README.md

---
language:
- en
license: apache-2.0
tags:
- browser-use
- web-agent
- recurrent-moe
- edge-llm
- lightpanda
- obscura
- multi-agent
- robotics-web
pipeline_tag: text-generation
library_name: transformers
---

# Canopy-258M-R3 v6: Autonomous Tri-Engine Browser Agent & Swarm Coordinator

**Canopy-258M-R3 v6** is a 258.56M parameter Recurrent Mixture-of-Experts (MoE) browser automation agent optimized for high-speed edge navigation, anti-bot stealth, and verified persistence state read-backs.

v6 introduces a modular **Tri-Engine Execution Ecosystem**, an invariant **URL Token Normalization Preprocessor**, and deep syntheses of 2026 frontier multi-agent reasoning literature.

---

## Key Innovations in v6

### 1. Cooperative Tri-Engine Browser Swarm
Modern web navigation presents divergent requirements: speed, visual fidelity, and anti-bot evasion. v6 unifies three specialized engines via `TriEngineSwarmCoordinator`:
- **Lightpanda Engine** (*Headless Zig*): Ultra-fast headless execution (~27.5 MB standalone RSS, ~15ms cold navigation, 2.04x speedup). Deployed for high-throughput initial crawling, link harvesting, and structural AXTree indexing.
- **Obscura Engine** (*Native Rust CDP*): Stealth browser with native TLS fingerprint impersonation, Canvas/Audio spoofing, and native PNG rasterization (~49.3 MB RSS vs ~656 MiB Chromium). Deployed for visual Set-of-Marks grounding and form submissions on protected sites.
- **Chromium Engine**: Desktop Blink fallback for heavy client-side single-page applications.

### 2. URL Token Normalization Preprocessor
As documented in controlled sensitivity evaluations, small language models (258M) exhibit sensitivity to ephemeral address tokens (such as random loopback ports `http://127.0.0.1:40809` vs `http://127.0.0.1:33987`), which alter prompt tokenizations and can divert causal attention. 
v6 implements invariant URL normalization:
$$\text{URL}_{\text{ephemeral}} \to \text{http://app.local/path}$$
The agent reasoning context sees deterministic canonical tokens, while the execution layer resolves targets to live page origins transparently.

### 3. Frontier Multi-Agent Reasoning Syntheses
- **Thought Communication Bus** (*CMU / Meta AI*): Disentangles shared coordination thoughts $\hat{Z}_{\text{shared}}$ from agent-private intent $\hat{Z}_{\text{private}}$, avoiding brittle string serialization.
- **Flow Reasoning Refiner** (*Georgia Tech / MIT, EqR*): Iterative recurrent flow refinement toward stable spatial coordinate attractors, eliminating coordinate hallucinations.
- **Graph Machine DOM Referral Engine** (*Iter Labs*): $O(n)$ referral graph with dynamic 2-hop pointer chasing for $O(1)$ element retrieval.
- **Saved-Record State Verifier**: Explicit goal-purpose/field/value binding and persistent read-back contract, eliminating false completion claims ($18 \to 0$).

---

## Live Performance & Swarm Telemetry

| Engine / Modality | Memory (RSS) | Navigation / Step Latency | Visual Receipts | Anti-Bot Stealth |
| :--- | :---: | :---: | :---: | :---: |
| **Lightpanda (Scraping)** | **27.5 MB** | **~15 - 35 ms (2.04x speedup)** | No (Text Only) | Basic |
| **Obscura (Interactions)** | **49.3 MB** | **~85 - 120 ms** | **Yes (Full PNG SoM)** | **Advanced (Sannysoft 100%)** |
| **Chromium (Baseline)** | 656.0 MiB | ~180 - 320 ms | Yes | Standard CDP |
| **Cooperative Swarm** | **<80 MB combined** | **Optimal Split** | **Yes** | **Active** |

---

## Model Architecture Specifications

| Hyperparameter | Value | Description |
| :--- | :--- | :--- |
| **Total Parameters** | **258,555,654** | Standalone weights with tied embeddings |
| **Active Parameters** | **~112,000,000** | Active parameter compute per token |
| **Recurrent Layers** | **18 effective layers** | 3 Prelude + 6 Recurrent (visited 2x) + 3 Coda |
| **Recurrent Scaling** | **$1/\sqrt{2} \approx 0.7071$** | SMELT recurrence variance stabilization |
| **KV-Cache Engine** | **Prefix Sliding** | 128 prefix tokens + 512 sliding window tokens |
| **MoE Routing** | **Top-2 of 8 Experts** | Dense first 3 layers, MoE middle/coda layers |
| **Context Window** | **2,048 tokens** | RoPE position embeddings |
| **Vocabulary Size** | **49,152** | Byte-level BPE tokenizer (Cosmo-2) |

---

## Quickstart: Python Swarm Inference

```python
import asyncio
from miniswardbower.browser.swarm_coordinator import TriEngineSwarmCoordinator
from miniswardbower.core.schemas import BrowserAction, BrowserActionType

async def run_swarm():
    swarm = TriEngineSwarmCoordinator()
    try:
        # Stage 1: High-speed scraping via Lightpanda (~25ms)
        tree, text = await swarm.scrape_and_index("https://news.ycombinator.com")
        print(f"Indexed {len(tree.elements)} interactive elements via Lightpanda.")

        # Stage 2: Stealth interaction & visual audit via Obscura
        actions = [
            BrowserAction(op=BrowserActionType.TYPE, target="input[name='q']", text="Canopy MoE"),
            BrowserAction(op=BrowserActionType.PRESS, key="Enter")
        ]
        results, receipt = await swarm.visual_interact_and_submit(
            "https://news.ycombinator.com", actions, capture_audit_screenshot=True
        )
        print(f"Executed {len(results)} actions. Visual receipt: {len(receipt)} bytes.")
    finally:
        await swarm.close_all()

asyncio.run(run_swarm())
```

---

## License
Released under the **Apache 2.0** License.