130 lines
4.1 KiB
Python
130 lines
4.1 KiB
Python
# agent.py — Agent loop dan tool execution.
|
|
# submit() adalah entry point: membaca input buffer, mengirim ke LLM,
|
|
# memproses tool calls, dan menampilkan hasil di log.
|
|
|
|
import json
|
|
from datetime import datetime
|
|
from .render import draw
|
|
from scripts import ntro
|
|
|
|
|
|
def log(app, role, text):
|
|
# Simpan item ke app.log untuk di-render oleh render.py
|
|
app.log.append({
|
|
"role": role,
|
|
"text": text,
|
|
"time": datetime.now().strftime("%H:%M"),
|
|
})
|
|
|
|
|
|
def submit(app, stdscr):
|
|
# Kirim query dari input buffer ke LLM.
|
|
# Loop sampai LLM mengembalikan final answer (tanpa tool_calls)
|
|
# atau mencapai max_iterations.
|
|
|
|
query = "\n".join(app.input_buffer).strip()
|
|
if not query:
|
|
return
|
|
|
|
log(app, "user", query)
|
|
# Reset input buffer
|
|
app.input_buffer = [""]
|
|
app.input_line = 0
|
|
app.input_col = 0
|
|
app.scroll = 999999 # scroll ke paling bawah
|
|
app.processing = True
|
|
draw(app, stdscr)
|
|
stdscr.refresh()
|
|
|
|
app.messages.append({"role": "user", "content": query})
|
|
|
|
stamp = ntro.start()
|
|
|
|
for step in range(app.agent_max_iterations):
|
|
stamp_step = ntro.start()
|
|
log(app, "system", f" step {step + 1} \u2014 Thinking...")
|
|
app.scroll = 999999
|
|
draw(app, stdscr)
|
|
stdscr.refresh()
|
|
|
|
response = app.llm.chat(app.messages, tools=app.TOOLS)
|
|
|
|
# Hapus "step N — LLM..." log, ganti dengan hasil aktual
|
|
app.log.pop()
|
|
|
|
if response.tool_calls:
|
|
# LLM meminta menjalankan tool(s)
|
|
amsg = {
|
|
"role": "assistant",
|
|
"content": response.content,
|
|
"tool_calls": response.tool_calls,
|
|
}
|
|
app.messages.append(amsg)
|
|
if response.content and response.content.strip():
|
|
log(app, "ai", response.content)
|
|
app.scroll = 999999
|
|
draw(app, stdscr)
|
|
stdscr.refresh()
|
|
for tc in response.tool_calls:
|
|
tname = tc["function"]["name"]
|
|
log(app, "system", f" \u2192 {tname}")
|
|
app.scroll = 999999
|
|
draw(app, stdscr)
|
|
stdscr.refresh()
|
|
execute_tool(app, tc)
|
|
else:
|
|
# Final answer — tidak ada tool_calls
|
|
if response.content:
|
|
app.messages.append({
|
|
"role": "assistant",
|
|
"content": response.content,
|
|
})
|
|
log(app, "ai", response.content)
|
|
log(app, "sep", "")
|
|
app.processing = False
|
|
app.scroll = 999999
|
|
draw(app, stdscr)
|
|
stdscr.refresh()
|
|
ntro.end(stamp)
|
|
return
|
|
ntro.end(stamp_step)
|
|
|
|
# Timeout — max iterations tercapai tanpa final answer
|
|
log(app, "error", "Max iterations reached without final answer.")
|
|
app.messages.append({"role": "assistant",
|
|
"content": "Max iterations reached without final answer."})
|
|
app.processing = False
|
|
ntro.end(stamp)
|
|
|
|
|
|
def execute_tool(app, tool_call):
|
|
# Dispatch tool_call ke handler yang terdaftar di TOOL_HANDLERS.
|
|
# search_code dan git_operation butuh penanganan argumen khusus.
|
|
tname = tool_call["function"]["name"]
|
|
targs = json.loads(tool_call["function"]["arguments"])
|
|
handler = app.TOOL_HANDLERS.get(tname)
|
|
|
|
if not handler:
|
|
result = f"Tool {tname} not found"
|
|
else:
|
|
try:
|
|
if tname == "search_code":
|
|
result = handler(
|
|
pattern=targs["pattern"],
|
|
search_type=targs["search_type"],
|
|
path=targs.get("path", "."),
|
|
)
|
|
elif tname == "git_operation":
|
|
result = handler(args=targs["args"])
|
|
else:
|
|
result = handler(**targs)
|
|
except Exception as e:
|
|
result = f"Error executing tool: {str(e)}"
|
|
|
|
# Hasil tool disimpan ke messages agar bisa dikirim balik ke LLM
|
|
app.messages.append({
|
|
"role": "tool",
|
|
"tool_call_id": tool_call["id"],
|
|
"content": str(result),
|
|
})
|