Spaces:
Paused
Paused
File size: 18,817 Bytes
cb8b80f 1c96c02 911cadf cb8b80f e73dcb3 cb8b80f d00d60c cb8b80f 1c96c02 cb8b80f 1c96c02 cb8b80f 1c96c02 cb8b80f 1c96c02 cb8b80f 1c96c02 cb8b80f 1c96c02 cb8b80f 1c96c02 cb8b80f 1c96c02 cb8b80f e73dcb3 cb8b80f e73dcb3 cb8b80f e73dcb3 cb8b80f e73dcb3 cb8b80f e73dcb3 cb8b80f e73dcb3 cb8b80f e73dcb3 cb8b80f e73dcb3 cb8b80f 450c3f0 cb8b80f e73dcb3 450c3f0 e73dcb3 cb8b80f e73dcb3 cb8b80f 450c3f0 c891949 450c3f0 cb8b80f e73dcb3 cb8b80f e73dcb3 cb8b80f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 | """P2.5 agents backend: agents/teams persisted in the user's bucket, real LLM function-calling
(Inference Providers `tools=` with the user's token; fallback: JSON tool-calls via the allow-listed
Llama ZeroGPU Space on the user's quota) and a server-side job list with approval gating for
external/persisting actions. Tokens never leave the server."""
import json, re, secrets, time, tempfile
from pathlib import Path
from fastapi import APIRouter, Request, HTTPException
from huggingface_hub import HfApi, InferenceClient
PROVIDER_MODELS = ["meta-llama/Llama-3.3-70B-Instruct", "Qwen/Qwen2.5-72B-Instruct"]
LLAMA = "huggingface-projects/llama-3.2-3B-Instruct"
GEMMA = "huggingface-projects/gemma-3-12b-it"
EXTERNAL = {"workspace.save", "avatar.save", "shop.draft", "docs.save", "board.create"} # actions that persist/publish → need approval
JOBS: dict[str, dict] = {}
NOPROV: dict[str, float] = {} # token-hash -> until (provider returned 402: skip for a while) # job id -> job (scoped by user)
ID = re.compile(r"[a-z0-9][a-z0-9-]{0,47}")
def slug(s): return re.sub(r"[^a-z0-9]+", "-", (s or "").lower()).strip("-")[:48] or "x" + secrets.token_hex(3)
def to_schema(tools): # [{name, desc, params:{k:type}}] -> OpenAI tools
return [{"type": "function", "function": {"name": t["name"].replace(".", "__"), "description": t.get("desc", "")[:300],
"parameters": {"type": "object", "properties": {k: {"type": "number" if v == "number" else "string"} for k, v in (t.get("params") or {}).items()},
"required": list((t.get("params") or {}).keys())}}} for t in tools]
def parse_json(txt):
m = re.search(r"\{[\s\S]*\}", txt or "")
if not m: return None
s = m.group(0)
for end in range(len(s), 0, -1): # tolerate trailing junk
if s[end - 1] != "}": continue
try: return json.loads(s[:end])
except Exception: pass
return None
import threading
JD = ".studius/jobs/"
LOCKS: dict[str, threading.Lock] = {}
def persist(j, token):
"""Write the job's latest state to the bucket in a background thread (serialized per job, newest wins)."""
