m3taversal
7bfce6b706
Pulled from /opt/teleo-eval/telegram/ on VPS. Includes: - bot.py (92K), kb_retrieval.py, kb_tools.py (agentic retrieval) - retrieval.py (RRF merge, query decomposition, entity traversal) - response.py (system prompt builder, response parser) - agent_config.py, agent_runner.py (multi-agent template unit support) - approval_stages.py, approvals.py, digest.py (approval workflow) - eval_checks.py, eval.py (response quality checks) - output_gate.py, x_publisher.py, x_client.py, x_search.py (X pipeline) - market_data.py, worktree_lock.py (utilities) - rio.yaml, theseus.yaml (agent configs) These files were deployed to VPS but never committed to the repo. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2069 lines
90 KiB
Python
2069 lines
90 KiB
Python
#!/usr/bin/env python3
|
|
"""Teleo Telegram Bot — Rio as analytical agent in community groups.
|
|
|
|
Architecture:
|
|
- Always-on ingestion: captures all messages, batch triage every N minutes
|
|
- Tag-based response: Opus-quality KB-grounded responses when @tagged
|
|
- Conversation-window triage: identifies coherent claims across message threads
|
|
- Full eval tracing: Rio's responses are logged as KB claims, accountable
|
|
|
|
Two paths (Ganymede architecture):
|
|
- Fast path (read): tag → KB query → Opus response → post to group
|
|
- Slow path (write): batch triage → archive to inbox/ → pipeline extracts
|
|
|
|
Separate systemd service: teleo-telegram.service
|
|
Does NOT integrate with pipeline daemon.
|
|
|
|
Epimetheus owns this module.
|
|
"""
|
|
|
|
import argparse
|
|
import asyncio
|
|
import logging
|
|
import os
|
|
import re
|
|
import sqlite3
|
|
import sys
|
|
import time
|
|
|
|
import yaml
|
|
from collections import defaultdict
|
|
from datetime import datetime, timezone
|
|
from pathlib import Path
|
|
|
|
# Add pipeline lib to path for shared modules
|
|
sys.path.insert(0, "/opt/teleo-eval/pipeline")
|
|
|
|
from telegram import Update
|
|
from telegram.ext import (
|
|
Application,
|
|
CommandHandler,
|
|
ContextTypes,
|
|
MessageHandler,
|
|
filters,
|
|
)
|
|
|
|
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
|
|
import json as _json
|
|
from kb_retrieval import KBIndex, retrieve_context, retrieve_vector_context
|
|
from retrieval import orchestrate_retrieval
|
|
from market_data import get_token_price, format_price_context
|
|
from worktree_lock import main_worktree_lock
|
|
from x_client import search_tweets, fetch_from_url, check_research_rate_limit, record_research_usage, get_research_remaining
|
|
|
|
# ─── Config ─────────────────────────────────────────────────────────────
|
|
|
|
BOT_TOKEN_FILE = "/opt/teleo-eval/secrets/telegram-bot-token"
|
|
OPENROUTER_KEY_FILE = "/opt/teleo-eval/secrets/openrouter-key"
|
|
PIPELINE_DB = "/opt/teleo-eval/pipeline/pipeline.db"
|
|
KB_READ_DIR = "/opt/teleo-eval/workspaces/main" # For KB retrieval (clean main branch)
|
|
ARCHIVE_DIR = "/opt/teleo-eval/telegram-archives" # Write outside worktree to avoid read-only errors
|
|
MAIN_WORKTREE = "/opt/teleo-eval/workspaces/main" # For git operations only
|
|
LEARNINGS_FILE = "/opt/teleo-eval/workspaces/main/agents/rio/learnings.md" # Agent memory (Option D)
|
|
LOG_FILE = "/opt/teleo-eval/logs/telegram-bot.log"
|
|
|
|
# Persistent audit connection — opened once at startup, reused for all writes
|
|
# (Ganymede + Rhea: no per-response sqlite3.connect / migrate)
|
|
_audit_conn: sqlite3.Connection | None = None
|
|
|
|
# Triage interval (seconds)
|
|
TRIAGE_INTERVAL = 900 # 15 minutes
|
|
|
|
# Models
|
|
RESPONSE_MODEL = "anthropic/claude-opus-4-6" # Opus for tagged responses
|
|
TRIAGE_MODEL = "anthropic/claude-haiku-4.5" # Haiku for batch triage
|
|
|
|
# KB scope — None means all domains (Rio default). Set from YAML config for other agents.
|
|
AGENT_KB_SCOPE: list[str] | None = None
|
|
|
|
# Rate limits
|
|
MAX_RESPONSE_PER_USER_PER_HOUR = 30
|
|
MIN_MESSAGE_LENGTH = 20 # Skip very short messages
|
|
|
|
# ─── Logging ────────────────────────────────────────────────────────────
|
|
|
|
logging.basicConfig(
|
|
level=logging.INFO,
|
|
format="%(asctime)s %(name)s [%(levelname)s] %(message)s",
|
|
handlers=[
|
|
logging.FileHandler(LOG_FILE),
|
|
logging.StreamHandler(),
|
|
],
|
|
)
|
|
logger = logging.getLogger("telegram-bot")
|
|
|
|
# ─── State ──────────────────────────────────────────────────────────────
|
|
|
|
# Message buffer for batch triage
|
|
message_buffer: list[dict] = []
|
|
|
|
# Rate limiting
|
|
user_response_times: dict[int, list[float]] = defaultdict(list)
|
|
|
|
# Allowed group IDs (set after first message received, or configure)
|
|
allowed_groups: set[int] = set()
|
|
|
|
# Shared KB index (built once, refreshed on mtime change)
|
|
kb_index = KBIndex(KB_READ_DIR)
|
|
|
|
# Conversation windows — track active conversations per (chat_id, user_id)
|
|
# Rhea's model: count unanswered messages, reset on bot response, expire at threshold
|
|
CONVERSATION_WINDOW = 5 # expire after 5 unanswered messages
|
|
unanswered_count: dict[tuple[int, int], int] = {} # (chat_id, user_id) → unanswered count
|
|
|
|
# Conversation history — last N exchanges for prompt context (Ganymede: high-value change)
|
|
MAX_HISTORY_USER = 5
|
|
MAX_HISTORY_CHAT = 30 # Group chats: multiple users, longer threads
|
|
conversation_history: dict[tuple[int, int], list[dict]] = {} # (chat_id, user_id) → [{user, bot}]
|
|
|
|
# Full transcript store — all messages in all chats, dumped every 6 hours
|
|
# Keyed by chat_id. No cap — dumped and cleared on schedule.
|
|
chat_transcripts: dict[int, list[dict]] = {}
|
|
TRANSCRIPT_DIR = "/opt/teleo-eval/transcripts"
|
|
|
|
|
|
# ─── Content Classification ─────────────────────────────────────────────
|
|
|
|
# Sub-topic keywords for internet-finance sources
|
|
_TOPIC_KEYWORDS = {
|
|
"futarchy": ["futarchy", "autocrat", "conditional market", "twap", "pass/fail",
|
|
"decision market", "futard", "metadao governance"],
|
|
"ownership-coins": ["ownership coin", "ico", "fundraise", "launch", "launchpad",
|
|
"permissioned", "permissionless", "unruggable", "treasury management",
|
|
"buyback", "token split"],
|
|
"defi": ["amm", "liquidity", "swap", "lending", "borrowing", "yield", "tvl",
|
|
"dex", "lp", "staking", "vault", "protocol"],
|
|
"governance": ["proposal", "vote", "governance", "dao", "subcommittee",
|
|
"treasury", "resolution", "benevolent dictator"],
|
|
"market-analysis": ["price", "market cap", "fdv", "oversubscribed", "committed",
|
|
"trading", "volume", "bullish", "bearish", "thesis"],
|
|
"crypto-infra": ["solana", "ethereum", "base", "bridge", "wallet", "on-ramp",
|
|
"off-ramp", "fiat", "stablecoin", "usdc"],
|
|
}
|
|
|
|
# Domain keywords for non-internet-finance content
|
|
_DOMAIN_KEYWORDS = {
|
|
"ai-alignment": ["ai safety", "alignment", "superintelligence", "llm", "frontier model",
|
|
"interpretability", "rlhf", "anthropic", "openai", "deepmind"],
|
|
"health": ["glp-1", "healthcare", "clinical", "pharma", "biotech", "fda",
|
|
"medicare", "hospital", "diagnosis", "therapeutic"],
|
|
"space-development": ["spacex", "starship", "orbital", "lunar", "satellite",
|
|
"launch cost", "rocket", "nasa", "artemis"],
|
|
"entertainment": ["streaming", "creator economy", "ip", "nft", "gaming",
|
|
"content", "media", "studio", "audience"],
|
|
}
|
|
|
|
|
|
# Author handle → domain map (Ganymede: counts as 1 keyword match)
|
|
_AUTHOR_DOMAIN_MAP = {
|
|
"metadaoproject": "internet-finance",
|
|
"metadaofi": "internet-finance",
|
|
"futardio": "internet-finance",
|
|
"p2pdotme": "internet-finance",
|
|
"oxranga": "internet-finance",
|
|
"metanallok": "internet-finance",
|
|
"proph3t_": "internet-finance",
|
|
"01resolved": "internet-finance",
|
|
"anthropicai": "ai-alignment",
|
|
"openai": "ai-alignment",
|
|
"daborai": "ai-alignment",
|
|
"deepmind": "ai-alignment",
|
|
"spacex": "space-development",
|
|
"blaborig": "space-development",
|
|
"nasa": "space-development",
|
|
}
|
|
|
|
|
|
def _classify_content(text: str, author: str = "") -> tuple[str, list[str]]:
|
|
"""Classify content into domain + sub-tags based on keywords + author.
|
|
|
|
Returns (domain, [sub-tags]). Default: internet-finance with no sub-tags.
|
|
"""
|
|
text_lower = text.lower()
|
|
author_lower = author.lower().lstrip("@")
|
|
|
|
# Author handle gives 1 keyword match toward domain threshold
|
|
author_domain = _AUTHOR_DOMAIN_MAP.get(author_lower, "")
|
|
|
|
# Check non-IF domains first
|
|
for domain, keywords in _DOMAIN_KEYWORDS.items():
|
|
matches = sum(1 for kw in keywords if kw in text_lower)
|
|
if author_domain == domain:
|
|
matches += 1 # Author signal counts as 1 match
|
|
if matches >= 2:
|
|
return domain, []
|
|
|
|
# Default to internet-finance, classify sub-topics
|
|
sub_tags = []
|
|
for tag, keywords in _TOPIC_KEYWORDS.items():
|
|
if any(kw in text_lower for kw in keywords):
|
|
sub_tags.append(tag)
|
|
|
|
return "internet-finance", sub_tags
|
|
|
|
|
|
# ─── Transcript Management ──────────────────────────────────────────────
|
|
|
|
|
|
def _record_transcript(msg, text: str, is_bot: bool = False,
|
|
rio_response: str = None, internal: dict = None):
|
|
"""Record a message to the full transcript for this chat."""
