dota-draft / agent.py
asusevski's picture
Initial Dota draft Space
231c410
Raw
History Blame Contribute Delete
25.2 kB
from __future__ import annotations
import json
import os
import asyncio
import time
from collections import deque
from dataclasses import dataclass, field
from json import JSONDecodeError
from typing import Any, Literal
import httpx
from openai import AsyncOpenAI
from mcp import OpenDotaMCPServer
HF_ROUTER_BASE_URL = "https://router.huggingface.co/v1"
OPENAI_MODEL = os.getenv("OPENAI_MODEL", "Qwen/Qwen3.6-27B:featherless-ai")
OPENAI_CALLS_PER_MINUTE = int(os.getenv("OPENAI_CALLS_PER_MINUTE", "60"))
OPENAI_STREAM = os.getenv("OPENAI_STREAM", "").lower() in {"1", "true", "yes", "on"}
MCP_PROTOCOL_VERSION = "2025-06-18"
MAX_PROMPT_VALUE_CHARS = int(os.getenv("MAX_PROMPT_VALUE_CHARS", "4000"))
OPENAI_MAX_TOKENS = int(os.getenv("OPENAI_MAX_TOKENS", "160"))
FAST_CANDIDATE_COUNT = int(os.getenv("FAST_CANDIDATE_COUNT", "6"))
FAST_DETAIL_COUNT = int(os.getenv("FAST_DETAIL_COUNT", "2"))
FAST_MATCHUP_COUNT = int(os.getenv("FAST_MATCHUP_COUNT", "0"))
MCP_PREFETCH_TIMEOUT_SECONDS = float(os.getenv("MCP_PREFETCH_TIMEOUT_SECONDS", "4"))
AGENT_DEBUG = os.getenv("AGENT_DEBUG", "").lower() in {"1", "true", "yes", "on"}
DEFAULT_HERO_POOL = [
"Puck",
"Pudge",
"Invoker",
"Shadow Fiend",
"Rubick",
"Earthshaker",
"Storm Spirit",
"Ember Spirit",
"Crystal Maiden",
"Anti-Mage",
"Phantom Assassin",
"Juggernaut",
"Lina",
"Lion",
"Axe",
"Doom",
"Batrider",
"Beastmaster",
"Mars",
"Tidehunter",
"Dark Seer",
"Chen",
"Enchantress",
"Io",
"Disruptor",
"Shadow Demon",
]
STATIC_HERO_POOL: list[str] = []
_TOOL_RESULT_CACHE: dict[str, Any] = {}
_OPENAI_CLIENTS: dict[str, AsyncOpenAI] = {}
_MCP_CLIENTS: dict[str, OpenDotaMCPHttpClient] = {}
_CACHE_LOCK = asyncio.Lock()
class AsyncSlidingWindowRateLimiter:
def __init__(self, max_calls: int, window_seconds: float):
self.max_calls = max_calls
self.window_seconds = window_seconds
self._calls: deque[float] = deque()
self._lock = asyncio.Lock()
async def acquire(self) -> None:
if self.max_calls <= 0:
return
while True:
async with self._lock:
now = time.monotonic()
while self._calls and now - self._calls[0] >= self.window_seconds:
self._calls.popleft()
if len(self._calls) < self.max_calls:
self._calls.append(now)
return
sleep_for = self.window_seconds - (now - self._calls[0])
await asyncio.sleep(max(sleep_for, 0.05))
OPENAI_RATE_LIMITER = AsyncSlidingWindowRateLimiter(
max_calls=OPENAI_CALLS_PER_MINUTE,
window_seconds=60,
)
@dataclass(frozen=True)
class DraftTeam:
name: str
side: str
picks: list[str] = field(default_factory=list)
bans: list[str] = field(default_factory=list)
player_ids: list[str] = field(default_factory=list)
@dataclass(frozen=True)
class DraftRecommendationRequest:
action: Literal["Pick", "Ban"]
step_number: int
phase: str
slot: int
you: DraftTeam
opponent: DraftTeam
unavailable_heroes: list[str]
opendota_api_key: str = ""
class OpenDotaMCPHttpClient:
def __init__(self, server: OpenDotaMCPServer | None = None, api_key: str = ""):
self.server = server or OpenDotaMCPServer()
self.api_key = api_key
self._client: httpx.AsyncClient | None = None
self._session_id = ""
self._request_id = 0
async def __aenter__(self) -> OpenDotaMCPHttpClient:
if not self.server.is_http:
raise ValueError("Only the OpenDota MCP HTTP transport is supported by this client.")