if not j.get("bucket"): return
j["_rev"] = j.get("_rev", 0) + 1
lock = LOCKS.setdefault(j["id"], threading.Lock())
def w():
with lock:
rev = j["_rev"]
if j.get("persisted", 0) >= rev: return
data = json.dumps({k: v for k, v in j.items() if k not in ("user", "_rev", "persisted", "persist_error")} | {"owner": j["user"]}, indent=1, ensure_ascii=False).encode()
try: HfApi(token=token).batch_bucket_files(j["bucket"], add=[(data, f"{JD}{j['id']}.json")]); j["persisted"] = rev; j.pop("persist_error", None)
except Exception as e: j["persist_error"] = str(e)[:120]
threading.Thread(target=w, daemon=True).start()
def load_jobs(api, bucket, user):
out = []
try:
files = sorted([f for f in api.list_bucket_tree(bucket, prefix=JD.rstrip("/"), recursive=True) if getattr(f, "type", "") == "file" and f.path.endswith(".json")], key=lambda f: str(f.mtime), reverse=True)[:40]
with tempfile.TemporaryDirectory() as t:
api.download_bucket_files(bucket, [(f.path, f"{t}/{i}.json") for i, f in enumerate(files)])
for i, _ in enumerate(files):
try: j = json.loads(Path(f"{t}/{i}.json").read_text())
except Exception: continue
if j.get("owner") == user: out.append(j)
except Exception: pass
return out
def make_agents_router(session, need, bid, client_for):
r = APIRouter()
def uid(s): return s["user"]["name"]
# ---------- agents & teams in the bucket ----------
def kind_dir(kind):
if kind not in ("agents", "teams"): raise HTTPException(400, "kind")
return f".studius/{kind}/"
@r.get("/api/crew/{kind}")
def crew_list(kind: str, req: Request, bucket: str):
s = need(req); api = HfApi(token=s["token"]); bucket = bid(bucket, s); d = kind_dir(kind)
try:
files = [f.path for f in api.list_bucket_tree(bucket, prefix=d.rstrip("/"), recursive=True) if getattr(f, "type", "") == "file" and f.path.endswith(".json")]
out = []
with tempfile.TemporaryDirectory() as t:
api.download_bucket_files(bucket, [(p, f"{t}/{i}.json") for i, p in enumerate(files)])
for i, _ in enumerate(files):
try: out.append(json.loads(Path(f"{t}/{i}.json").read_text()))
except Exception: pass
return {kind: out}
except HTTPException: raise
except Exception as e: raise HTTPException(502, str(e)[:200])
@r.post("/api/crew/{kind}")
async def crew_save(kind: str, req: Request):
s = need(req); body = await req.json(); api = HfApi(token=s["token"]); bucket = bid(body.get("bucket"), s); d = kind_dir(kind)
obj = body.get("item") or {}
obj["id"] = obj.get("id") if ID.fullmatch(str(obj.get("id", ""))) else slug(obj.get("name"))
obj["name"] = str(obj.get("name", obj["id"]))[:60]; obj["updated"] = int(time.time())
if kind == "agents":
obj["modules"] = [m for m in obj.get("modules", []) if re.fullmatch(r"[a-z]{2,20}", str(m))][:12]
obj["role"] = str(obj.get("role", ""))[:400]; obj["avatar"] = str(obj.get("avatar", ""))[:60]
else:
obj["members"] = [m for m in obj.get("members", []) if ID.fullmatch(str(m))][:8]; obj["task"] = str(obj.get("task", ""))[:300]
try: api.batch_bucket_files(bucket, add=[(json.dumps(obj, indent=1, ensure_ascii=False).encode(), f"{d}{obj['id']}.json")])
except Exception as e: raise HTTPException(502, str(e)[:200])
return {"saved": obj}
# ---------- LLM function-calling: best available model, validated plans, retries ----------
def gr_llm(space, token, prompt, system):
from gradio_client import Client # fresh session per call: chat Spaces keep history per session
c = Client(space, hf_token=token, verbose=False)
if space == GEMMA: out = c.predict({"text": prompt, "files": []}, system, 500, api_name="/run")
else: out = c.predict((system + "\n\n" if system else "") + prompt, 500, 0.2, 0.9, 50, 1.1, api_name="/generate")
return out if isinstance(out, str) else json.dumps(out)
def validate(call, tools, hist, known):
"""Return (call, problem). Coerces types, repairs unknown asset urls to the latest known one."""