|
|
chat_id = msg.chat_id
|
|
transcript = chat_transcripts.setdefault(chat_id, [])
|
|
|
|
entry = {
|
|
"ts": msg.date.isoformat() if hasattr(msg, "date") and msg.date else datetime.now(timezone.utc).isoformat(),
|
|
"chat_id": chat_id,
|
|
"chat_title": msg.chat.title if hasattr(msg, "chat") and msg.chat else str(chat_id),
|
|
"message_id": msg.message_id if hasattr(msg, "message_id") else None,
|
|
}
|
|
|
|
if is_bot:
|
|
entry["type"] = "bot_response"
|
|
entry["rio_response"] = rio_response or text
|
|
if internal:
|
|
entry["internal"] = internal # KB matches, searches, learnings
|
|
else:
|
|
user = msg.from_user if hasattr(msg, "from_user") else None
|
|
entry["type"] = "user_message"
|
|
entry["username"] = f"@{user.username}" if user and user.username else "unknown"
|
|
entry["display_name"] = user.full_name if user else "unknown"
|
|
entry["user_id"] = user.id if user else None
|
|
entry["message"] = text[:2000]
|
|
entry["reply_to"] = msg.reply_to_message.message_id if hasattr(msg, "reply_to_message") and msg.reply_to_message else None
|
|
|
|
transcript.append(entry)
|
|
|
|
|
|
_last_dump_index: dict[int, int] = {} # chat_id → index of last dumped message
|
|
|
|
|
|
async def _dump_transcripts(context=None):
|
|
"""Append new transcript entries to per-chat JSONL files. Runs every hour.
|
|
|
|
Append-only: each dump writes only new messages since last dump (Ganymede review).
|
|
One JSONL file per chat per day. Each line is one message.
|
|
"""
|
|
if not chat_transcripts:
|
|
return
|
|
|
|
os.makedirs(TRANSCRIPT_DIR, exist_ok=True)
|
|
now = datetime.now(timezone.utc)
|
|
today = now.strftime("%Y-%m-%d")
|
|
|
|
import json as _json
|
|
for chat_id, entries in list(chat_transcripts.items()):
|
|
if not entries:
|
|
continue
|
|
|
|
# Only write new entries since last dump
|
|
last_idx = _last_dump_index.get(chat_id, 0)
|
|
new_entries = entries[last_idx:]
|
|
if not new_entries:
|
|
continue
|
|
|
|
# Get chat title from first entry
|
|
chat_title = entries[0].get("chat_title", str(chat_id))
|
|
chat_slug = re.sub(r"[^a-z0-9]+", "-", chat_title.lower()).strip("-") or str(chat_id)
|
|
|
|
# Create per-chat directory
|
|
chat_dir = os.path.join(TRANSCRIPT_DIR, chat_slug)
|
|
os.makedirs(chat_dir, exist_ok=True)
|
|
|
|
# Append to today's JSONL file
|
|
filename = f"{today}.jsonl"
|
|
filepath = os.path.join(chat_dir, filename)
|
|
|
|
try:
|
|
with open(filepath, "a") as f:
|
|
for entry in new_entries:
|
|
f.write(_json.dumps(entry, default=str) + "\n")
|
|
_last_dump_index[chat_id] = len(entries)
|
|
logger.info("Transcript appended: %s (+%d messages, %d total)",
|
|
filepath, len(new_entries), len(entries))
|
|
except Exception as e:
|
|
logger.warning("Failed to dump transcript for %s: %s", chat_slug, e)
|
|
|
|
|
|
def _create_inline_source(source_text: str, user_message: str, user, msg):
|
|
"""Create a source file from Rio's SOURCE: tag. Verbatim user content, attributed."""
|
|
try:
|
|
username = user.username if user else "anonymous"
|
|
date_str = datetime.now(timezone.utc).strftime("%Y-%m-%d")
|
|
slug = re.sub(r"[^a-z0-9]+", "-", source_text[:50].lower()).strip("-")
|
|
filename = f"{date_str}-tg-source-{username}-{slug}.md"
|
|
source_path = Path(ARCHIVE_DIR) / filename
|
|
if source_path.exists():
|
|
return
|
|
|
|
content = f"""---
|
|
type: source
|
|
source_type: telegram-contribution
|
|
title: "Source from @{username} — {source_text[:80]}"
|
|
author: "@{username}"
|
|
date: {date_str}
|
|
domain: {_classify_content(source_text + " " + user_message)[0]}
|
|
format: contribution
|
|
status: unprocessed
|
|
proposed_by: "@{username}"
|
|
contribution_type: source-submission
|
|
tags: {["telegram-contribution", "inline-source"] + _classify_content(source_text + " " + user_message)[1]}
|
|
---
|
|
|
|
# Source: {source_text[:100]}
|
|
|
|
Contributed by @{username} in Telegram chat.
|
|
Flagged by Rio as relevant source material.
|
|
|
|
## Verbatim User Message
|
|
|
|
{user_message}
|
|
|
|
## Rio's Context
|
|
|
|
{source_text}
|
|
"""
|
|
source_path.write_text(content)
|
|
logger.info("Inline source created: %s (by @%s)", filename, username)
|
|
except Exception as e:
|
|
logger.warning("Failed to create inline source: %s", e)
|
|
|
|
|
|
def _create_inline_claim(claim_text: str, user_message: str, user, msg):
|
|
"""Create a draft claim file from Rio's CLAIM: tag. Attributed to contributor."""
|
|
try:
|
|
username = user.username if user else "anonymous"
|
|
date_str = datetime.now(timezone.utc).strftime("%Y-%m-%d")
|
|
slug = re.sub(r"[^a-z0-9]+", "-", claim_text[:60].lower()).strip("-")
|
|
filename = f"{date_str}-tg-claim-{username}-{slug}.md"
|
|
source_path = Path(ARCHIVE_DIR) / filename
|
|
if source_path.exists():
|
|
return
|
|
|
|
domain, sub_tags = _classify_content(claim_text + " " + user_message)
|
|
|
|
content = f"""---
|
|
type: source
|
|
source_type: telegram-claim
|
|
title: "Claim from @{username} — {claim_text[:80]}"
|
|
author: "@{username}"
|
|
date: {date_str}
|
|
domain: {domain}
|
|
format: claim-draft
|
|
status: unprocessed
|
|
proposed_by: "@{username}"
|
|
contribution_type: claim-proposal
|
|
tags: [telegram-claim, inline-claim]
|
|
---
|
|
|
|
# Draft Claim: {claim_text}
|
|
|
|
Contributed by @{username} in Telegram chat.
|
|
Flagged by Rio as a specific, disagreeable assertion worth extracting.
|
|
|
|
## Verbatim User Message
|
|
|
|
{user_message}
|
|
|
|
## Proposed Claim
|
|
|
|
{claim_text}
|
|
"""
|
|
source_path.write_text(content)
|
|
logger.info("Inline claim drafted: %s (by @%s)", filename, username)
|
|
except Exception as e:
|
|
logger.warning("Failed to create inline claim: %s", e)
|
|
|
|
|
|
# ─── Helpers ────────────────────────────────────────────────────────────
|
|
|
|
|
|
|
|
def get_db_stats() -> dict:
|
|
"""Get basic KB stats from pipeline DB."""
|
|
try:
|
|
conn = sqlite3.connect(PIPELINE_DB, timeout=5)
|
|
conn.row_factory = sqlite3.Row
|
|
conn.execute("PRAGMA query_only=ON")
|
|
merged = conn.execute("SELECT COUNT(*) as n FROM prs WHERE status='merged'").fetchone()["n"]
|
|
contributors = conn.execute("SELECT COUNT(*) as n FROM contributors").fetchone()["n"]
|
|
conn.close()
|
|
return {"merged_claims": merged, "contributors": contributors}
|
|
except Exception:
|
|
return {"merged_claims": "?", "contributors": "?"}
|
|
|
|
|
|
from eval_checks import (
|
|
_LLMResponse, estimate_cost, check_url_fabrication, apply_confidence_floor,
|
|
CONFIDENCE_FLOOR, COST_ALERT_THRESHOLD,
|
|
)
|
|
|
|
|
|
async def call_openrouter(model: str, prompt: str, max_tokens: int = 2048) -> _LLMResponse | None:
|
|
"""Call OpenRouter API. Returns _LLMResponse with token counts and cost."""
|
|
import aiohttp
|
|
|
|
key = Path(OPENROUTER_KEY_FILE).read_text().strip()
|
|
payload = {
|
|
"model": model,
|
|
"messages": [{"role": "user", "content": prompt}],
|
|
"max_tokens": max_tokens,
|
|
"temperature": 0.3,
|
|
}
|
|
try:
|
|
async with aiohttp.ClientSession() as session:
|
|
async with session.post(
|
|
"https://openrouter.ai/api/v1/chat/completions",
|
|
headers={"Authorization": f"Bearer {key}", "Content-Type": "application/json"},
|
|
json=payload,
|
|
timeout=aiohttp.ClientTimeout(total=120),
|
|
) as resp:
|
|
if resp.status >= 400:
|
|
logger.error("OpenRouter %s → %d", model, resp.status)
|
|
return None
|
|
data = await resp.json()
|
|
content = data.get("choices", [{}])[0].get("message", {}).get("content")
|
|
if content is None:
|
|
return None
|
|
# Extract token usage from OpenRouter response
|
|
usage = data.get("usage", {})
|
|
pt = usage.get("prompt_tokens", 0)
|
|
ct = usage.get("completion_tokens", 0)
|
|
cost = estimate_cost(model, pt, ct)
|
|
return _LLMResponse(content, prompt_tokens=pt, completion_tokens=ct,
|
|
cost=cost, model=model)
|
|
except Exception as e:
|
|
logger.error("OpenRouter error: %s", e)
|
|
return None
|
|
|
|
|
|
async def call_openrouter_with_tools(model: str, prompt: str, tools: list[dict],
|
|
tool_executor, max_tokens: int = 2048,
|
|
max_iterations: int = 3) -> tuple[_LLMResponse | None, list[dict]]:
|
|
"""Agentic loop: call LLM with tools, execute tool calls, feed back results.
|
|
|
|
Returns (final_response, tool_call_audit_list).
|
|
Token counts and cost are ACCUMULATED across all iterations, not just the final call.
|
|
Tool audit includes LLM reasoning text between tool calls for full observability.
|
|
Falls back to plain call_openrouter if model returns 400 with tool errors.