self._client = httpx.AsyncClient(timeout=30)
await self._initialize()
return self
async def __aexit__(self, *_: object) -> None:
if self._client:
await self._client.aclose()
async def list_tools(self) -> list[dict[str, Any]]:
result = await self._request("tools/list", {})
return result.get("tools", [])
async def call_tool(self, name: str, arguments: dict[str, Any]) -> Any:
return await self._request("tools/call", {"name": name, "arguments": arguments})
async def _initialize(self) -> None:
result, headers = await self._post(
"initialize",
{
"protocolVersion": MCP_PROTOCOL_VERSION,
"capabilities": {"tools": {}},
"clientInfo": {"name": "dota-draft-agent", "version": "0.1.0"},
},
include_session=False,
)
self._session_id = headers.get("mcp-session-id", "")
if not result:
raise RuntimeError("OpenDota MCP server did not return an initialize result.")
await self._notify_initialized()
async def _notify_initialized(self) -> None:
assert self._client is not None
headers = self._headers(include_session=True)
message = {"jsonrpc": "2.0", "method": "notifications/initialized"}
await self._client.post(self.server.url, json=message, headers=headers)
async def _request(self, method: str, params: dict[str, Any]) -> Any:
result, _ = await self._post(method, params, include_session=True)
return result
async def _post(
self,
method: str,
params: dict[str, Any],
include_session: bool,
) -> tuple[Any, httpx.Headers]:
assert self._client is not None
self._request_id += 1
message = {
"jsonrpc": "2.0",
"id": self._request_id,
"method": method,
"params": params,
}
response = await self._client.post(
self.server.url,
json=message,
headers=self._headers(include_session=include_session),
)
response.raise_for_status()
payload = _decode_mcp_response(response)
if "error" in payload:
raise RuntimeError(payload["error"])
return payload.get("result"), response.headers
def _headers(self, include_session: bool) -> dict[str, str]:
headers = {
"Accept": "application/json, text/event-stream",
"Content-Type": "application/json",
}
if include_session and self._session_id:
headers["Mcp-Session-Id"] = self._session_id
if self.api_key:
headers["X-OpenDota-API-Key"] = self.api_key
return headers
async def recommend_draft_action(request: DraftRecommendationRequest) -> str:
context = await build_fast_recommendation_context(request)
try:
recommendation = await _call_openai_for_recommendation(request, context)
except Exception as exc:
return f"Recommendation unavailable: OpenAI chat completions request failed ({exc})"
if not recommendation:
return "Recommendation unavailable: the agent did not return a final pick or ban."
return _format_recommendation(request, recommendation, context.get("observations", []))
def warm_static_dota_data(hero_names: list[str] | None = None) -> None:
"""Seed static hero data at app startup without network calls."""
global STATIC_HERO_POOL
names = hero_names or DEFAULT_HERO_POOL
seen: set[str] = set()
STATIC_HERO_POOL = []
for name in [*names, *DEFAULT_HERO_POOL]:
normalized = _normalize_hero_name(name)
if normalized and normalized not in seen:
STATIC_HERO_POOL.append(name)
seen.add(normalized)
async def build_fast_recommendation_context(
request: DraftRecommendationRequest,
) -> dict[str, Any]:
candidates = _select_candidate_heroes(request)
context: dict[str, Any] = {
"candidate_heroes": candidates,
"observations": [],
"prefetch_errors": [],
}
try:
client = await _get_mcp_client(request.opendota_api_key)
context["observations"] = await asyncio.wait_for(
_prefetch_draft_observations(
client,
request,
candidates,
),
timeout=MCP_PREFETCH_TIMEOUT_SECONDS,
)
except asyncio.TimeoutError:
context["prefetch_errors"].append(
f"OpenDota MCP prefetch timed out after {MCP_PREFETCH_TIMEOUT_SECONDS:g}s"
)
except Exception as exc:
context["prefetch_errors"].append(str(exc))
if AGENT_DEBUG:
_debug_log(
"fast context",
{
"candidates": len(candidates),
"observations": len(context["observations"]),
"prefetch_errors": context["prefetch_errors"],
},
)
return context
async def _get_mcp_client(api_key: str) -> OpenDotaMCPHttpClient:
probe = OpenDotaMCPHttpClient(api_key=api_key)
cache_key = f"{probe.server.url}\0{api_key}"
async with _CACHE_LOCK:
cached = _MCP_CLIENTS.get(cache_key)
if cached is not None:
return cached
try:
client = await probe.__aenter__()
except Exception:
await probe.__aexit__(None, None, None)
raise
async with _CACHE_LOCK:
cached = _MCP_CLIENTS.get(cache_key)
if cached is not None:
await client.__aexit__(None, None, None)
return cached
_MCP_CLIENTS[cache_key] = client
return client
def _select_candidate_heroes(
request: DraftRecommendationRequest,
limit: int | None = FAST_CANDIDATE_COUNT,
) -> list[str]:
unavailable = {_normalize_hero_name(hero) for hero in request.unavailable_heroes}
pool = STATIC_HERO_POOL or DEFAULT_HERO_POOL
candidates: list[str] = []
candidate_names: set[str] = set()
priority_names = [
*request.opponent.picks,
*request.you.picks,
*pool,
]
for hero in priority_names:
normalized = _normalize_hero_name(hero)
if not normalized or normalized in unavailable:
continue
if normalized in candidate_names:
continue
candidates.append(hero)
candidate_names.add(normalized)
if limit is not None and len(candidates) >= limit:
break
return candidates
def _normalize_hero_name(name: str) -> str:
return "".join(char for char in name.lower() if char.isalnum())
async def _prefetch_draft_observations(
client: OpenDotaMCPHttpClient,
request: DraftRecommendationRequest,
candidates: list[str],
) -> list[dict[str, Any]]:
observations: list[dict[str, Any]] = []
detail_targets = candidates[:FAST_DETAIL_COUNT]
matchup_targets = [
*request.opponent.picks,
*request.you.picks,
*candidates[:1],
][:FAST_MATCHUP_COUNT]
detail_tasks = [
_call_tool_observation(
client,
"get_hero_details",
{"hero": hero},
compact_keys=(
"id",
"localized_name",
"primary_attr",
"attack_type",
"roles",
"base_health",
"base_mana",
"base_armor",
"base_attack_min",
"base_attack_max",
"move_speed",
),
)
for hero in detail_targets
]
relevant_names = {
_normalize_hero_name(hero)
for hero in [
*request.you.picks,
*request.opponent.picks,
*candidates,
]
}
matchup_tasks = [
_call_tool_observation(client, "get_hero_matchups", {"hero": hero})
for hero in matchup_targets
]
detail_observations, matchup_observations = await asyncio.gather(
asyncio.gather(*detail_tasks),
asyncio.gather(*matchup_tasks),
)
observations.extend(detail_observations)
for observation in matchup_observations:
observation["result"] = _compact_matchups(observation.get("result"), relevant_names)
observations.append(observation)
return observations
async def _call_tool_observation(
client: OpenDotaMCPHttpClient,
name: str,
arguments: dict[str, Any],
compact_keys: tuple[str, ...] = (),
) -> dict[str, Any]:
try:
result = await _call_tool_cached(client, name, arguments)
result = _normalize_tool_result(result)
if compact_keys and isinstance(result, dict):
result = {key: result.get(key) for key in compact_keys if key in result}
return {"tool": name, "arguments": arguments, "result": result}
except Exception as exc:
return {"tool": name, "arguments": arguments, "result": {"tool_error": str(exc)}}
async def _call_tool_cached(
client: OpenDotaMCPHttpClient,
name: str,
arguments: dict[str, Any],
) -> Any:
cache_key = json.dumps(
{
"server": client.server.url,
"tool": name,
"arguments": arguments,
},
sort_keys=True,
ensure_ascii=True,
)
async with _CACHE_LOCK:
cached = _TOOL_RESULT_CACHE.get(cache_key)
if cached is not None:
return cached
result = await client.call_tool(name, arguments)
async with _CACHE_LOCK:
_TOOL_RESULT_CACHE[cache_key] = result
return result
def _compact_matchups(result: Any, relevant_names: set[str]) -> Any:
if not isinstance(result, list):
return result
compact = []
for item in result:
if not isinstance(item, dict):
continue
hero_name = item.get("hero_name", "")
if _normalize_hero_name(hero_name) not in relevant_names:
continue
compact.append(
{
"hero_name": hero_name,
"games": item.get("games"),
"win_rate": item.get("win_rate"),
}
)
return compact[:12]
async def _call_openai_for_recommendation(
request: DraftRecommendationRequest,
context: dict[str, Any],
) -> dict[str, Any]:
api_key = os.getenv("HF_TOKEN", "")
if not api_key:
raise RuntimeError("HF_TOKEN is not set")
messages = [
{
"role": "system",
"content": (
"Respond with JSON only. Your first character must be `{` and your last "
"character must be `}`. Do not include analysis, planning, markdown, "
"code fences, or explanatory text. You are a Dota 2 Captains Mode draft "
"advisor. Make a final recommendation now from the supplied draft state "
"and prefetched OpenDota evidence. Do not request tools. Do not recommend "
"a hero that has already been picked or banned. Keep rationale under 25 "
"words and stats_used to at most two short items."