t = next((x for x in tools if x["name"] == call.get("tool")), None)
if not t: return None, f"Werkzeug '{call.get('tool')}' existiert nicht. Erlaubt: {', '.join(x['name'] for x in tools)}"
args = call.get("args") if isinstance(call.get("args"), dict) else {}
params = t.get("params") or {}; fixed = []
args = {k: v for k, v in args.items() if k in params}
for k, typ in params.items():
v = args.get(k)
if typ == "number":
try: args[k] = float(v) if v not in (None, "") else None
except Exception: args[k] = None
if args[k] is None: args.pop(k)
elif k in ("url", "source"):
if v not in known and known: args[k] = known[-1]; fixed.append(k)
elif not known and not v: return None, f"'{k}' braucht die url eines vorherigen Ergebnisses – zuerst ein Werkzeug aufrufen, das eines erzeugt"
elif k == "prefix": args.setdefault(k, "")
elif not str(v or "").strip(): return None, f"Parameter '{k}' fehlt für {t['name']}"
if any(h.get("tool") == t["name"] and h.get("args") == args for h in hist): return None, f"{t['name']} wurde mit diesen Argumenten schon ausgeführt – nächster Schritt oder final"
return {"tool": t["name"], "args": args, **({"repaired": fixed} if fixed else {})}, None
@r.post("/api/agent/step")
async def step(req: Request):
"""One planning step. Body: {task, tools:[{name,desc,params}], history:[{tool,args,result}], system, known_urls, require_tool}.
Returns {call:{tool,args}} | {final}, plus via + attempts (model/validation trail)."""
s = session(req)
if not s: raise HTTPException(401, "not signed in")
b = await req.json(); tools = b.get("tools") or []
hist = b.get("history") or []; system = str(b.get("system") or "")[:2000]; task = str(b.get("task") or "")[:2000]
known = [u for u in (b.get("known_urls") or []) if isinstance(u, str)][-10:]; need_tool = bool(b.get("require_tool"))
if not tools: raise HTTPException(400, "no tools")
sysmsg = (system + "\n\n" if system else "") + "Du bist ein STUDIUS-Mitarbeiter. Erledige den Auftrag mit den Werkzeugen, Schritt für Schritt. Nutze Ergebnisse vorheriger Schritte (deren url) als Argumente. Wenn alles erledigt ist, antworte kurz ohne Werkzeug."
attempts = []
def accept(call, via):
c, prob = validate(call, tools, hist, known)
attempts.append({"via": via, "tool": call.get("tool"), "ok": not prob, **({"problem": prob} if prob else {})})
return c, prob
def final_ok(via):
ok = not need_tool
attempts.append({"via": via, "final": True, "ok": ok, **({} if ok else {"problem": "Werkzeug nötig"})}); return ok
# 1) Inference Providers with native tools (needs credits/PRO on the user's account)
th = str(hash(s["token"]))
msgs = [{"role": "system", "content": sysmsg}, {"role": "user", "content": task}]
for i, h in enumerate(hist):
msgs.append({"role": "assistant", "content": "", "tool_calls": [{"id": f"c{i}", "type": "function", "function": {"name": h["tool"].replace(".", "__"), "arguments": json.dumps(h.get("args", {}))}}]})
msgs.append({"role": "tool", "tool_call_id": f"c{i}", "content": json.dumps(h.get("result", {}))[:1500]})
for m in ([] if NOPROV.get(th, 0) > time.time() else PROVIDER_MODELS):
for k in range(2):
try:
out = InferenceClient(token=s["token"], timeout=60).chat_completion(model=m, messages=msgs, tools=to_schema(tools), tool_choice="required" if need_tool else "auto", max_tokens=400)
msg = out.choices[0].message
if msg.tool_calls:
tc = msg.tool_calls[0].function; a = tc.arguments; a = json.loads(a) if isinstance(a, str) else (a or {})
c, prob = accept({"tool": tc.name.replace("__", "."), "args": a}, f"tools:{m}")
if c: return {"call": c, "via": f"tools:{m}", "attempts": attempts}
msgs.append({"role": "user", "content": "Ungültiger Aufruf: " + prob})
elif final_ok(f"tools:{m}"): return {"final": msg.content or "", "via": f"tools:{m}", "attempts": attempts}