|
|
"""
|
|
import aiohttp
|
|
import json
|
|
|
|
key = Path(OPENROUTER_KEY_FILE).read_text().strip()
|
|
messages = [{"role": "user", "content": prompt}]
|
|
tool_audit = []
|
|
|
|
# Accumulate tokens/cost across ALL iterations (not just final call)
|
|
total_prompt_tokens = 0
|
|
total_completion_tokens = 0
|
|
total_cost = 0.0
|
|
|
|
for iteration in range(max_iterations):
|
|
payload = {
|
|
"model": model,
|
|
"messages": messages,
|
|
"max_tokens": max_tokens,
|
|
"temperature": 0.3,
|
|
"tools": tools,
|
|
}
|
|
try:
|
|
async with aiohttp.ClientSession() as session:
|
|
async with session.post(
|
|
"https://openrouter.ai/api/v1/chat/completions",
|
|
headers={"Authorization": f"Bearer {key}", "Content-Type": "application/json"},
|
|
json=payload,
|
|
timeout=aiohttp.ClientTimeout(total=120),
|
|
) as resp:
|
|
if resp.status >= 400:
|
|
body = await resp.text()
|
|
if "tool" in body.lower():
|
|
logger.warning("Model doesn't support tools, falling back to plain call")
|
|
result = await call_openrouter(model, prompt, max_tokens)
|
|
return result, tool_audit
|
|
logger.error("OpenRouter with tools %s → %d", model, resp.status)
|
|
return None, tool_audit
|
|
data = await resp.json()
|
|
except Exception as e:
|
|
logger.error("OpenRouter with tools error: %s", e)
|
|
return None, tool_audit
|
|
|
|
# Accumulate this iteration's token usage
|
|
usage = data.get("usage", {})
|
|
iter_pt = usage.get("prompt_tokens", 0)
|
|
iter_ct = usage.get("completion_tokens", 0)
|
|
iter_cost = estimate_cost(model, iter_pt, iter_ct)
|
|
total_prompt_tokens += iter_pt
|
|
total_completion_tokens += iter_ct
|
|
total_cost += iter_cost
|
|
|
|
choice = data.get("choices", [{}])[0]
|
|
message = choice.get("message", {})
|
|
|
|
# If model wants to call tools (check presence only — finish_reason varies by model)
|
|
tool_calls_in_response = message.get("tool_calls", [])
|
|
if tool_calls_in_response:
|
|
# Capture LLM reasoning text alongside tool calls (the "thinking" between searches)
|
|
reasoning_text = message.get("content", "")
|
|
if reasoning_text:
|
|
tool_audit.append({
|
|
"type": "reasoning", "iteration": iteration + 1,
|
|
"text": reasoning_text[:2000],
|
|
"tokens": {"prompt": iter_pt, "completion": iter_ct, "cost": round(iter_cost, 6)},
|
|
})
|
|
|
|
messages.append(message) # Add assistant message with tool calls
|
|
for tc in tool_calls_in_response:
|
|
fn_name = tc["function"]["name"]
|
|
try:
|
|
fn_args = json.loads(tc["function"]["arguments"])
|
|
except (json.JSONDecodeError, KeyError):
|
|
fn_args = {}
|
|
|
|
t0 = time.monotonic()
|
|
result = tool_executor(fn_name, fn_args)
|
|
duration_ms = int((time.monotonic() - t0) * 1000)
|
|
|
|
# Truncate tool results
|
|
result_str = str(result)[:4000]
|
|
tool_audit.append({
|
|
"type": "tool_call", "iteration": iteration + 1,
|
|
"tool": fn_name, "input": fn_args,
|
|
"output_preview": result_str[:500],
|
|
"output_length": len(result_str), "duration_ms": duration_ms,
|
|
})
|
|
messages.append({
|
|
"role": "tool",
|
|
"tool_call_id": tc["id"],
|
|
"content": result_str,
|
|
})
|
|
continue # Next iteration with tool results
|
|
|
|
# Model returned a text response (done)
|
|
content = message.get("content")
|
|
if content is None:
|
|
return None, tool_audit
|
|
return _LLMResponse(content, prompt_tokens=total_prompt_tokens,
|
|
completion_tokens=total_completion_tokens,
|
|
cost=total_cost, model=model), tool_audit
|
|
|
|
# Exhausted iterations — force one final call WITHOUT tools to get a text answer
|
|
logger.warning("Tool loop exhausted %d iterations, forcing final plain call", max_iterations)
|
|
try:
|
|
messages.append({"role": "user", "content": "Please provide your final answer now based on the information gathered."})
|
|
payload_final = {
|
|
"model": model,
|
|
"messages": messages,
|
|
"max_tokens": max_tokens,
|
|
"temperature": 0.3,
|
|
}
|
|
async with aiohttp.ClientSession() as session:
|
|
async with session.post(
|
|
"https://openrouter.ai/api/v1/chat/completions",
|
|
headers={"Authorization": f"Bearer {key}", "Content-Type": "application/json"},
|
|
json=payload_final,
|
|
timeout=aiohttp.ClientTimeout(total=120),
|
|
) as resp:
|
|
if resp.status < 400:
|
|
data = await resp.json()
|
|
content = data.get("choices", [{}])[0].get("message", {}).get("content")
|
|
if content:
|
|
usage = data.get("usage", {})
|
|
total_prompt_tokens += usage.get("prompt_tokens", 0)
|
|
total_completion_tokens += usage.get("completion_tokens", 0)
|
|
total_cost += estimate_cost(model, usage.get("prompt_tokens", 0),
|
|
usage.get("completion_tokens", 0))
|
|
return _LLMResponse(content, prompt_tokens=total_prompt_tokens,
|
|
completion_tokens=total_completion_tokens,
|
|
cost=total_cost, model=model), tool_audit
|
|
except Exception as e:
|
|
logger.error("Final plain call after tool exhaustion failed: %s", e)
|
|
return None, tool_audit
|
|
|
|
|
|
def is_rate_limited(user_id: int) -> bool:
|
|
"""Check if a user has exceeded the response rate limit."""
|
|
now = time.time()
|
|
times = user_response_times[user_id]
|
|
# Prune old entries
|
|
times[:] = [t for t in times if now - t < 3600]
|
|
return len(times) >= MAX_RESPONSE_PER_USER_PER_HOUR
|
|
|
|
|
|
def sanitize_message(text: str) -> str:
|
|
"""Sanitize message content before sending to LLM. (Ganymede: security)"""
|
|
# Strip code blocks (potential prompt injection)
|
|
text = re.sub(r"```.*?```", "[code block removed]", text, flags=re.DOTALL)
|
|
# Strip anything that looks like system instructions
|
|
text = re.sub(r"(system:|assistant:|human:|<\|.*?\|>)", "", text, flags=re.IGNORECASE)
|
|
# Truncate
|
|
return text[:2000]
|
|
|
|
|
|
def _git_commit_archive(archive_path, filename: str):
|
|
"""Commit archived source to git so it survives git clean. (Rio review: data loss bug)"""
|
|
import subprocess
|
|
try:
|
|
cwd = MAIN_WORKTREE
|
|
subprocess.run(["git", "add", str(archive_path)], cwd=cwd, timeout=10,
|
|
capture_output=True, check=False)
|
|
result = subprocess.run(
|
|
["git", "commit", "-m", f"telegram: archive {filename}\n\n"
|
|
"Pentagon-Agent: Epimetheus <3D35839A-7722-4740-B93D-51157F7D5E70>"],
|
|
cwd=cwd, timeout=10, capture_output=True, check=False,
|
|
)
|
|
if result.returncode == 0:
|
|
# Push with retry (Ganymede: abort rebase on failure, don't lose the file)
|
|
for attempt in range(3):
|
|
rebase = subprocess.run(["git", "pull", "--rebase", "origin", "main"],
|
|
cwd=cwd, timeout=30, capture_output=True, check=False)
|
|
if rebase.returncode != 0:
|
|
subprocess.run(["git", "rebase", "--abort"], cwd=cwd, timeout=10,
|
|
capture_output=True, check=False)
|
|
logger.warning("Git rebase failed for archive %s (attempt %d), aborted", filename, attempt + 1)
|
|
continue
|
|
push = subprocess.run(["git", "push", "origin", "main"],
|
|
cwd=cwd, timeout=30, capture_output=True, check=False)
|
|
if push.returncode == 0:
|
|
logger.info("Git committed archive: %s", filename)
|
|
return
|
|
# All retries failed — file is still on filesystem (safety net), commit is uncommitted
|
|
logger.warning("Git push failed for archive %s after 3 attempts (file preserved on disk)", filename)
|
|
except Exception as e:
|
|
logger.warning("Git commit archive failed: %s", e)
|
|
|
|
|
|
def _load_learnings() -> str:
|
|
"""Load Rio's learnings file for prompt injection. Sanitized (Ganymede: prompt injection risk).
|
|
|
|
Dated entries older than 7 days are filtered out (Ganymede: stale learning TTL).
|
|
Permanent entries (undated) always included.
|
|
"""
|
|
try:
|
|
raw = Path(LEARNINGS_FILE).read_text()[:4000]
|
|
today = datetime.now(timezone.utc).date()
|
|
lines = []
|
|
for line in raw.split("\n"):
|
|
# Check for dated entries [YYYY-MM-DD]
|
|
date_match = re.search(r"\[(\d{4}-\d{2}-\d{2})\]", line)
|
|
if date_match:
|
|
try:
|
|
entry_date = datetime.strptime(date_match.group(1), "%Y-%m-%d").date()
|
|
if (today - entry_date).days > 7:
|
|
continue # stale, skip
|
|
except ValueError:
|
|
pass
|
|
lines.append(line)
|
|
return sanitize_message("\n".join(lines))
|
|
except Exception:
|
|
return ""
|
|
|
|
|
|
def _save_learning(correction: str, category: str = "factual"):
|
|
"""Append a learning to staging file. Cron syncs to git (same as archives).
|
|
|
|
Categories: communication, factual, structured_data
|
|
"""
|
|
try:
|
|
# Write to staging file outside worktree (avoids read-only errors)
|
|
staging_file = Path(ARCHIVE_DIR) / "pending-learnings.jsonl"
|
|
import json as _json
|
|
entry = _json.dumps({"category": category, "correction": correction,
|
|
"ts": datetime.now(timezone.utc).isoformat()})
|
|
with open(staging_file, "a") as f:
|
|
f.write(entry + "\n")
|
|
logger.info("Learning staged: [%s] %s", category, correction[:80])
|
|
return
|
|
except Exception as e:
|
|
logger.warning("Learning staging failed: %s", e)
|
|
|
|
# No fallback — staging is the only write path. Cron syncs to git.
|
|
|
|
|
|
def _compress_history(history: list[dict]) -> str:
|
|
"""Extract key context from conversation history — 20 tokens, unmissable (Ganymede)."""
|
|
if not history:
|
|
return ""
|
|
# Combine all text for entity/number extraction
|
|
all_text = " ".join(h.get("user", "") + " " + h.get("bot", "") for h in history)
|
|
tickers = sorted(set(re.findall(r"\$[A-Z]{2,10}", all_text)))
|
|
numbers = re.findall(r"\$[\d,.]+[KMB]?|\d+\.?\d*%", all_text)
|
|
parts = []
|
|
if tickers:
|
|
parts.append(f"Discussing: {', '.join(tickers)}")
|
|
if numbers:
|
|
parts.append(f"Key figures: {', '.join(numbers[:5])}")
|
|
parts.append(f"Exchanges: {len(history)}")
|
|
return " | ".join(parts)
|
|
|
|
|
|
def _format_conversation_history(chat_id: int, user_id: int) -> str:
|
|
"""Format conversation history with compressed context summary (Ganymede: Option C+A).
|
|
|
|
In group chats, merges user-specific history with chat-level history
|
|
so the bot sees exchanges from other users in the same chat.
|
|
"""
|
|
user_key = (chat_id, user_id)
|
|
chat_key = (chat_id, 0) # chat-level history (all users)
|
|
|
|
# Merge: chat-level history gives full group context
|
|
chat_history = conversation_history.get(chat_key, [])
|
|
user_history = conversation_history.get(user_key, [])
|
|
|
|
# Use chat-level if available (group chats), otherwise user-level (DMs)
|
|
history = chat_history if chat_history else user_history
|
|
if not history:
|
|
return "(No prior conversation)"
|
|
|
|
# Compressed context first — hard for the model to miss
|
|
summary = _compress_history(history)
|
|
lines = [summary, ""]
|
|
|
|
# Full exchange log for reference
|
|
for exchange in history:
|
|
who = exchange.get("username", "User")
|
|
if exchange.get("user"):
|
|
lines.append(f"@{who}: {exchange['user']}")
|
|
if exchange.get("bot"):
|
|
lines.append(f"Rio: {exchange['bot']}")
|
|
lines.append("")
|
|
return "\n".join(lines)
|
|
|
|
|
|
# Research intent patterns (Rhea: explicit /research + natural language fallback)
|
|
# Telegram appends @botname to commands in groups (Ganymede: /research@FutAIrdBot query)
|
|
RESEARCH_PATTERN = re.compile(r'/research(?:@\w+)?\s+(.+)', re.IGNORECASE)
|
|
|
|
|
|
async def _research_and_followup(msg, query: str, user):
|
|
"""Run X search and send a follow-up message with findings.
|
|
|
|
Used when Opus triggers RESEARCH: tag — the user expects results back,
|
|
not silent archival.
|
|
"""
|
|
from x_client import search_tweets as _search
|
|
logger.info("Research follow-up: searching X for '%s'", query)
|
|
tweets = await _search(query, max_results=10, min_engagement=0)
|
|
if not tweets:
|
|
await msg.reply_text(f"Searched X for '{query}' — nothing recent found.")
|
|
return
|
|
|
|
# Build concise summary of findings
|
|
lines = [f"Found {len(tweets)} recent posts about '{query}':\n"]
|
|
for t in tweets[:5]:
|
|
author = t.get("author", "?")