),
},
{
"role": "user",
"content": "/no_think\n"
+ json.dumps(
{
"draft_request": _request_to_json(request),
"candidate_heroes": context.get("candidate_heroes", []),
"prefetched_evidence": _trim_for_prompt(context.get("observations", [])),
"prefetch_errors": context.get("prefetch_errors", []),
"response_schema": {
"recommendation": {
"hero": "Hero name",
"action": request.action,
"confidence": "low|medium|high",
"rationale": "under 25 words",
"stats_used": ["up to two brief facts from prefetched evidence"],
}
},
},
ensure_ascii=True,
),
},
]
await OPENAI_RATE_LIMITER.acquire()
if AGENT_DEBUG:
_debug_log(
"final model request",
{
"model": OPENAI_MODEL,
"observations": len(context.get("observations", [])),
"prompt_chars": sum(len(message["content"]) for message in messages),
},
)
content = await _create_chat_completion(api_key, messages)
try:
decision = _parse_json_object(content)
except JSONDecodeError:
if AGENT_DEBUG:
_debug_log("JSON parse failed; retrying final prompt", {"prefix": content[:240]})
retry_messages = [
{
**messages[0],
"content": (
"Return exactly one valid JSON object and nothing else. "
"No reasoning. No prose. Start with `{`. End with `}`. "
+ messages[0]["content"]
),
},
messages[1],
]
decision = _parse_json_object(await _create_chat_completion(api_key, retry_messages))
return decision.get("recommendation", decision)
async def _create_chat_completion(api_key: str, messages: list[dict[str, str]]) -> str:
if not OPENAI_STREAM:
return await _create_non_stream_chat_completion(api_key, messages, 0, [])
content_chunks: list[str] = []
alternate_chunks: list[str] = []
chunk_count = 0
finish_reasons: list[str] = []
client = await _get_openai_client(api_key)
stream = await client.chat.completions.create(
model=OPENAI_MODEL,
messages=messages,
temperature=0.2,
max_tokens=OPENAI_MAX_TOKENS,
stream=True,
**_model_extra_kwargs(),
)
async for chunk in stream:
chunk_count += 1
if chunk.choices:
choice = chunk.choices[0]
if choice.finish_reason:
finish_reasons.append(choice.finish_reason)
content = _extract_primary_text(choice.delta)
if content:
content_chunks.append(content)
else:
alternate = _extract_alternate_text(choice.delta)
if alternate:
alternate_chunks.append(alternate)
content = "".join(content_chunks)
if content:
return content
alternate_content = "".join(alternate_chunks)
if alternate_content:
return alternate_content
if AGENT_DEBUG:
_debug_log(
"stream response was empty; retrying non-stream",
{"chunks": chunk_count, "finish_reasons": finish_reasons},
)
return await _create_non_stream_chat_completion(
api_key,
messages,
chunk_count,
finish_reasons,
)
async def _create_non_stream_chat_completion(
api_key: str,
messages: list[dict[str, str]],
empty_stream_chunks: int,
empty_stream_finish_reasons: list[str],
) -> str:
client = await _get_openai_client(api_key)
response = await client.chat.completions.create(
model=OPENAI_MODEL,
messages=messages,
temperature=0.2,
max_tokens=OPENAI_MAX_TOKENS,
stream=False,
**_model_extra_kwargs(),
)
message = response.choices[0].message if response.choices else None
content = _extract_primary_text(message) if message else ""
if not content and message:
content = _extract_alternate_text(message)
if content:
return content
finish_reasons = [
choice.finish_reason
for choice in response.choices
if choice.finish_reason
]
raise RuntimeError(
"model returned an empty response "
f"(stream chunks={empty_stream_chunks}, "
f"stream finish_reasons={empty_stream_finish_reasons or ['unknown']}, "
f"non_stream finish_reasons={finish_reasons or ['unknown']})"
)
async def _get_openai_client(api_key: str) -> AsyncOpenAI:
async with _CACHE_LOCK:
client = _OPENAI_CLIENTS.get(api_key)
if client is None:
client = AsyncOpenAI(
base_url=HF_ROUTER_BASE_URL,
api_key=api_key,
timeout=30,
max_retries=0,
)
_OPENAI_CLIENTS[api_key] = client
return client
def _extract_primary_text(message_part: Any) -> str:
if not message_part:
return ""
for field in ("content", "text"):
value = getattr(message_part, field, None)