else: msgs.append({"role": "user", "content": "Rufe jetzt genau ein Werkzeug auf."})
except Exception as e:
attempts.append({"via": f"tools:{m}", "ok": False, "problem": str(e)[:90]})
if "402" in str(e): NOPROV[th] = time.time() + 600
break
if NOPROV.get(th, 0) > time.time(): break
# 2) JSON function-calling on ZeroGPU LLM Spaces (user's quota): strongest first
tl = "\n".join(f'- {t["name"]}({", ".join(f"{k}: {v}" for k, v in (t.get("params") or {}).items())}): {t.get("desc", "")}' for t in tools)
done = "\n".join(f'{i+1}. {h["tool"]}({json.dumps(h.get("args", {}), ensure_ascii=False)}) -> {json.dumps(h.get("result", {}), ensure_ascii=False)[:300]}' for i, h in enumerate(hist)) or "(noch nichts)"
base = (f"Werkzeuge:\n{tl}\n\nAuftrag: {task}\n\nBisher ausgeführt:\n{done}\n\nVerfügbare Ergebnis-URLs: {', '.join(known) or '(keine)'}\n\n"
'Antworte NUR mit einem JSON-Objekt in einer Zeile. Nächster Werkzeugaufruf: {"tool":"<name>","args":{...}} '
+ ('(ein Werkzeugaufruf ist jetzt Pflicht).' if need_tool else 'oder, wenn der Auftrag vollständig erledigt ist: {"final":"<kurze Zusammenfassung>"}.'))
for space, label in ((GEMMA, "json:gemma-3-12b"), (LLAMA, "json:llama-3.2-3B")):
prompt = base
for k in range(2):
try: txt = gr_llm(space, s["token"], prompt, sysmsg)
except Exception as e:
attempts.append({"via": label, "ok": False, "problem": str(e)[:110]}); break
j = parse_json(txt) or {}
if j.get("tool"):
c, prob = accept(j, label)
if c: return {"call": c, "via": label, "attempts": attempts}
elif "final" in j and final_ok(label): return {"final": str(j["final"]), "via": label, "attempts": attempts}
else:
prob = "Werkzeug nötig" if "final" in j else "keine gültige JSON-Antwort"
if "final" not in j: attempts.append({"via": label, "ok": False, "problem": prob, "raw": txt[:120]})
prompt = base + f"\n\nDein letzter Vorschlag war ungültig: {prob}. Korrigiere ihn."
raise HTTPException(502, "Kein Planer-Modell lieferte einen gültigen Plan: " + "; ".join(f"{a['via']}: {a.get('problem','')}" for a in attempts)[-400:])
# ---------- server-side job list with approval ----------
def restore(s, bucket):
for j in load_jobs(HfApi(token=s["token"]), bucket, uid(s)):
if j["id"] not in JOBS:
j["user"] = j.pop("owner", uid(s)); j["restored"] = True
if j.get("state") in ("running", "awaiting_approval"): j["state"] = "interrupted" # server restarted mid-run; runner (browser) is gone
JOBS[j["id"]] = j
def job_of(s, jid, bucket=None):
j = JOBS.get(jid)
if not j and bucket: restore(s, bid(bucket, s)); j = JOBS.get(jid)
if not j or j["user"] != uid(s): raise HTTPException(404, "job")
return j
def pub(j): return {k: v for k, v in j.items() if k not in ("user", "_rev")} | {"saved": j.get("persisted", 0) >= j.get("_rev", 0)}
def save(s, j): persist(j, s["token"]); return pub(j)
@r.get("/api/jobs")
def jobs(req: Request, bucket: str = ""):
s = need(req)
if bucket: restore(s, bid(bucket, s))
return {"jobs": [pub(j) for j in sorted(JOBS.values(), key=lambda x: -x["created"]) if j["user"] == uid(s) and (not bucket or j.get("bucket") == bucket)][:50]}
@r.post("/api/jobs")
async def job_new(req: Request):
s = need(req); b = await req.json(); jid = "j" + secrets.token_hex(5)
bk = bid(b["bucket"], s) if b.get("bucket") else ""
JOBS[jid] = {"id": jid, "user": uid(s), "bucket": bk, "task": str(b.get("task", ""))[:500], "agent": str(b.get("agent", ""))[:60], "team": str(b.get("team", ""))[:60],
"state": "running", "steps": [], "created": int(time.time())}
return save(s, JOBS[jid])
@r.post("/api/jobs/{jid}/step")
async def job_step(jid: str, req: Request):
"""Register a tool call. External actions are held as 'awaiting_approval' until approved."""