|
|
text = t.get("text", "")[:200]
|
|
url = t.get("url", "")
|
|
lines.append(f"@{author}: {text}")
|
|
if url:
|
|
lines.append(f" {url}")
|
|
lines.append("")
|
|
|
|
followup = "\n".join(lines)
|
|
# Split if needed
|
|
if len(followup) <= 4096:
|
|
await msg.reply_text(followup)
|
|
else:
|
|
chunks = []
|
|
remaining = followup
|
|
while remaining:
|
|
if len(remaining) <= 4096:
|
|
chunks.append(remaining)
|
|
break
|
|
split_at = remaining.rfind("\n\n", 0, 4000)
|
|
if split_at == -1:
|
|
split_at = remaining.rfind("\n", 0, 4096)
|
|
if split_at == -1:
|
|
split_at = 4096
|
|
chunks.append(remaining[:split_at])
|
|
remaining = remaining[split_at:].lstrip("\n")
|
|
for chunk in chunks:
|
|
if chunk.strip():
|
|
await msg.reply_text(chunk)
|
|
|
|
# Also archive for pipeline
|
|
await handle_research(msg, query, user, silent=True)
|
|
|
|
|
|
async def handle_research(msg, query: str, user, silent: bool = False):
|
|
"""Handle a research request — search X and archive results as sources.
|
|
|
|
If silent=True, archive only — no messages posted. Used when triggered
|
|
by RESEARCH: tag after Opus already responded.
|
|
"""
|
|
username = user.username if user else "unknown"
|
|
|
|
if not silent and not check_research_rate_limit(user.id if user else 0):
|
|
remaining = get_research_remaining(user.id if user else 0)
|
|
await msg.reply_text(f"Research limit reached (3/day). Resets at midnight UTC. {remaining} remaining.")
|
|
return
|
|
|
|
if not silent:
|
|
await msg.chat.send_action("typing")
|
|
|
|
logger.info("Research: searching X for '%s'", query)
|
|
tweets = await search_tweets(query, max_results=15, min_engagement=0)
|
|
logger.info("Research: got %d tweets for '%s'", len(tweets), query)
|
|
if not tweets:
|
|
if not silent:
|
|
await msg.reply_text(f"No recent tweets found for '{query}'.")
|
|
return
|
|
|
|
# Fetch full content for top tweets (not just search snippets)
|
|
from x_client import fetch_from_url
|
|
for i, tweet in enumerate(tweets[:5]): # Top 5 by engagement
|
|
if i > 0:
|
|
await asyncio.sleep(0.5) # Ganymede: 500ms between calls, polite to Ben's API
|
|
url = tweet.get("url", "")
|
|
if url:
|
|
try:
|
|
full_data = await fetch_from_url(url)
|
|
if full_data:
|
|
# Replace snippet with full text
|
|
full_text = full_data.get("text", "")
|
|
if full_text and len(full_text) > len(tweet.get("text", "")):
|
|
tweet["text"] = full_text
|
|
# Include article content if available
|
|
contents = full_data.get("contents", [])
|
|
if contents:
|
|
article_parts = []
|
|
for block in contents:
|
|
block_text = block.get("text", "")
|
|
if not block_text:
|
|
continue
|
|
block_type = block.get("type", "unstyled")
|
|
if block_type in ("header-one", "header-two", "header-three"):
|
|
article_parts.append(f"\n## {block_text}\n")
|
|
elif block_type == "blockquote":
|
|
article_parts.append(f"> {block_text}")
|
|
elif block_type == "list-item":
|
|
article_parts.append(f"- {block_text}")
|
|
else:
|
|
article_parts.append(block_text)
|
|
if article_parts:
|
|
tweet["text"] += "\n\n--- Article Content ---\n" + "\n".join(article_parts)
|
|
except Exception as e:
|
|
logger.warning("Failed to fetch full content for %s: %s", url, e)
|
|
|
|
# Archive all tweets as ONE source file per research query
|
|
# (not per-tweet — one extraction PR produces claims from the best material)
|
|
try:
|
|
# Write to staging dir (outside worktree — no read-only errors)
|
|
date_str = datetime.now(timezone.utc).strftime("%Y-%m-%d")
|
|
slug = re.sub(r"[^a-z0-9]+", "-", query[:60].lower()).strip("-")
|
|
filename = f"{date_str}-x-research-{slug}.md"
|
|
source_path = Path(ARCHIVE_DIR) / filename
|
|
source_path.parent.mkdir(parents=True, exist_ok=True)
|
|
|
|
# Build consolidated source file
|
|
tweets_body = ""
|
|
for i, tweet in enumerate(tweets, 1):
|
|
tweets_body += f"\n### Tweet {i} — @{tweet['author']} ({tweet.get('engagement', 0)} engagement)\n"
|
|
tweets_body += f"**URL:** {tweet.get('url', '')}\n"
|
|
tweets_body += f"**Followers:** {tweet.get('author_followers', 0)} | "
|
|
tweets_body += f"**Likes:** {tweet.get('likes', 0)} | **RT:** {tweet.get('retweets', 0)}\n\n"
|
|
tweets_body += f"{tweet['text']}\n"
|
|
|
|
source_content = f"""---
|
|
type: source
|
|
source_type: x-research
|
|
title: "X research: {query}"
|
|
url: ""
|
|
author: "multiple"
|
|
date: {date_str}
|
|
domain: internet-finance
|
|
format: social-media-collection
|
|
status: unprocessed
|
|
proposed_by: "@{username}"
|
|
contribution_type: research-direction
|
|
research_query: "{query.replace('"', "'")}"
|
|
tweet_count: {len(tweets)}
|
|
tags: [x-research, telegram-research]
|
|
---
|
|
|
|
# X Research: {query}
|
|
|
|
Submitted by @{username} via Telegram /research command.
|
|
{len(tweets)} tweets found, sorted by engagement.
|
|
|
|
{tweets_body}
|
|
"""
|
|
source_path.write_text(source_content)
|
|
archived = len(tweets)
|
|
logger.info("Research archived: %s (%d tweets)", filename, archived)
|
|
except Exception as e:
|
|
logger.warning("Research archive failed: %s", e)
|
|
|
|
if not silent:
|
|
record_research_usage(user.id if user else 0)
|
|
remaining = get_research_remaining(user.id if user else 0)
|
|
top_authors = list(set(t["author"] for t in tweets[:5]))
|
|
await msg.reply_text(
|
|
f"Queued {archived} tweets about '{query}' for extraction. "
|
|
f"Top voices: @{', @'.join(top_authors[:3])}. "
|
|
f"Results will appear in the KB within ~30 minutes. "
|
|
f"({remaining} research requests remaining today.)"
|
|
)
|
|
logger.info("Research: @%s queried '%s', archived %d tweets (silent=%s)", username, query, archived, silent)
|
|
|
|
|
|
# ─── Message Handlers ───────────────────────────────────────────────────
|
|
|
|
|
|
def _is_reply_to_bot(update: Update, context: ContextTypes.DEFAULT_TYPE) -> bool:
|
|
"""Check if a message is a reply to one of the bot's own messages."""
|
|
msg = update.message
|
|
if not msg or not msg.reply_to_message:
|
|
return False
|
|
replied = msg.reply_to_message
|
|
return replied.from_user is not None and replied.from_user.id == context.bot.id
|
|
|
|
|
|
async def handle_reply_to_bot(update: Update, context: ContextTypes.DEFAULT_TYPE):
|
|
"""Handle replies to the bot's messages — treat as tagged conversation."""
|
|
if not _is_reply_to_bot(update, context):
|
|
# Not a reply to us — fall through to buffer handler
|
|
await handle_message(update, context)
|
|
return
|
|
logger.info("Reply to bot from @%s",
|
|
update.message.from_user.username if update.message.from_user else "unknown")
|
|
await handle_tagged(update, context)
|
|
|
|
|
|
async def handle_message(update: Update, context: ContextTypes.DEFAULT_TYPE):
|
|
"""Handle ALL incoming group messages — buffer for triage."""
|
|
if not update.message or not update.message.text:
|
|
return
|
|
|
|
msg = update.message
|
|
text = msg.text.strip()
|
|
|
|
# Skip very short messages
|
|
if len(text) < MIN_MESSAGE_LENGTH:
|
|
return
|
|
|
|
# Conversation window behavior depends on chat type (Rio: DMs vs groups)
|
|
# DMs: auto-respond (always 1-on-1, no false positives)
|
|
# Groups: silent context only (reply-to is the only follow-up trigger)
|
|
user = msg.from_user
|
|
is_dm = msg.chat.type == "private"
|
|
|
|
if user:
|
|
key = (msg.chat_id, user.id)
|
|
if key in unanswered_count:
|
|
unanswered_count[key] += 1
|
|
|
|
if is_dm and unanswered_count[key] < CONVERSATION_WINDOW:
|
|
# DM: auto-respond — conversation window fires
|
|
logger.info("DM conversation window: @%s msg %d/%d",
|
|
user.username or "?", unanswered_count[key], CONVERSATION_WINDOW)
|
|
await handle_tagged(update, context)
|
|
return
|
|
# Group: don't track silent messages in history (Ganymede: Option A)
|
|
# History should be the actual conversation, not a log of everything said in the group
|
|
# Expire window after CONVERSATION_WINDOW unanswered messages
|
|
if unanswered_count[key] >= CONVERSATION_WINDOW:
|
|
del unanswered_count[key]
|
|
conversation_history.pop(key, None)
|
|
logger.info("Conversation window expired for @%s", user.username or "?")
|
|
|
|
# Capture to full transcript (all messages, all chats)
|
|
_record_transcript(msg, text, is_bot=False)
|
|
|
|
# Buffer for batch triage
|
|
message_buffer.append({
|
|
"text": sanitize_message(text),
|
|
"user_id": msg.from_user.id if msg.from_user else None,
|
|
"username": msg.from_user.username if msg.from_user else None,
|
|
"display_name": msg.from_user.full_name if msg.from_user else None,
|
|
"chat_id": msg.chat_id,
|
|
"message_id": msg.message_id,
|
|
"timestamp": msg.date.isoformat() if msg.date else datetime.now(timezone.utc).isoformat(),
|
|
"reply_to": msg.reply_to_message.message_id if msg.reply_to_message else None,
|
|
})
|
|
|
|
|
|
async def handle_tagged(update: Update, context: ContextTypes.DEFAULT_TYPE):
|
|
"""Handle messages that tag the bot — Rio responds with Opus."""
