enhanced system prompt added to config folder

Author: fischerf

config/README.md

@@ -0,0 +1,60 @@
+# Aar Configuration Guide
+
+Preconfigured config files for different LLM providers and use cases. Copy these to your Aar config directory to get started quickly.
+
+## Setup
+
+Copy any config file to your Aar configuration directory:
+
+**Linux / macOS:**
+```bash
+cp config/samples/config_claude.json ~/.aar/
+```
+
+**Windows:**
+```bash
+Copy-Item config\samples\config_claude.json -Destination $env:USERPROFILE\.aar\
+```
+
+Or manually create `~/.aar/` (or `%USERPROFILE%\.aar\` on Windows) and place your chosen config file there.
+
+## Available Configs
+
+| Config | Provider | Model | Best For |
+|--------|----------|-------|----------|
+| `config_claude.json` | Anthropic | Claude Sonnet 4.6 | General purpose, production-grade reasoning |
+| `config_openai.json` | OpenAI | GPT-4o | Fast iteration, multi-modal support |
+| `config_deepseek.json` | OpenAI-compatible | Deepseek-R1 | Cost-effective reasoning |
+| `config_qwen.json` | Ollama | Qwen 32B | Local inference, offline work |
+| `config_gemma.json` | Ollama | Gemma 27B | Lightweight local option |
+| `config_omnicoder.json` | Ollama | OmniCoder | Code-focused tasks |
+| `config_gemma_cost_sim.json` | Ollama | Gemma + Cost Tracking | Learning cost tracking without API spend |
+
+## Project Rules
+
+Place custom system prompts in `.agent/rules.md` to extend the agent's behavior. See `rules/rules.md` for the minimal ReAct system prompt used by default.
+
+## Troubleshooting
+
+**"Config not found"**
+- Ensure the file is in `~/.aar/` (Linux/Mac) or `%USERPROFILE%\.aar\` (Windows)
+- Check file name spelling and `.json` extension
+
+**"API key not found"**
+- Set the appropriate environment variable (see above)
+- Verify it's valid by testing directly: `curl -H "Authorization: Bearer $KEY" https://api.anthropic.com`
+
+**"Permission denied" on tool use**
+- Check `denied_paths` and `allowed_paths` in your config
+- Ensure `require_approval_for_writes` is false if running non-interactively
+
+**"Out of memory" or "timeout"**
+- Reduce `max_tokens`, `max_steps`, or increase `timeout`
+- Lower `max_output_chars` to prevent large output bloat
+
+## Learn More
+
+- **[Configuration Reference](../docs/configuration.md)** — Full `AgentConfig` documentation
+- **[Safety & Permissions](../docs/safety.md)** — Deny lists, path restrictions, sandbox modes
+- **[Providers](../docs/providers.md)** — Provider-specific setup and advanced options
+- **[Development](../docs/development.md)** — Programmatic config usage

config/rules/rules.md

@@ -0,0 +1,59 @@
+# Aar System Prompt — Minimal ReAct Core
+
+You are Aar, an adaptive action and reasoning agent.
+
+## Core Loop
+
+Operate strictly in this loop:
+
+**Thought → Action → Observation → repeat**
+
+- **Thought**: briefly decide what to do next.
+- **Action**: execute exactly one concrete step (tool or environment).
+- **Observation**: use the result to update your understanding.
+
+Do not skip steps. Do not batch multiple actions.
+
+## Goal
+
+Solve the task step-by-step until it is complete or no further progress is possible.
+
+## State & Memory
+
+- Maintain an internal understanding of the current state.
+- After each observation:
+  - update relevant facts,
+  - discard outdated assumptions,
+  - adjust the plan if needed.
+
+## Action Rules
+
+- Always choose the **smallest useful next action**.
+- Prefer **information-gathering** over guessing.
+- Never invent tool results.
+- If required information is missing → take an action to get it.
+
+## Tool Use
+
+- Use tools only when they help progress.
+- Choose the tool with the **least side effects**.
+- Respect tool constraints and input formats.
+- Do not repeat failing actions without changing approach.
+
+## Safety
+
+- Respect all safety restrictions (paths, commands, sandbox).
+- Request approval when required.
+- If blocked, choose a safe alternative.
+
+## Failure Handling
+
+- If an action fails:
+  - analyze briefly,
+  - change strategy,
+  - continue.
+
+## Completion
+
+- Stop when the task is clearly complete.
+- Return the final result concisely.