dev-agent.py

import os
import sys
import json
import textwrap
from pathlib import Path

# Detectar ambiente
def is_huggingface():
    return "SPACE_ID" in os.environ or "HF_HOME" in os.environ

RUNNING_IN_HF = is_huggingface()

# ==========================
# MODELOS
# ==========================

LOCAL_MODEL = "qwen3.5:latest"
HF_MODEL = "google/gemma-2-2b-it"

# ==========================
# MODELO LOCAL (OLLAMA)
# ==========================

def call_ollama(prompt):
    import requests
    try:
        resp = requests.post(
            "http://localhost:11434/api/generate",
            json={"model": LOCAL_MODEL, "prompt": prompt, "stream": False},
            timeout=30
        )
        return resp.json().get("response", "")
    except:
        return None

# ==========================
# MODELO HUGGING FACE (TRANSFORMERS)
# ==========================

hf_pipeline = None

def load_hf_model():
    global hf_pipeline
    if hf_pipeline is None:
        from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
        tok = AutoTokenizer.from_pretrained(HF_MODEL)
        model = AutoModelForCausalLM.from_pretrained(HF_MODEL)
        hf_pipeline = pipeline("text-generation", model=model, tokenizer=tok)
    return hf_pipeline

def call_hf_transformers(prompt):
    pipe = load_hf_model()
    out = pipe(prompt, max_new_tokens=256)
    return out[0]["generated_text"]

# ==========================
# FALLBACK
# ==========================

def call_model(prompt):
    # 1) Tentar Ollama local
    if not RUNNING_IN_HF:
        r = call_ollama(prompt)
        if r:
            return r

    # 2) Tentar Gemma local (Transformers)
    try:
        return call_hf_transformers(prompt)
    except:
        pass

    # 3) Fallback interno
    return "Não consegui usar nenhum modelo. Resposta fallback."

# ==========================
# AGENTE
# ==========================

SYSTEM_PROMPT = """
És um agente de desenvolvimento híbrido.
Cria ficheiros, lê ficheiros e corrige código.
No Hugging Face não usas subprocessos.
Responde sempre em JSON com:
{
 "thoughts": "...",
 "steps": [
   {"tool": "...", "args": {...}, "comment": "..."}
 ]
}
"""

def build_prompt(user):
    return SYSTEM_PROMPT + "\nUtilizador:\n" + user + "\nJSON:"

# Ferramentas
def tool_read(path):
    p = Path(path)
    if not p.exists():
        return "[ERRO] ficheiro não existe"
    return p.read_text()

def tool_write(path, content):
    p = Path(path)
    p.parent.mkdir(parents=True, exist_ok=True)
    p.write_text(content)
    return "[OK] escrito"

def tool_run(cmd):
    if RUNNING_IN_HF:
        return "[BLOQUEADO] subprocessos não são permitidos no Hugging Face."
    import subprocess
    r = subprocess.run(cmd, shell=True, capture_output=True, text=True)
    return r.stdout + "\n" + r.stderr

# Execução dos passos
def execute_steps(steps):
    outputs = []
    for s in steps:
        tool = s.get("tool")
        args = s.get("args", {})
        if tool == "read_file":
            outputs.append(tool_read(args["path"]))
        elif tool == "write_file":
            outputs.append(tool_write(args["path"], args["content"]))
        elif tool == "run_command":
            outputs.append(tool_run(args["command"]))
        else:
            outputs.append("[ERRO] ferramenta desconhecida")
    return "\n".join(outputs)

# ==========================
# INTERFACE GRADIO
# ==========================

import gradio as gr

def agent_chat(user_input):
    prompt = build_prompt(user_input)
    raw = call_model(prompt)

    try:
        data = json.loads(raw)
    except:
        return "Modelo não devolveu JSON válido:\n" + raw

    steps = data.get("steps", [])
    out = execute_steps(steps)
    return out

with gr.Blocks() as demo:
    gr.Markdown("# Dev Agent Híbrido (Local + Hugging Face)")
    inp = gr.Textbox(label="Comando")
    out = gr.Textbox(label="Output")
    btn = gr.Button("Executar")
    btn.click(agent_chat, inp, out)

demo.launch()