|
|
if not update.message or not update.message.text:
|
|
return
|
|
|
|
msg = update.message
|
|
user = msg.from_user
|
|
text = sanitize_message(msg.text)
|
|
|
|
# Rate limit check
|
|
if user and is_rate_limited(user.id):
|
|
await msg.reply_text("I'm processing other requests — try again in a few minutes.", do_quote=True)
|
|
return
|
|
|
|
logger.info("Tagged by @%s: %s", user.username if user else "unknown", text[:100])
|
|
|
|
# ─── Audit: init timing and tool call tracking ──────────────────
|
|
response_start = time.monotonic()
|
|
tool_calls = []
|
|
|
|
# Check for /research command — run search BEFORE Opus so results are in context
|
|
research_context = ""
|
|
research_match = RESEARCH_PATTERN.search(text)
|
|
if research_match:
|
|
query = research_match.group(1).strip()
|
|
logger.info("Research: searching X for '%s'", query)
|
|
from x_client import search_tweets, check_research_rate_limit, record_research_usage
|
|
if check_research_rate_limit(user.id if user else 0):
|
|
tweets = await search_tweets(query, max_results=10, min_engagement=0)
|
|
logger.info("Research: got %d tweets for '%s'", len(tweets), query)
|
|
if tweets:
|
|
# Archive as source file (staging dir)
|
|
try:
|
|
slug = re.sub(r"[^a-z0-9]+", "-", query[:60].lower()).strip("-")
|
|
filename = f"{datetime.now(timezone.utc).strftime('%Y-%m-%d')}-x-research-{slug}.md"
|
|
source_path = Path(ARCHIVE_DIR) / filename
|
|
tweets_body = "\n".join(
|
|
f"@{t['author']} ({t.get('engagement',0)} eng): {t['text'][:200]}"
|
|
for t in tweets[:10]
|
|
)
|
|
source_path.write_text(f"---\ntype: source\nsource_type: x-research\ntitle: \"X research: {query}\"\ndate: {datetime.now(timezone.utc).strftime('%Y-%m-%d')}\ndomain: internet-finance\nstatus: unprocessed\nproposed_by: \"@{user.username if user else 'unknown'}\"\ncontribution_type: research-direction\n---\n\n{tweets_body}\n")
|
|
logger.info("Research archived: %s", filename)
|
|
except Exception as e:
|
|
logger.warning("Research archive failed: %s", e)
|
|
|
|
# Build context for Opus prompt
|
|
research_context = f"\n## Fresh X Research Results for '{query}'\n"
|
|
for t in tweets[:7]:
|
|
research_context += f"- @{t['author']}: {t['text'][:150]}\n"
|
|
record_research_usage(user.id if user else 0)
|
|
# Strip the /research command from text so Opus responds to the topic, not the command
|
|
text = re.sub(r'/research(?:@\w+)?\s+', '', text).strip()
|
|
if not text:
|
|
text = query
|
|
|
|
# Send typing indicator
|
|
await msg.chat.send_action("typing")
|
|
|
|
# Fetch any X/Twitter links in the message (tweet or article)
|
|
x_link_context = ""
|
|
x_urls = re.findall(r'https?://(?:twitter\.com|x\.com)/\w+/status/\d+', text)
|
|
if x_urls:
|
|
from x_client import fetch_from_url
|
|
for url in x_urls[:3]: # Cap at 3 links
|
|
try:
|
|
tweet_data = await fetch_from_url(url)
|
|
if tweet_data:
|
|
x_link_context += f"\n## Linked Tweet by @{tweet_data['author']}\n"
|
|
if tweet_data.get("title"):
|
|
x_link_context += f"Title: {tweet_data['title']}\n"
|
|
x_link_context += f"{tweet_data['text'][:500]}\n"
|
|
x_link_context += f"Engagement: {tweet_data.get('engagement', 0)} | URL: {url}\n"
|
|
logger.info("Fetched X link: @%s — %s", tweet_data['author'], tweet_data['text'][:60])
|
|
except Exception as e:
|
|
logger.warning("Failed to fetch X link %s: %s", url, e)
|
|
|
|
# Haiku pre-pass: does this message need an X search? (Option A: two-pass)
|
|
t_haiku = time.monotonic()
|
|
if not research_context: # Skip if /research already ran
|
|
try:
|
|
haiku_prompt = (
|
|
f"Does this Telegram message need a live X/Twitter search to answer well? "
|
|
f"Only say YES if the user is asking about recent sentiment, community takes, "
|
|
f"what people are saying, or emerging discussions.\n\n"
|
|
f"Message: {text}\n\n"
|
|
f"If YES, provide a SHORT search query (2-3 words max, like 'P2P.me' or 'MetaDAO buyback'). "
|
|
f"Twitter search works best with simple queries — too many words returns nothing.\n\n"
|
|
f"Respond with ONLY one of:\n"
|
|
f"YES: [2-3 word query]\n"
|
|
f"NO"
|
|
)
|
|
haiku_result = await call_openrouter("anthropic/claude-haiku-4.5", haiku_prompt, max_tokens=50)
|
|
if haiku_result and haiku_result.strip().upper().startswith("YES:"):
|
|
search_query = haiku_result.strip()[4:].strip()
|
|
logger.info("Haiku pre-pass: research needed — '%s'", search_query)
|
|
from x_client import search_tweets, check_research_rate_limit, record_research_usage
|
|
if check_research_rate_limit(user.id if user else 0):
|
|
tweets = await search_tweets(search_query, max_results=10, min_engagement=0)
|
|
logger.info("Haiku research: got %d tweets", len(tweets))
|
|
if tweets:
|
|
research_context = f"\n## LIVE X Search Results (you just searched for '{search_query}' — cite these directly)\n"
|
|
for t in tweets[:7]:
|
|
research_context += f"- @{t['author']}: {t['text'][:200]}\n"
|
|
# Don't burn user's rate limit on autonomous searches (Ganymede)
|
|
# Archive as source
|
|
try:
|
|
slug = re.sub(r"[^a-z0-9]+", "-", search_query[:60].lower()).strip("-")
|
|
filename = f"{datetime.now(timezone.utc).strftime('%Y-%m-%d')}-x-research-{slug}.md"
|
|
source_path = Path(ARCHIVE_DIR) / filename
|
|
tweets_body = "\n".join(f"@{t['author']}: {t['text'][:200]}" for t in tweets[:10])
|
|
source_path.write_text(f"---\ntype: source\nsource_type: x-research\ntitle: \"X research: {search_query}\"\ndate: {datetime.now(timezone.utc).strftime('%Y-%m-%d')}\ndomain: internet-finance\nstatus: unprocessed\nproposed_by: \"@{user.username if user else 'unknown'}\"\ncontribution_type: research-direction\n---\n\n{tweets_body}\n")
|
|
except Exception as e:
|
|
logger.warning("Haiku research archive failed: %s", e)
|
|
except Exception as e:
|
|
logger.warning("Haiku pre-pass failed: %s", e)
|
|
haiku_duration = int((time.monotonic() - t_haiku) * 1000)
|
|
if research_context:
|
|
tool_calls.append({
|
|
"tool": "haiku_prepass", "input": {"query": text[:200]},
|
|
"output": {"triggered": True, "result_length": len(research_context)},
|
|
"duration_ms": haiku_duration,
|
|
})
|
|
|
|
# ─── Query reformulation for follow-ups ────────────────────────
|
|
# Conversational follow-ups ("you're wrong", "tell me more") are unsearchable.
|
|
# Use Haiku to rewrite them into standalone queries using conversation context.
|
|
search_query_text = text # default: use raw message
|
|
user_key = (msg.chat_id, user.id if user else 0)
|
|
hist = conversation_history.get(user_key, [])
|
|
if hist:
|
|
# There's conversation history — check if this is a follow-up
|
|
try:
|
|
last_exchange = hist[-1]
|
|
recent_context = ""
|
|
if last_exchange.get("user"):
|
|
recent_context += f"User: {last_exchange['user'][:300]}\n"
|
|
if last_exchange.get("bot"):
|
|
recent_context += f"Bot: {last_exchange['bot'][:300]}\n"
|
|
reformulate_prompt = (
|
|
f"A user is in a conversation. Given the recent exchange and their new message, "
|
|
f"rewrite the new message as a STANDALONE search query that captures what they're "
|
|
f"actually asking about. The query should work for semantic search — specific topics, "
|
|
f"entities, and concepts.\n\n"
|
|
f"Recent exchange:\n{recent_context}\n"
|
|
f"New message: {text}\n\n"
|
|
f"If the message is already a clear standalone question or topic, return it unchanged.\n"
|
|
f"If it's a follow-up, correction, or reference to the conversation, rewrite it.\n\n"
|
|
f"Return ONLY the rewritten query, nothing else. Max 30 words."
|
|
)
|
|
reformulated = await call_openrouter("anthropic/claude-haiku-4.5", reformulate_prompt, max_tokens=80)
|
|
if reformulated and reformulated.strip() and len(reformulated.strip()) > 3:
|
|
search_query_text = reformulated.strip()
|
|
logger.info("Query reformulated: '%s' → '%s'", text[:60], search_query_text[:60])
|
|
tool_calls.append({
|
|
"tool": "query_reformulate", "input": {"original": text[:200], "history_turns": len(hist)},
|
|
"output": {"reformulated": search_query_text[:200]},
|
|
"duration_ms": 0, # included in haiku timing
|
|
})
|
|
except Exception as e:
|
|
logger.warning("Query reformulation failed: %s", e)
|
|
# Fall through — use raw text
|
|
|
|
# Full retrieval pipeline: keyword → decompose → vector → RRF merge
|
|
retrieval = await orchestrate_retrieval(
|
|
text=text,
|
|
search_query=search_query_text,
|
|
kb_read_dir=KB_READ_DIR,
|
|
kb_index=kb_index,
|
|
llm_fn=call_openrouter,
|
|
triage_model=TRIAGE_MODEL,
|
|
retrieve_context_fn=retrieve_context,
|
|
retrieve_vector_fn=retrieve_vector_context,
|
|
kb_scope=AGENT_KB_SCOPE,
|
|
)
|
|
kb_context_text = retrieval["kb_context_text"]
|
|
kb_ctx = retrieval["kb_ctx"]
|
|
retrieval_layers = retrieval["retrieval_layers"]
|
|
tool_calls.extend(retrieval["tool_calls"])
|
|
|
|
stats = get_db_stats()
|
|
|
|
# Fetch live market data for any tokens mentioned (Rhea: market-data API)
|
|
market_context = ""
|
|
market_data_audit = {}
|
|
token_mentions = re.findall(r"\$([A-Z]{2,10})", text.upper())
|
|
# Entity name → token mapping for natural language mentions
|
|
ENTITY_TOKEN_MAP = {
|
|
"omnipair": "OMFG", "metadao": "META", "sanctum": "CLOUD",
|
|
"drift": "DRIFT", "ore": "ORE", "jupiter": "JUP",
|
|
}
|
|
text_lower = text.lower()
|
|
for name, ticker in ENTITY_TOKEN_MAP.items():
|
|
if name in text_lower:
|
|
token_mentions.append(ticker)
|
|
# Also check entity matches from KB retrieval
|
|
for ent in kb_ctx.entities:
|
|
for tag in ent.tags:
|
|
if tag.upper() in ENTITY_TOKEN_MAP.values():
|
|
token_mentions.append(tag.upper())
|
|
t_market = time.monotonic()
|
|
for token in set(token_mentions):
|
|
try:
|
|
data = await get_token_price(token)
|
|
if data:
|
|
price_str = format_price_context(data, token)
|
|
if price_str:
|
|
market_context += price_str + "\n"
|
|
market_data_audit[token] = data
|
|
except Exception:
|
|
pass # Market data is supplementary — never block on failure
|
|
market_duration = int((time.monotonic() - t_market) * 1000)
|
|
if token_mentions:
|
|
tool_calls.append({
|
|
"tool": "market_data", "input": {"tickers": list(set(token_mentions))},
|
|
"output": market_data_audit,
|
|
"duration_ms": market_duration,
|
|
})
|
|
|
|
# Build Opus prompt — Rio's voice
|
|
prompt = f"""You are Rio, the Teleo internet finance agent. Your Telegram handle is @FutAIrdBot — that IS you. Users tag @FutAIrdBot to reach you. Never say "I'm not FutAIrdBot." You are also @futaRdIO on X. You have deep knowledge about futarchy, prediction markets, token governance, and the MetaDAO ecosystem.
|
|
|
|
## How to sound
|
|
Write like a sharp analyst talking to peers, not like an AI. Specifically:
|
|
- Use your knowledge naturally. Don't say "the KB tracks" or "at experimental confidence" or "our claims show." Just state what you know and how confident you are in plain language.
|
|
- Have a take. You're an analyst, not a summarizer. Say what you actually think.
|
|
- Before you respond, ask yourself: "Does every sentence here add something the user doesn't already know?" If a sentence just restates context, agrees without adding insight, or pads with filler — cut it. Your goal is signal density, not word count.
