Upload SKILLS/kilo_agent/references/howto.md

SKILLS/kilo_agent/references/howto.md

@@ -0,0 +1,69 @@
+# Kilo Agent — How To
+
+Purpose
+- Use Kilo (headless DeepSeek CLI) as the reasoning engine for ARC-AGI tasks.
+- Kilo identifies transforms, shapes, and parameters from rendered ARC grid images.
+- Output is structured markdown that feeds into the NN executor.
+
+## Setup
+
+```bash
+# Install Kilo CLI (if not already)
+# See: https://kilo.org/docs/install
+
+# Verify
+kilo --version
+```
+
+## Usage Patterns
+
+### 1. Direct CLI (headless, free tier)
+```bash
+kilo run "Analyze this ARC-AGI task image. Identify the transformation rule." \
+  --image task_007bbfb7.png \
+  --format default
+```
+
+### 2. Via SDK / tunnel (Kaggle → local server)
+```python
+from kilo_sdk import KiloClient
+
+client = KiloClient(server_url="http://localhost:8765")  # tunneled
+response = client.run(
+    prompt="Analyze this ARC-AGI task. Output transform name and parameters.",
+    image="task_007bbfb7.png",
+    format="default"
+)
+```
+
+### 3. Structured Output Format
+
+Kilo should output:
+
+```markdown
+## Transform
+name: kron_self_similar
+
+## Parameters
+- scale: 3
+- input_shape: [3, 3]
+- output_shape: [9, 9]
+
+## Objects
+- object_1: {shape: "square", color: 4, position: [0, 0], size: [3, 3]}
+```
+
+The NN executor parses this markdown to configure the transform.
+
+## Verification
+
+```bash
+# Test that Kilo can see and analyze an ARC task
+python SKILLS/kilo_agent/references/examples/test_kilo_arc.py
+```
+
+## Notes
+
+- Free tier has latency (~10-60s per request). Batch tasks for overnight runs.
+- Headless mode (`kilo run`) is faster than API — no HTTP overhead.
+- Image rendering: use `SKILLS/arc_agi_project/references/examples/minimal_runner.py` color scheme.