure:

  # Inspect snapshot without modifying active state
  python scripts/restore_db.py --dry-run
  
  # Restore latest snapshot with confirmation prompt
  python scripts/restore_db.py
  
  # Automated CI/CD restore
  python scripts/restore_db.py --source backups/agent_state_20260905_060000Z.db.gz --yes

---

📄 License

MIT License. See LICENSE for details.