|
|
- Short questions deserve short answers. If someone asks a factual question, give the fact. Don't surround it with caveats, context, and "the honest picture is" framing.
|
|
- Long answers are fine when the question is genuinely complex or the user asks for depth. But earn every paragraph — each one should contain a distinct insight the previous one didn't cover.
|
|
- Match the user's energy. If they wrote one line, respond in kind.
|
|
- Sound human. No em dashes, no "That said", no "It's worth noting." Just say the thing.
|
|
- No markdown. Plain text only.
|
|
- When you're uncertain, just say so simply. "I'm not sure about X" beats "we don't have data on this yet."
|
|
|
|
## Your learnings (corrections from past conversations — prioritize these over KB data when they conflict)
|
|
{_load_learnings()}
|
|
|
|
## What you know about this topic
|
|
{kb_context_text}
|
|
|
|
## KB Tools — SEARCH UNTIL YOU HAVE ENOUGH
|
|
|
|
You have 8 tools to search the knowledge base. The context above is an initial retrieval pass — it is almost never sufficient on its own. You MUST use tools to verify and deepen your understanding before answering.
|
|
|
|
**Your retrieval loop (follow this every time):**
|
|
1. Review the initial context above. Identify what's missing or unclear.
|
|
2. Use tools to fill gaps — search for sources, explore graph edges, read full claims.
|
|
3. After each tool result, ask yourself: "Do I have enough to give a substantive, grounded answer?"
|
|
4. If NO — search again with different terms, follow more graph edges, read the original source.
|
|
5. If YES — compose your answer. You have up to 6 tool calls, use them.
|
|
|
|
**Tool selection rules:**
|
|
- Someone asks about a specific author/paper/research → call find_by_source AND search_sources to find ALL material from that source
|
|
- You see a claim but need the original article → call read_source with the source title
|
|
- You want to understand the argument structure around a claim → call explore_graph to see what supports, challenges, and depends on it
|
|
- Initial claims don't cover the topic well → call search_kb with refined keywords
|
|
- You want to trace an entity's full network → call list_entity_links then read linked items
|
|
- You want to find original research documents → call search_sources by topic/author
|
|
|
|
**Critical rules:**
|
|
- DO NOT guess or hallucinate details about specific research — use tools to get actual data
|
|
- DO NOT answer from just the initial retrieval context if the question asks about specific research — always trace back to the source
|
|
- When you find a claim, explore its graph edges — connected claims often contain the nuance the user needs
|
|
- If search_kb returns poor results, try search_sources or find_by_source with different keywords
|
|
|
|
{f"## Live Market Data{chr(10)}{market_context}" if market_context else ""}
|
|
|
|
{research_context}
|
|
|
|
{x_link_context}
|
|
|
|
## Conversation History (NEVER ask a question your history already answers)
|
|
{_format_conversation_history(msg.chat_id, user.id if user else 0)}
|
|
|
|
## The message you're responding to
|
|
From: @{user.username if user else 'unknown'}
|
|
Message: {text}
|
|
|
|
Respond now. Be substantive but concise. If they're wrong about something, say so directly. If they know something you don't, tell them it's worth digging into. If they correct you, accept it and build on the correction. Do NOT respond to messages that aren't directed at you — only respond when tagged or replied to.
|
|
|
|
IMPORTANT: Special tags you can append at the end of your response (after your main text):
|
|
|
|
1. LEARNING: [category] [what you learned]
|
|
Categories: factual, communication, structured_data
|
|
Only when genuinely learned something. Most responses have none.
|
|
NEVER save a learning about what data you do or don't have access to.
|
|
|
|
2. RESEARCH: [search query]
|
|
Triggers a live X search and sends results back to the chat. Use when the user asks about recent activity, sentiment, or discussions.
|
|
|
|
3. SOURCE: [description of what to ingest]
|
|
When a user shares valuable source material (X posts, articles, data). Creates a source file in the ingestion pipeline, attributed to the user. Include the verbatim content — don't alter or summarize the user's contribution. Use this when someone drops a link or shares original analysis worth preserving.
|
|
|
|
4. CLAIM: [specific, disagreeable assertion]
|
|
When a user makes a specific claim with evidence that could enter the KB. Creates a draft claim file attributed to them. Only for genuine claims — not opinions or questions.
|
|
|
|
5. CONFIDENCE: [0.0-1.0]
|
|
ALWAYS include this tag. Rate how well the KB context above actually helped you answer this question. 1.0 = KB had exactly what was needed. 0.5 = KB had partial/tangential info. 0.0 = KB had nothing relevant, you answered from general knowledge. This is for internal audit only — never visible to users."""
|
|
|
|
# Call Opus with KB tools — agent can drill into claims, entities, and sources
|
|
from kb_tools import TOOL_DEFINITIONS, execute_tool
|
|
_tool_executor = lambda name, args: execute_tool(name, args, KB_READ_DIR)
|
|
response, kb_tool_audit = await call_openrouter_with_tools(
|
|
RESPONSE_MODEL, prompt, TOOL_DEFINITIONS, _tool_executor, max_tokens=1024,
|
|
max_iterations=6)
|
|
if kb_tool_audit:
|
|
for t in kb_tool_audit:
|
|
if t.get("type") == "reasoning":
|
|
tool_calls.append({"type": "kb_reasoning", **t})
|
|
else:
|
|
tool_calls.append({"tool": f"kb:{t.get('tool', 'unknown')}", **{k: v for k, v in t.items() if k != "tool"}})
|
|
|
|
if not response:
|
|
await msg.reply_text("Processing error — I'll get back to you.", do_quote=True)
|
|
return
|
|
|
|
# Parse LEARNING and RESEARCH tags before posting
|
|
display_response = response
|
|
|
|
# Auto-learning (Rhea: zero-cost self-write trigger)
|
|
learning_lines = re.findall(r'^LEARNING:\s*(factual|communication|structured_data)\s+(.+)$',
|
|
response, re.MULTILINE)
|
|
if learning_lines:
|
|
display_response = re.sub(r'\nLEARNING:\s*\S+\s+.+$', '', display_response, flags=re.MULTILINE).rstrip()
|
|
for category, correction in learning_lines:
|
|
_save_learning(correction.strip(), category.strip())
|
|
logger.info("Auto-learned [%s]: %s", category, correction[:80])
|
|
|
|
# Auto-research (Ganymede: LLM-driven research trigger)
|
|
# Skip if Haiku pre-pass already searched (prevents double-fire + duplicate "No tweets found" messages)
|
|
research_lines = re.findall(r'^RESEARCH:\s+(.+)$', response, re.MULTILINE)
|
|
if research_lines:
|
|
display_response = re.sub(r'\nRESEARCH:\s+.+$', '', display_response, flags=re.MULTILINE).rstrip()
|
|
if not research_context: # Only fire if Haiku didn't already search
|
|
for query in research_lines:
|
|
# Send follow-up with findings (not silent — user expects results)
|
|
asyncio.get_event_loop().create_task(
|
|
_research_and_followup(msg, query.strip(), user))
|
|
logger.info("Auto-research triggered (will follow up): %s", query[:80])
|
|
|
|
# SOURCE: tag — Rio flags content for pipeline ingestion (verbatim, attributed)
|
|
source_lines = re.findall(r'^SOURCE:\s+(.+)$', response, re.MULTILINE)
|
|
if source_lines:
|
|
display_response = re.sub(r'\nSOURCE:\s+.+$', '', display_response, flags=re.MULTILINE).rstrip()
|
|
for source_text in source_lines:
|
|
_create_inline_source(source_text.strip(), text, user, msg)
|
|
logger.info("Inline SOURCE created: %s", source_text[:80])
|
|
|
|
# CLAIM: tag — Rio flags a specific assertion for claim drafting
|
|
claim_lines = re.findall(r'^CLAIM:\s+(.+)$', response, re.MULTILINE)
|
|
if claim_lines:
|
|
display_response = re.sub(r'\nCLAIM:\s+.+$', '', display_response, flags=re.MULTILINE).rstrip()
|
|
for claim_text in claim_lines:
|
|
_create_inline_claim(claim_text.strip(), text, user, msg)
|
|
logger.info("Inline CLAIM drafted: %s", claim_text[:80])
|
|
|
|
# CONFIDENCE: tag — model self-rated retrieval quality (audit only)
|
|
# Handles: "CONFIDENCE: 0.8", "CONFIDENCE: [0.8]", "Confidence: 0.8", case-insensitive
|
|
# Ganymede: must strip from display even if the model deviates from exact format
|
|
confidence_score = None
|
|
confidence_match = re.search(r'^CONFIDENCE:\s*\[?([\d.]+)\]?', response, re.MULTILINE | re.IGNORECASE)
|
|
if confidence_match:
|
|
try:
|
|
confidence_score = max(0.0, min(1.0, float(confidence_match.group(1))))
|
|
except ValueError:
|
|
pass
|
|
# Strip ANY line starting with CONFIDENCE (broad match — catches format deviations)
|
|
display_response = re.sub(r'\n?^CONFIDENCE\s*:.*$', '', display_response, flags=re.MULTILINE | re.IGNORECASE).rstrip()
|
|
|
|
# ─── Audit: write response_audit record ────────────────────────
|
|
response_time_ms = int((time.monotonic() - response_start) * 1000)
|
|
tool_calls.append({
|
|
"tool": "llm_call", "input": {"model": RESPONSE_MODEL},
|
|
"output": {"response_length": len(response), "tags_found": {
|
|
"learning": len(learning_lines) if learning_lines else 0,
|
|
"research": len(research_lines) if research_lines else 0,
|
|
"source": len(source_lines) if source_lines else 0,
|
|
"claim": len(claim_lines) if claim_lines else 0,
|
|
}},
|
|
"duration_ms": response_time_ms - sum(tc.get("duration_ms", 0) for tc in tool_calls),
|
|
})
|
|
|
|
# Claims audit — already built by orchestrate_retrieval with RRF ranking
|
|
claims_audit = retrieval.get("claims_audit", [])
|
|
|
|
# ─── Eval: URL fabrication check ──────────────────────────────
|
|
blocked = False
|
|
block_reason = None
|
|
display_response, fabricated_urls = check_url_fabrication(display_response, kb_context_text)
|
|
if fabricated_urls:
|
|
logger.warning("URL fabrication detected (%d URLs removed): %s", len(fabricated_urls), text[:80])
|
|
|
|
# ─── Eval: confidence floor ────────────────────────────────────
|
|
display_response, blocked, block_reason = apply_confidence_floor(display_response, confidence_score)
|
|
if blocked:
|
|
logger.warning("Confidence floor triggered: %.2f for query: %s", confidence_score, text[:100])
|
|
|
|
# ─── Eval: cost alert ──────────────────────────────────────────
|
|
response_cost = getattr(response, 'cost', 0.0) if response else 0.0
|
|
response_prompt_tokens = getattr(response, 'prompt_tokens', 0) if response else 0
|
|
response_completion_tokens = getattr(response, 'completion_tokens', 0) if response else 0
|
|
if response_cost > COST_ALERT_THRESHOLD:
|
|
logger.warning("Cost alert: $%.4f for query: %s (model=%s)", response_cost, text[:80], RESPONSE_MODEL)
