Nottybro's picture
Add daily token cap: 2.5M tokens per UTC day
8c37db9
Raw
History Blame Contribute Delete
9.62 kB
import os, json, time, tempfile, httpx
from fastapi import FastAPI, Request
from fastapi.responses import JSONResponse, StreamingResponse, HTMLResponse
import uvicorn
app = FastAPI()
KIMI_BASE = os.environ.get("KIMI_BASE_URL", "https://api.kimi.com/coding").rstrip("/")
DEFAULT_MODEL = os.environ.get("KIMI_DEFAULT_MODEL", "k3")
DAILY_CAP = int(os.environ.get("DAILY_TOKEN_CAP", "2500000"))
USAGE_FILE = os.path.join(tempfile.gettempdir(), "wirelessaudit_token_usage.json")
_usage = {"date": "", "used": 0}
def _today():
return time.strftime("%Y-%m-%d", time.gmtime())
def _load_usage():
try:
with open(USAGE_FILE) as f:
d = json.load(f)
if d.get("date") == _today() and isinstance(d.get("used"), int):
_usage.update(d)
except Exception:
pass
def _save_usage():
try:
with open(USAGE_FILE, "w") as f:
json.dump(_usage, f)
except Exception:
pass
def usage_today():
"""Tokens used today (UTC); rolls the counter at the day boundary."""
if _usage["date"] != _today():
_usage["date"] = _today()
_usage["used"] = 0
_save_usage()
return _usage["used"]
def usage_add(n):
usage_today()
_usage["used"] += max(0, int(n))
_save_usage()
def estimate_prompt_tokens(body):
n = 0
for m in body.get("messages", []):
c = m.get("content")
if isinstance(c, str):
n += len(c)
elif isinstance(c, list):
n += sum(len(p.get("text", "")) for p in c if isinstance(p, dict))
return n // 4 + 8
_load_usage()
def get_key():
key = os.environ.get("KIMI_API_KEY", "")
if not key:
raise ValueError("KIMI_API_KEY not set")
return key
def check_client_auth(request: Request):
"""Clients must present the proxy's own key. Returns a 401/503 response or None."""
proxy_key = os.environ.get("PROXY_KEY", "")
if not proxy_key:
return JSONResponse({"error": {"message": "PROXY_KEY not configured", "code": 503}}, status_code=503)
auth = request.headers.get("Authorization", "")
x_key = request.headers.get("x-api-key", "")
if auth != f"Bearer {proxy_key}" and x_key != proxy_key:
return JSONResponse({"error": {"message": "Invalid or missing API key", "code": 401}}, status_code=401)
return None
def auth_headers():
return {"Authorization": f"Bearer {get_key()}", "Content-Type": "application/json"}
async def kimi_alive():
"""Returns (ok, models_list | error_message)."""
try:
async with httpx.AsyncClient(timeout=8) as c:
r = await c.get(f"{KIMI_BASE}/v1/models", headers=auth_headers())
if r.status_code == 200:
return True, [m.get("id") for m in r.json().get("data", [])]
return False, f"HTTP {r.status_code}: {r.text[:200]}"
except Exception as e:
return False, str(e)
@app.get("/")
async def root():
if not os.environ.get("KIMI_API_KEY"):
alive, info = False, "KIMI_API_KEY secret not set"
else:
alive, info = await kimi_alive()
color = "#00ff88" if alive else "#ff4444"
status = "ONLINE" if alive else "OFFLINE"
models = ", ".join(info) if isinstance(info, list) else info
used = usage_today()
cap_color = "#ff4444" if used >= DAILY_CAP else "#00ff88"
return HTMLResponse(f"""<html><body style="font-family:monospace;padding:2rem;background:#0d1117;color:#c9d1d9">
<h2>WirelessAudit Proxy</h2>
<p>Backend: <code>Kimi K3</code> — <code>{KIMI_BASE}</code></p>
<p>Status: <b style="color:{color}">{status}</b></p>
<p>Models: <code>{models}</code></p>
<p>Daily tokens (UTC): <b style="color:{cap_color}">{used:,}</b> / {DAILY_CAP:,} — resets 00:00 UTC</p>
<p>Cline Base URL: <code>/v1</code> (requires API key)</p>
</body></html>""")
@app.get("/health")
async def health():
try:
get_key()
except ValueError as e:
return JSONResponse({"status": "offline", "error": str(e)}, status_code=503)
alive, info = await kimi_alive()
usage = {"used": usage_today(), "cap": DAILY_CAP}
if alive:
return {"status": "ok", "backend": "kimi", "models": info, "daily_tokens": usage}
return JSONResponse({"status": "offline", "error": info, "daily_tokens": usage}, status_code=503)
@app.get("/v1/models")
async def models(request: Request):
if (deny := check_client_auth(request)) is not None:
return deny
try:
async with httpx.AsyncClient(timeout=10) as c:
r = await c.get(f"{KIMI_BASE}/v1/models", headers=auth_headers())
return JSONResponse(r.json(), status_code=r.status_code)
except ValueError as e:
return JSONResponse({"error": str(e)}, status_code=503)
except Exception as e:
return JSONResponse({"error": str(e)}, status_code=502)
@app.post("/v1/chat/completions")
async def chat(request: Request):
if (deny := check_client_auth(request)) is not None:
return deny
try:
headers = auth_headers()
except ValueError as e:
return JSONResponse({"error": str(e)}, status_code=503)
body = await request.json()
if not body.get("model"):
body["model"] = DEFAULT_MODEL
stream = body.get("stream", False)
est_prompt = estimate_prompt_tokens(body)
used = usage_today()
if used + est_prompt >= DAILY_CAP:
return JSONResponse({"error": {"message": f"Daily token cap reached ({used:,}/{DAILY_CAP:,} tokens, UTC day). Resets at 00:00 UTC.",
"type": "rate_limit_exceeded", "code": 429}}, status_code=429)
url = f"{KIMI_BASE}/v1/chat/completions"
timeout = httpx.Timeout(300.0, connect=15.0)
try:
if stream:
so = dict(body.get("stream_options") or {})
so["include_usage"] = True
body["stream_options"] = so
c = httpx.AsyncClient(timeout=timeout)
req = c.build_request("POST", url, json=body, headers=headers)
r = await c.send(req, stream=True)
if r.status_code != 200:
err_body = await r.aread()
code = r.status_code
await r.aclose()
await c.aclose()
try:
payload = json.loads(err_body)
except Exception:
payload = {"error": {"message": err_body.decode(errors="replace"), "code": code}}
return JSONResponse(payload, status_code=code)
async def proxy():
usage = None
completion_chars = 0
buf = b""
try:
async for chunk in r.aiter_bytes():
yield chunk
buf += chunk
if len(buf) > 65536:
buf = buf[-4096:]
while b"\n" in buf:
line, buf = buf.split(b"\n", 1)
line = line.strip()
if not line.startswith(b"data:"):
continue
data = line[5:].strip()
if not data or data == b"[DONE]":
continue
try:
ev = json.loads(data)
except Exception:
continue
if ev.get("usage"):
usage = ev["usage"]
for ch in ev.get("choices") or []:
d = ch.get("delta") or {}
completion_chars += len(d.get("content") or "")
completion_chars += len(d.get("reasoning_content") or "")
finally:
await r.aclose()
await c.aclose()
if usage and usage.get("total_tokens"):
usage_add(usage["total_tokens"])
else:
usage_add(est_prompt + completion_chars // 4)
return StreamingResponse(proxy(), media_type="text/event-stream",
headers={"Cache-Control": "no-cache", "X-Accel-Buffering": "no"})
async with httpx.AsyncClient(timeout=timeout) as c:
r = await c.post(url, json=body, headers=headers)
try:
payload = r.json()
except Exception:
return JSONResponse({"error": {"message": r.text, "code": r.status_code}},
status_code=r.status_code)
if r.status_code == 200:
u = payload.get("usage") or {}
if u.get("total_tokens"):
usage_add(u["total_tokens"])
else:
out = "".join((ch.get("message") or {}).get("content") or ""
for ch in payload.get("choices", []))
usage_add(est_prompt + len(out) // 4)
return JSONResponse(payload, status_code=r.status_code)
except httpx.ConnectError:
return JSONResponse({"error": {"message": "Kimi unreachable", "code": 503}}, status_code=503)
except httpx.ReadTimeout:
return JSONResponse({"error": {"message": "Kimi timed out", "code": 504}}, status_code=504)
except Exception as e:
return JSONResponse({"error": {"message": str(e)}}, status_code=500)
if __name__ == "__main__":
uvicorn.run(app, host="0.0.0.0", port=7860)