if isinstance(value, str) and value:
return value
if isinstance(message_part, dict):
for field in ("content", "text"):
value = message_part.get(field)
if isinstance(value, str) and value:
return value
return ""
def _extract_alternate_text(message_part: Any) -> str:
if not message_part:
return ""
for field in ("reasoning_content", "reasoning"):
value = getattr(message_part, field, None)
if isinstance(value, str) and value:
return value
if isinstance(message_part, dict):
for field in ("reasoning_content", "reasoning"):
value = message_part.get(field)
if isinstance(value, str) and value:
return value
return ""
def _model_extra_kwargs() -> dict[str, Any]:
if "qwen" not in OPENAI_MODEL.lower():
return {}
return {"extra_body": {"chat_template_kwargs": {"enable_thinking": False}}}
def _trim_for_prompt(value: Any, max_chars: int = MAX_PROMPT_VALUE_CHARS) -> Any:
if isinstance(value, str):
if len(value) <= max_chars:
return value
return value[:max_chars] + f"... [truncated {len(value) - max_chars} chars]"
if isinstance(value, list):
return [_trim_for_prompt(item, max_chars) for item in value]
if isinstance(value, dict):
return {key: _trim_for_prompt(item, max_chars) for key, item in value.items()}
return value
def _debug_log(message: str, details: dict[str, Any]) -> None:
print(f"[agent] {message}: {json.dumps(details, ensure_ascii=True)}", flush=True)
def _parse_json_object(content: str) -> dict[str, Any]:
text = content.strip()
if text.startswith("```"):
lines = text.splitlines()
if lines and lines[0].startswith("```"):
lines = lines[1:]
if lines and lines[-1].startswith("```"):
lines = lines[:-1]
text = "\n".join(lines).strip()
if not text.startswith("{"):
extracted = _extract_json_object_text(text)
if extracted:
text = extracted
return json.loads(text)
def _extract_json_object_text(text: str) -> str:
start = text.find("{")
if start < 0:
return ""
depth = 0
in_string = False
escaped = False
for index in range(start, len(text)):
char = text[index]
if escaped:
escaped = False
continue
if char == "\\":
escaped = True
continue
if char == '"':
in_string = not in_string
continue
if in_string:
continue
if char == "{":
depth += 1
elif char == "}":
depth -= 1
if depth == 0:
return text[start : index + 1]
return ""
def _decode_mcp_response(response: httpx.Response) -> dict[str, Any]:
text = response.text.strip()
if not text:
return {}
if "text/event-stream" in response.headers.get("content-type", ""):
data_lines = [
line.removeprefix("data:").strip()
for line in text.splitlines()
if line.startswith("data:")
]
text = data_lines[-1] if data_lines else "{}"
return json.loads(text)
def _normalize_tool_result(result: Any) -> Any:
if isinstance(result, dict) and "content" in result:
content = result["content"]
if isinstance(content, list):
return [
item.get("text", item)
if isinstance(item, dict)
else getattr(item, "text", str(item))
for item in content
]
return result
def _request_to_json(request: DraftRecommendationRequest) -> dict[str, Any]:
return {
"action": request.action,
"step_number": request.step_number,
"phase": request.phase,
"slot": request.slot,
"you": request.you.__dict__,
"opponent": request.opponent.__dict__,
"unavailable_heroes": request.unavailable_heroes,
}
def _format_recommendation(
request: DraftRecommendationRequest,
recommendation: dict[str, Any],
observations: list[dict[str, Any]],
) -> str:
hero = recommendation.get("hero", "Unknown hero")
action = recommendation.get("action", request.action)
confidence = recommendation.get("confidence", "unknown")
rationale = recommendation.get("rationale", "No rationale returned.")
stats = recommendation.get("stats_used") or []
stats_text = "\n".join(f"- {item}" for item in stats) if stats else "- No stat summary returned."
tool_text = ", ".join(item["tool"] for item in observations) or "none"
return (
f"Recommended {action}: **{hero}**\n\n"
f"Confidence: `{confidence}`\n\n"
f"{rationale}\n\n"
f"Stats used:\n{stats_text}\n\n"
f"OpenDota MCP tools called: {tool_text}"
)
warm_static_dota_data()