feat: improve agent behavior && add new physics knowledge (#34)

Signed-off-by: wiseaidev <oss@wiseai.dev>

Author: wiseaidev

.gitignore

@@ -35,4 +35,7 @@ build/
 *.egg-info/
 dist/
 wheels/
-*.so
+*.so
+recordings/
+server.*
+environment_files/

Cargo.lock

@@ -7,6 +7,8 @@ members = [
 ]
 exclude = [
     "target",
+    "examples/arc-lmm-agent",
+    "examples/chat",
 ]
 
 [workspace.dependencies]

Cargo.toml

@@ -5,7 +5,7 @@
 [![LMM](https://wiseai.dev/assets/logo.png)](https://wiseai.dev)
 
 [![Work In Progress](https://img.shields.io/badge/Work%20In%20Progress-orange)](https://github.com/wiseaidotdev/lmm)
-[![ASI](https://img.shields.io/badge/ASI-10.71%25-brown)](https://arcprize.org/replay/8471c865-4c54-40c5-a523-dcaa681aa4f1)
+[![ASI (Best Run)](https://img.shields.io/badge/ASI-14.55%25-brown)](https://arcprize.org/replay/69c86b04-c9ff-4ae2-98e8-eade2e4c2214)
 [![Crates.io](https://img.shields.io/crates/v/lmm.svg)](https://crates.io/crates/lmm)
 [![Docs.rs](https://docs.rs/lmm/badge.svg)](https://docs.rs/lmm)
 [![Crates.io Downloads](https://img.shields.io/crates/d/lmm)](https://crates.io/crates/lmm)

README.md

@@ -15,7 +15,7 @@ name = "arc-lmm-agent"
 path = "src/main.rs"
 
 [dependencies]
-lmm-agent = { path = "../../lmm-agent" }
+lmm-agent = { version = "0.1.2" }
 arc-agi-rs = { version = "0.1.0" }
 tokio = { version = "1.52.1", features = ["full"] }
 anyhow = "1.0.102"
@@ -24,6 +24,6 @@ serde_json = "1.0.149"
 tracing = "0.1"
 tracing-subscriber = { version = "0.3", features = ["fmt", "env-filter"] }
 clap = { version = "4.6.1",  features = ["derive"] }
-rand = "0.8.5"
-indicatif = "0.18"
-owo-colors = "4"
+rand = "0.10.1"
+indicatif = "0.18.4"
+owo-colors = "4.3.0"

examples/arc-lmm-agent/Cargo.lock

@@ -2,10 +2,10 @@
 
 # 🕹️ arc-lmm-agent
 
-[![ASI](https://img.shields.io/badge/ASI-10.71%25-brown)](https://arcprize.org/replay/8471c865-4c54-40c5-a523-dcaa681aa4f1)
+[![ASI (Best Run)](https://img.shields.io/badge/ASI-14.55%25-brown)](https://arcprize.org/replay/69c86b04-c9ff-4ae2-98e8-eade2e4c2214)
 [![MIT License](https://img.shields.io/badge/license-MIT-blue.svg)](../../LICENSE)
 
-[![ls20-arc-lmm.gif](./assets/ls20-arc-lmm.gif)](https://arcprize.org/replay/8471c865-4c54-40c5-a523-dcaa681aa4f1)
+[![ls20-arc-lmm.gif](./assets/ls20-arc-lmm.gif)](https://arcprize.org/replay/69c86b04-c9ff-4ae2-98e8-eade2e4c2214)
 
 > `arc-lmm-agent` is an autonomous navigation solver for ARC-AGI interactive environments (`ls20` game atm). It uses an episodic framework, progressive strategy learning, and robust world modeling to dynamically maneuver through complex grids, interact with rotation modifiers, systematically collect step-boosters, and reach the target zones across escalating levels.
 
@@ -94,8 +94,8 @@ When all else fails (no plan, no known targets, nothing visible on radar), the a
 The solver natively utilizes the overarching `lmm-agent` architecture for generalized intelligence logic:
 
 1. **`InternalDrive`**: The agent fires intrinsic reward/motivation signals. If the agent finds a new bonus position or discovers a completely unvisited tile, the `Curiosity` drive spikes. If the agent bumps into a newly discovered wall and loses a turn, the `Incoherence` drive registers the penalty, adjusting future behavioral tolerances.
-2. **`KnowledgeIndex` (Cross-Level Transfer)**: As the agent completes `Level N`, it synthesizes the trial's metadata into narrative English (e.g. *"Level 0 completed after 1 mod interactions and 0 bonuses... "*). This raw text is dynamically ingested into the localized `KnowledgeIndex`. When `Level N+1` begins, this long-term semantic memory primes the agent about the nature of the puzzles it will likely encounter.
-3. **`LearningEngine` (HELM)**: Traditional tabular Q-learning shapes underlying values. The agent emits a continuous localized Bellman reward stream (+10 for activating a modifier, +50 for moving closer to the target post-modifier, -1.0 for wall collisions) to fine-tune the `NOVELTY` fallback recommendations.
+1. **`KnowledgeIndex` (Cross-Level Transfer)**: As the agent completes `Level N`, it synthesizes the trial's metadata into narrative English (e.g. *"Level 0 completed after 1 mod interactions and 0 bonuses... "*). This raw text is dynamically ingested into the localized `KnowledgeIndex`. When `Level N+1` begins, this long-term semantic memory primes the agent about the nature of the puzzles it will likely encounter.
+1. **`LearningEngine` (HELM)**: Traditional tabular Q-learning shapes underlying values. The agent emits a continuous localized Bellman reward stream (+10 for activating a modifier, +50 for moving closer to the target post-modifier, -1.0 for wall collisions) to fine-tune the `NOVELTY` fallback recommendations.
 
 
 ## 🕹️ Run the agent

examples/arc-lmm-agent/Cargo.toml

@@ -400,3 +400,36 @@ pub fn print_plan_invalidated() {
         "Plan invalidated (wall/unavailable)".dimmed()
     );
 }
+
+/// Prints when a launch-pedal is detected from a large position jump.
+pub fn print_pedal_detected(from: (usize, usize), to: (usize, usize), delta: usize) {
+    eprintln!(
+        "  {} {} {} → {} {}",
+        "⚡".yellow().bold(),
+        "PEDAL detected".bright_yellow().bold(),
+        format!("({},{})", from.0, from.1).bright_white(),
+        format!("({},{})", to.0, to.1).bright_white(),
+        format!("Δ={delta}px").dimmed(),
+    );
+}
+
+/// Prints when a colorful novel object is discovered during exploration.
+pub fn print_novel_object_found(pos: (usize, usize)) {
+    eprintln!(
+        "  {} {} {}",
+        "🎨".dimmed(),
+        "Novel object found".bright_cyan().bold(),
+        format!("at ({},{})", pos.0, pos.1).bright_white(),
+    );
+}
+
+/// Prints when the agent learns that touching a novel object re-colors the target.
+pub fn print_novel_object_learned() {
+    eprintln!(
+        "  {} {}",
+        "🧩".dimmed(),
+        "LEARNED: novel object changes target color!"
+            .bright_green()
+            .bold(),
+    );
+}