|
|
|
|
# Detect retrieval gap (Rio: most valuable signal for KB improvement)
|
|
retrieval_gap = None
|
|
if not claims_audit and not (kb_ctx and kb_ctx.entities):
|
|
retrieval_gap = f"No KB matches for: {text[:200]}"
|
|
elif confidence_score is not None and confidence_score < 0.3:
|
|
retrieval_gap = f"Low confidence ({confidence_score}) — KB may lack coverage for: {text[:200]}"
|
|
|
|
# Conversation window (Ganymede + Rio: capture prior messages)
|
|
conv_window = None
|
|
if user:
|
|
hist = conversation_history.get((msg.chat_id, user.id), [])
|
|
if hist:
|
|
conv_window = _json.dumps(hist[-5:])
|
|
|
|
try:
|
|
from lib.db import insert_response_audit
|
|
insert_response_audit(
|
|
_audit_conn,
|
|
timestamp=datetime.now(timezone.utc).strftime("%Y-%m-%d %H:%M:%S"),
|
|
chat_id=msg.chat_id,
|
|
user=f"@{user.username}" if user and user.username else "unknown",
|
|
agent="rio",
|
|
model=RESPONSE_MODEL,
|
|
query=text[:2000],
|
|
conversation_window=conv_window,
|
|
entities_matched=_json.dumps([{"name": e.name, "path": e.path}
|
|
for e in (kb_ctx.entities if kb_ctx else [])]),
|
|
claims_matched=_json.dumps(claims_audit),
|
|
retrieval_layers_hit=_json.dumps(list(set(retrieval_layers))),
|
|
retrieval_gap=retrieval_gap,
|
|
market_data=_json.dumps(market_data_audit) if market_data_audit else None,
|
|
research_context=research_context[:2000] if research_context else None,
|
|
kb_context_text=kb_context_text[:10000],
|
|
tool_calls=_json.dumps(tool_calls),
|
|
raw_response=response[:5000],
|
|
display_response=display_response[:5000],
|
|
confidence_score=confidence_score,
|
|
response_time_ms=response_time_ms,
|
|
# Eval pipeline columns (schema v10)
|
|
prompt_tokens=response_prompt_tokens,
|
|
completion_tokens=response_completion_tokens,
|
|
generation_cost=response_cost,
|
|
total_cost=response_cost, # same as generation_cost until embedding cost tracked
|
|
blocked=1 if blocked else 0,
|
|
block_reason=block_reason,
|
|
)
|
|
_audit_conn.commit()
|
|
kb_tool_count = sum(1 for t in tool_calls if t.get("type") == "tool_call" or (t.get("tool", "").startswith("kb:") and t.get("type") != "kb_reasoning"))
|
|
kb_reasoning_count = sum(1 for t in tool_calls if t.get("type") in ("reasoning", "kb_reasoning"))
|
|
logger.info("Audit record written (confidence=%.2f, cost=$%.4f, layers=%s, %d claims, %d kb_tools, %d reasoning_steps, %dms%s)",
|
|
confidence_score or 0, response_cost, retrieval_layers,
|
|
len(claims_audit), kb_tool_count, kb_reasoning_count, response_time_ms,
|
|
", BLOCKED" if blocked else "")
|
|
except Exception as e:
|
|
logger.warning("Failed to write audit record: %s", e)
|
|
|
|
# Post response (without tag lines)
|
|
# Telegram has a 4096 char limit — split long messages
|
|
if len(display_response) <= 4096:
|
|
await msg.reply_text(display_response, do_quote=True)
|
|
else:
|
|
# Split on paragraph boundaries where possible
|
|
chunks = []
|
|
remaining = display_response
|
|
while remaining:
|
|
if len(remaining) <= 4096:
|
|
chunks.append(remaining)
|
|
break
|
|
# Find a good split point (paragraph break near 4000 chars)
|
|
split_at = remaining.rfind("\n\n", 0, 4000)
|
|
if split_at == -1:
|
|
split_at = remaining.rfind("\n", 0, 4096)
|
|
if split_at == -1:
|
|
split_at = 4096
|
|
chunks.append(remaining[:split_at])
|
|
remaining = remaining[split_at:].lstrip("\n")
|
|
# First chunk quotes the original message, rest are standalone follow-ups
|
|
first = True
|
|
for chunk in chunks:
|
|
if chunk.strip():
|
|
await msg.reply_text(chunk, quote=first)
|
|
first = False
|
|
|
|
# Update conversation state: reset window, store history (Ganymede+Rhea)
|
|
if user:
|
|
username = user.username or "anonymous"
|
|
key = (msg.chat_id, user.id)
|
|
unanswered_count[key] = 0 # reset — conversation alive
|
|
entry = {"user": text[:500], "bot": response[:500], "username": username}
|
|
# Per-user history
|
|
history = conversation_history.setdefault(key, [])
|
|
history.append(entry)
|
|
if len(history) > MAX_HISTORY_USER:
|
|
history.pop(0)
|
|
# Chat-level history (group context — all users visible)
|
|
chat_key = (msg.chat_id, 0)
|
|
chat_history = conversation_history.setdefault(chat_key, [])
|
|
chat_history.append(entry)
|
|
if len(chat_history) > MAX_HISTORY_CHAT:
|
|
chat_history.pop(0)
|
|
|
|
# Record rate limit
|
|
if user:
|
|
user_response_times[user.id].append(time.time())
|
|
|
|
# Log the exchange for audit trail
|
|
logger.info("Rio responded to @%s (msg_id=%d)", user.username if user else "?", msg.message_id)
|
|
|
|
# Record bot response to transcript (with internal reasoning)
|
|
_record_transcript(msg, display_response, is_bot=True, rio_response=display_response,
|
|
internal={
|
|
"entities_matched": [e.name for e in kb_ctx.entities] if kb_ctx else [],
|
|
"claims_matched": len(kb_ctx.claims) if kb_ctx else 0,
|
|
"search_triggered": bool(research_context),
|
|
"learnings_written": bool(learning_lines) if 'learning_lines' in dir() else False,
|
|
})
|
|
|
|
# Detect and fetch URLs for pipeline ingestion (all URLs, not just first)
|
|
urls = _extract_urls(text)
|
|
url_content = None
|
|
for url in urls[:5]: # Cap at 5 URLs per message
|
|
logger.info("Fetching URL: %s", url)
|
|
content = await _fetch_url_content(url)
|
|
if content:
|
|
logger.info("Fetched %d chars from %s", len(content), url)
|
|
if url_content is None:
|
|
url_content = content # First URL's content for conversation archive
|
|
_archive_standalone_source(url, content, user)
|
|
|
|
# Archive the exchange as a source for pipeline (slow path)
|
|
_archive_exchange(text, response, user, msg, url_content=url_content, urls=urls)
|
|
|
|
|
|
def _archive_standalone_source(url: str, content: str, user):
|
|
"""Create a standalone source file for a URL shared in Telegram.
|
|
|
|
Separate from the conversation archive — this is the actual article/tweet
|
|
entering the extraction pipeline as a proper source, attributed to the
|
|
contributor who shared it. Ganymede: keep pure (no Rio analysis), two
|
|
source_types (x-tweet vs x-article).
|
|
"""
|
|
try:
|
|
username = user.username if user else "anonymous"
|
|
date_str = datetime.now(timezone.utc).strftime("%Y-%m-%d")
|
|
|
|
# Extract author from URL or content
|
|
author = "unknown"
|
|
author_match = re.search(r"x\.com/(\w+)/", url) or re.search(r"twitter\.com/(\w+)/", url)
|
|
if author_match:
|
|
author = f"@{author_match.group(1)}"
|
|
|
|
# Distinguish tweet vs article (Ganymede: different extraction behavior)
|
|
is_article = "--- Article Content ---" in content and len(content) > 1000
|
|
source_type = "x-article" if is_article else "x-tweet"
|
|
fmt = "article" if is_article else "social-media"
|
|
|
|
slug = re.sub(r"[^a-z0-9]+", "-", f"{author}-{url.split('/')[-1][:30]}".lower()).strip("-")
|
|
filename = f"{date_str}-tg-shared-{slug}.md"
|
|
source_path = Path(ARCHIVE_DIR) / filename
|
|
|
|
# Don't overwrite if already archived
|
|
if source_path.exists():
|
|
return
|
|
|
|
domain, sub_tags = _classify_content(content)
|
|
all_tags = ["telegram-shared", source_type] + sub_tags
|
|
|
|
source_content = f"""---
|
|
type: source
|
|
source_type: {source_type}
|
|
title: "{author} — shared via Telegram by @{username}"
|
|
author: "{author}"
|
|
url: "{url}"
|
|
date: {date_str}
|
|
domain: {domain}
|
|
format: {fmt}
|
|
status: unprocessed
|
|
proposed_by: "@{username}"
|
|
contribution_type: source-submission
|
|
tags: {all_tags}
|
|
---
|
|
|
|
# {author} — {'Article' if is_article else 'Tweet/Thread'}
|
|
|
|
Shared by @{username} via Telegram.
|
|
Source URL: {url}
|
|
|
|
## Content
|
|
|
|
{content}
|
|
"""
|
|
source_path.write_text(source_content)
|
|
logger.info("Standalone source archived: %s (shared by @%s)", filename, username)
|
|
except Exception as e:
|
|
logger.warning("Failed to archive standalone source %s: %s", url, e)
|
|
|
|
|
|
async def _fetch_url_content(url: str) -> str | None:
|
|
"""Fetch article/page content from a URL for pipeline ingestion.
|
|
|
|
For X/Twitter URLs, uses Ben's API (x_client.fetch_from_url) which returns
|
|
structured article content. For other URLs, falls back to raw HTTP fetch.
|
|
"""
|
|
# X/Twitter URLs → use x_client for structured content
|
|
if "x.com/" in url or "twitter.com/" in url:
|
|
try:
|
|
from x_client import fetch_from_url
|
|
data = await fetch_from_url(url)
|
|
if not data:
|
|
logger.warning("x_client returned no data for %s", url)
|
|
return None
|
|
# Format structured content
|
|
parts = []
|
|
# Tweet text
|
|
tweet_text = data.get("text", "")
|
|
if tweet_text:
|
|
parts.append(tweet_text)
|
|
# Article content (contents[] array with typed blocks)
|
|
contents = data.get("contents", [])
|
|
if contents:
|
|
parts.append("\n--- Article Content ---\n")
|
|
for block in contents:
|
|
block_type = block.get("type", "unstyled")
|
|
block_text = block.get("text", "")
|
|
if not block_text:
|
|
continue
|
|
if block_type in ("header-one", "header-two", "header-three"):
|
|
parts.append(f"\n## {block_text}\n")
|
|
elif block_type == "blockquote":
|
|
parts.append(f"> {block_text}")
|
|
elif block_type == "list-item":
|
|
parts.append(f"- {block_text}")
|
|
else:
|
|
parts.append(block_text)
|
|
result = "\n".join(parts)
|
|
return result[:10000] if result else None
|
|
except Exception as e:
|
|
logger.warning("x_client fetch failed for %s: %s", url, e)
|
|
return None
|
|
|
|
# Non-X URLs → raw HTTP fetch with HTML stripping
|
|
import aiohttp
|
|
try:
|
|
async with aiohttp.ClientSession() as session:
|
|
async with session.get(url, timeout=aiohttp.ClientTimeout(total=30)) as resp:
|
|
if resp.status >= 400:
|
|
return None
|
|
html = await resp.text()
|
|
text = re.sub(r"<script.*?</script>", "", html, flags=re.DOTALL)
|
|
text = re.sub(r"<style.*?</style>", "", text, flags=re.DOTALL)
|
|
text = re.sub(r"<[^>]+>", " ", text)
|
|
text = re.sub(r"\s+", " ", text).strip()
|
|
return text[:10000]
|
|
except Exception as e:
|
|
logger.warning("Failed to fetch URL %s: %s", url, e)
|
|
return None
|
|
|
|
|
|
def _extract_urls(text: str) -> list[str]:
|
|
"""Extract URLs from message text."""
|
|
return re.findall(r"https?://[^\s<>\"']+", text)
|
|
|
|
|
|
def _archive_exchange(user_text: str, rio_response: str, user, msg,
|
|
url_content: str | None = None, urls: list[str] | None = None):
|
|
"""Archive a tagged exchange. Conversations go to telegram-archives/conversations/
|
|
(not queue — skips extraction). Sources with URLs already have standalone files."""