b = await req.json(); s = need(req); j = job_of(s, jid, b.get("bucket")); tool = str(b.get("tool", ""))[:60]
st = {"n": len(j["steps"]) + 1, "tool": tool, "agent": str(b.get("agent", ""))[:60], "args": b.get("args") or {}, "external": tool in EXTERNAL,
"state": "awaiting_approval" if tool in EXTERNAL else "approved", "t": int(time.time())}
j["steps"].append(st)
if st["external"]: j["state"] = "awaiting_approval"
return {"step": st, "job": save(s, j)}
@r.post("/api/jobs/{jid}/steps/{n}/{decision}")
async def job_decide(jid: str, n: int, decision: str, req: Request, bucket: str = ""):
try: body = await req.json()
except Exception: body = {}
s = need(req); j = job_of(s, jid, bucket or body.get("bucket"))
if decision not in ("approve", "reject", "done", "error"): raise HTTPException(400, "decision")
if not 0 < n <= len(j["steps"]): raise HTTPException(404, "step")
st = j["steps"][n - 1]
if decision in ("approve", "reject"):
if st["state"] != "awaiting_approval": raise HTTPException(409, "not awaiting approval")
st["state"] = "approved" if decision == "approve" else "rejected"; st["decided"] = int(time.time())
j["state"] = "running" if decision == "approve" else "rejected"
else:
if st["state"] != "approved": raise HTTPException(409, "step not approved")
st["state"] = decision; st["finished"] = int(time.time())
if isinstance(body.get("result"), dict): st["result"] = {k: (v[:600] if isinstance(v, str) else v) for k, v in body["result"].items() if k in ("kind", "url", "path", "text", "name", "error")}
return {"step": st, "job": save(s, j)}
@r.get("/api/jobs/{jid}/steps/{n}")
def job_poll(jid: str, n: int, req: Request, bucket: str = ""):
s = need(req); j = job_of(s, jid, bucket)
if not 0 < n <= len(j["steps"]): raise HTTPException(404, "step")
return {"step": j["steps"][n - 1], "job": pub(j)}
@r.post("/api/jobs/{jid}/cursor")
async def job_cursor(jid: str, req: Request):
"""Team progress pointer: which member is working (for resume)."""
b = await req.json(); s = need(req); j = job_of(s, jid, b.get("bucket"))
j["cursor"] = {"member": int(b.get("member", 0)), "agent": str(b.get("agent", ""))[:60]}
if isinstance(b.get("members"), list): j["members"] = [str(m)[:60] for m in b["members"]][:8]
return save(s, j)
@r.post("/api/jobs/{jid}/resume")
async def job_resume(jid: str, req: Request):
"""Resume an interrupted/errored job: completed steps are kept, unfinished ones are superseded."""
b = await req.json(); s = need(req); j = job_of(s, jid, b.get("bucket"))
if j["state"] not in ("interrupted", "error", "running", "awaiting_approval"): raise HTTPException(409, f"job is {j['state']}")
for st in j["steps"]:
if st["state"] in ("awaiting_approval", "approved"): st["state"] = "superseded"
j["state"] = "running"; j["resumed"] = j.get("resumed", 0) + 1; j.pop("finished", None)
return save(s, j)
@r.post("/api/jobs/{jid}/finish")
async def job_finish(jid: str, req: Request):
b = await req.json(); s = need(req); j = job_of(s, jid, b.get("bucket"))
j["state"] = b.get("state") if b.get("state") in ("done", "error", "rejected") else "done"; j["summary"] = str(b.get("summary", ""))[:1000]; j["finished"] = int(time.time())
return save(s, j)
return r
|