Agent7

A local AI coding agent powered by smolagents and Ollama. Run it from the terminal or via a web browser — both interfaces share the same core.

---

Folder Layout

agent7/
│
├── agent_core.py          # Shared brain — model, memory, tools, agent factory
├── Agent7.py              # Terminal interface (REPL loop, spinner, confirm prompts)
├── web_agent.py           # Web interface (FastAPI server, SSE streaming)
├── knowledge.py           # Self-teaching knowledge base (explore / discuss / save facts)
│
├── requirements.txt       # Python dependencies for the venv
├── README.md              # This file
│
├── templates/
│   └── index.html         # Browser UI — split panel layout, tabs, chat bubbles
│
├── static/
│   └── app.js             # Browser JS — SSE client, tab switching, file list
│
├── memory.json            # Auto-created — session history (what you asked, when)
├── facts.json             # Auto-created — knowledge base facts after discussions
└── agent_outputs/         # Auto-created — files generated by the agent (downloads)

Files marked auto-created are generated on first run. Do not edit them by hand unless you know what you're doing — they are plain JSON so you can inspect them freely.

---

Setup

1. Install Ollama

Download from https://ollama.com and pull a model:

ollama pull qwen2.5-coder:14b       # fast, good for most coding tasks
ollama pull qwen3-coder:30b         # slower but stronger reasoning
ollama pull deepseek-r1:14b         # good for science / knowledge tasks

2. Create the virtual environment

cd agent7
python3 -m venv venv
source venv/bin/activate            # Windows: venv\Scripts\activate
pip install -r requirements.txt

3. Run — terminal mode

python Agent7.py
python Agent7.py --model qwen3-coder:30b
python Agent7.py --model qwen2.5-coder:7b --max-steps 10 --ctx 16384
python Agent7.py --no-thinking      # hide chain-of-thought blocks

4. Run — web mode

python web_agent.py
python web_agent.py --model qwen3-coder:30b --port 8080

Then open http://localhost:7860 in your browser.

---

CLI Flags (both interfaces)

Flag	Default	Description
--model	qwen2.5-coder:14b	Ollama model name
--max-steps	15	Max reasoning steps per task
--ctx	8192	Context window size in tokens
--no-thinking	off	Hide <think> chain-of-thought blocks (terminal only)
--port	7860	Port for the web server (web only)
--host	127.0.0.1	Host for the web server (web only)

---

Terminal Commands

Command	What it does
explore: <topic>	Agent researches topic, discusses with you, saves a fact
facts list	Show all facts in the knowledge base
facts search <query>	Search facts by keyword
facts show <id>	Full detail on a specific fact
facts delete <id>	Remove a fact
quit / exit	Save session memory and exit

---

Architecture

agent_core.py          ← one source of truth
    │
    ├── Agent7.py      ← adds: Spinner, terminal confirm, REPL loop
    │
    └── web_agent.py   ← adds: FastAPI routes, SSE stream, TeeStream capture

agent_core.py defines ThinkingModel (with an on_thinking() hook) and BaseBashTool (with a confirm() hook). Each interface subclasses these to customise behaviour without duplicating logic.

---

Data Files

memory.json — grows over time, one entry per session:

[
  {
    "date": "2026-02-24 10:30",
    "prompts": ["write a snake game", "add high scores"],
    "summary": "Completed 2 task(s). Last task: add high scores"
  }
]

facts.json — grows via the explore: command or automatic gap detection:

[
  {
    "id": 1,
    "topic": "Python GIL",
    "finding": "The GIL prevents true parallel execution of Python threads...",
    "fact": "The GIL limits CPU-bound multithreading but not I/O-bound tasks.",
    "date": "2026-02-24 11:00",
    "source": "explored"
  }
]

---

Version Control Tips

Add this .gitignore to avoid committing generated files:

venv/
__pycache__/
*.pyc
agent_outputs/
memory.json
facts.json

Keep memory.json and facts.json out of git — they are personal runtime data, not source code. Back them up separately if you want to preserve your knowledge base.