|
|
try:
|
|
date_str = datetime.now(timezone.utc).strftime("%Y-%m-%d")
|
|
username = user.username if user else "anonymous"
|
|
slug = re.sub(r"[^a-z0-9]+", "-", user_text[:50].lower()).strip("-")
|
|
filename = f"{date_str}-telegram-{username}-{slug}.md"
|
|
|
|
# Conversations go to conversations/ subdir (Ganymede: skip extraction at source).
|
|
# The cron only moves top-level ARCHIVE_DIR/*.md to queue — subdirs are untouched.
|
|
conv_dir = Path(ARCHIVE_DIR) / "conversations"
|
|
conv_dir.mkdir(parents=True, exist_ok=True)
|
|
archive_path = conv_dir / filename
|
|
|
|
# Extract rationale (the user's text minus the @mention and URL)
|
|
rationale = re.sub(r"@\w+", "", user_text).strip()
|
|
for url in (urls or []):
|
|
rationale = rationale.replace(url, "").strip()
|
|
|
|
# Determine priority — directed contribution with rationale gets high priority
|
|
priority = "high" if rationale and len(rationale) > 20 else "medium"
|
|
intake_tier = "directed" if rationale and len(rationale) > 20 else "undirected"
|
|
|
|
url_section = ""
|
|
if url_content:
|
|
url_section = f"\n## Article Content (fetched)\n\n{url_content[:8000]}\n"
|
|
|
|
domain, sub_tags = _classify_content(user_text + " " + rio_response)
|
|
|
|
content = f"""---
|
|
type: source
|
|
source_type: telegram
|
|
title: "Telegram: @{username} — {slug}"
|
|
author: "@{username}"
|
|
url: "{urls[0] if urls else ''}"
|
|
date: {date_str}
|
|
domain: {domain}
|
|
format: conversation
|
|
status: unprocessed
|
|
priority: {priority}
|
|
intake_tier: {intake_tier}
|
|
rationale: "{rationale[:200]}"
|
|
proposed_by: "@{username}"
|
|
tags: [telegram, ownership-community]
|
|
---
|
|
|
|
## Conversation
|
|
|
|
**@{username}:**
|
|
{user_text}
|
|
|
|
**Rio (response):**
|
|
{rio_response}
|
|
{url_section}
|
|
## Agent Notes
|
|
**Why archived:** Tagged exchange in ownership community.
|
|
**Rationale from contributor:** {rationale if rationale else 'No rationale provided (bare link or question)'}
|
|
**Intake tier:** {intake_tier} — {'fast-tracked, contributor provided reasoning' if intake_tier == 'directed' else 'standard processing'}
|
|
**Triage:** Conversation may contain [CLAIM], [ENTITY], or [EVIDENCE] for extraction.
|
|
"""
|
|
# Write to telegram-archives/ (outside worktree — no read-only errors)
|
|
# A cron moves files into inbox/queue/ and commits them
|
|
archive_path.write_text(content)
|
|
logger.info("Archived exchange to %s (tier: %s, urls: %d)",
|
|
filename, intake_tier, len(urls or []))
|
|
except Exception as e:
|
|
logger.error("Failed to archive exchange: %s", e)
|
|
|
|
|
|
# ─── Batch Triage ───────────────────────────────────────────────────────
|
|
|
|
|
|
async def run_batch_triage(context: ContextTypes.DEFAULT_TYPE):
|
|
"""Batch triage of buffered messages every TRIAGE_INTERVAL seconds.
|
|
|
|
Groups messages into conversation windows, sends to Haiku for classification,
|
|
archives substantive findings.
|
|
"""
|
|
global message_buffer
|
|
|
|
if not message_buffer:
|
|
return
|
|
|
|
# Grab and clear buffer
|
|
messages = message_buffer[:]
|
|
message_buffer = []
|
|
|
|
logger.info("Batch triage: %d messages to process", len(messages))
|
|
|
|
# Group into conversation windows (messages within 5 min of each other)
|
|
windows = _group_into_windows(messages, window_seconds=300)
|
|
|
|
if not windows:
|
|
return
|
|
|
|
# Build triage prompt
|
|
windows_text = ""
|
|
for i, window in enumerate(windows):
|
|
window_msgs = "\n".join(
|
|
f" @{m.get('username', '?')}: {m['text'][:200]}"
|
|
for m in window
|
|
)
|
|
windows_text += f"\n--- Window {i+1} ({len(window)} messages) ---\n{window_msgs}\n"
|
|
|
|
prompt = f"""Classify each conversation window. For each, respond with ONE tag:
|
|
|
|
[CLAIM] — Contains a specific, disagreeable proposition about how something works
|
|
[ENTITY] — Contains factual data about a company, protocol, person, or market
|
|
[EVIDENCE] — Contains data or argument that supports or challenges an existing claim about internet finance, futarchy, prediction markets, or token governance
|
|
[SKIP] — Casual conversation, not relevant to the knowledge base
|
|
|
|
Be generous with EVIDENCE — even confirming evidence strengthens the KB.
|
|
|
|
{windows_text}
|
|
|
|
Respond with ONLY the window numbers and tags, one per line:
|
|
1: [TAG]
|
|
2: [TAG]
|
|
..."""
|
|
|
|
result = await call_openrouter(TRIAGE_MODEL, prompt, max_tokens=500)
|
|
|
|
if not result:
|
|
logger.warning("Triage LLM call failed — buffered messages dropped")
|
|
return
|
|
|
|
# Parse triage results — consolidate tagged windows per chat_id
|
|
# Priority: CLAIM > EVIDENCE > ENTITY when merging windows from same chat
|
|
TAG_PRIORITY = {"CLAIM": 3, "EVIDENCE": 2, "ENTITY": 1}
|
|
chat_tagged: dict[int, dict] = {} # chat_id -> {tag, messages}
|
|
|
|
for line in result.strip().split("\n"):
|
|
match = re.match(r"(\d+):\s*\[(\w+)\]", line)
|
|
if not match:
|
|
continue
|
|
idx = int(match.group(1)) - 1
|
|
tag = match.group(2).upper()
|
|
|
|
if idx < 0 or idx >= len(windows):
|
|
continue
|
|
if tag not in ("CLAIM", "ENTITY", "EVIDENCE"):
|
|
continue
|
|
|
|
window = windows[idx]
|
|
chat_id = window[0].get("chat_id", 0)
|
|
|
|
if chat_id not in chat_tagged:
|
|
chat_tagged[chat_id] = {"tag": tag, "messages": list(window)}
|
|
else:
|
|
# Merge windows from same chat — keep highest-priority tag
|
|
existing = chat_tagged[chat_id]
|
|
existing["messages"].extend(window)
|
|
if TAG_PRIORITY.get(tag, 0) > TAG_PRIORITY.get(existing["tag"], 0):
|
|
existing["tag"] = tag
|
|
|
|
# Archive one source per chat_id
|
|
for chat_id, data in chat_tagged.items():
|
|
_archive_window(data["messages"], data["tag"])
|
|
|
|
logger.info("Triage complete: %d windows → %d sources (%d chats)",
|
|
len(windows), len(chat_tagged), len(chat_tagged))
|
|
|
|
|
|
def _group_into_windows(messages: list[dict], window_seconds: int = 300) -> list[list[dict]]:
|
|
"""Group messages into conversation windows by chat_id and time proximity.
|
|
|
|
Groups by chat_id first, then splits on time gaps > window_seconds.
|
|
Cap per-window at 50 messages (not 10 — one conversation shouldn't become 12 branches).
|
|
"""
|
|
if not messages:
|
|
return []
|
|
|
|
# Group by chat_id first
|
|
by_chat: dict[int, list[dict]] = {}
|
|
for msg in messages:
|
|
cid = msg.get("chat_id", 0)
|
|
by_chat.setdefault(cid, []).append(msg)
|
|
|
|
windows = []
|
|
for chat_id, chat_msgs in by_chat.items():
|
|
# Sort by timestamp within each chat
|
|
chat_msgs.sort(key=lambda m: m.get("timestamp", ""))
|
|
|
|
current_window = [chat_msgs[0]]
|
|
for msg in chat_msgs[1:]:
|
|
# Check time gap
|
|
try:
|
|
prev_ts = datetime.fromisoformat(current_window[-1].get("timestamp", ""))
|
|
curr_ts = datetime.fromisoformat(msg.get("timestamp", ""))
|
|
gap = (curr_ts - prev_ts).total_seconds()
|
|
except (ValueError, TypeError):
|
|
gap = window_seconds + 1 # Unknown gap → force split
|
|
|
|
if gap > window_seconds or len(current_window) >= 50:
|
|
windows.append(current_window)
|
|
current_window = [msg]
|
|
else:
|
|
current_window.append(msg)
|
|
|
|
if current_window:
|
|
windows.append(current_window)
|
|
|
|
return windows
|
|
|
|
|
|
def _archive_window(window: list[dict], tag: str):
|
|
"""Archive a triaged conversation window to inbox/queue/."""
|
|
try:
|
|
date_str = datetime.now(timezone.utc).strftime("%Y-%m-%d")
|
|
first_user = window[0].get("username", "group")
|
|
slug = re.sub(r"[^a-z0-9]+", "-", window[0]["text"][:40].lower()).strip("-")
|
|
filename = f"{date_str}-telegram-{first_user}-{slug}.md"
|
|
|
|
archive_path = Path(ARCHIVE_DIR) / filename
|
|
archive_path.parent.mkdir(parents=True, exist_ok=True)
|
|
|
|
# Build conversation content
|
|
conversation = ""
|
|
contributors = set()
|
|
for msg in window:
|
|
username = msg.get("username", "anonymous")
|
|
contributors.add(username)
|
|
conversation += f"**@{username}:** {msg['text']}\n\n"
|
|
|
|
content = f"""---
|
|
type: source
|
|
source_type: telegram
|
|
title: "Telegram conversation: {slug}"
|
|
author: "{', '.join(contributors)}"
|
|
date: {date_str}
|
|
domain: internet-finance
|
|
format: conversation
|
|
status: unprocessed
|
|
priority: medium
|
|
triage_tag: {tag.lower()}
|
|
tags: [telegram, ownership-community]
|
|
---
|
|
|
|
## Conversation ({len(window)} messages, {len(contributors)} participants)
|
|
|
|
{conversation}
|
|
|
|
## Agent Notes
|
|
**Triage:** [{tag}] — classified by batch triage
|
|
**Participants:** {', '.join(f'@{u}' for u in contributors)}
|
|
"""
|
|
# Write to telegram-archives/ (outside worktree)
|
|
archive_path.write_text(content)
|
|
logger.info("Archived window [%s]: %s (%d msgs, %d participants)",
|
|
tag, filename, len(window), len(contributors))
|
|
except TimeoutError:
|
|
logger.warning("Failed to archive window: worktree lock timeout")
|
|
except Exception as e:
|
|
logger.error("Failed to archive window: %s", e)
|
|
|
|
|
|
# ─── Bot Setup ──────────────────────────────────────────────────────────
|
|
|
|
|
|
async def start_command(update: Update, context: ContextTypes.DEFAULT_TYPE):
|
|
"""Handle /start command."""
|
|
await update.message.reply_text(
|
|
"I'm Rio, the internet finance agent for TeleoHumanity's collective knowledge base. "
|
|
"Tag me with @teleo to ask about futarchy, prediction markets, token governance, "
|
|
"or anything in our domain. I'll ground my response in our KB's evidence."
|
|
)
|
|
|
|
|
|
async def stats_command(update: Update, context: ContextTypes.DEFAULT_TYPE):
|
|
"""Handle /stats command — show KB stats."""
|
|
kb_index.ensure_fresh()
|
|
stats = get_db_stats()
|
|
await update.message.reply_text(
|
|
f"📊 KB Stats:\n"
|
|
f"• {len(kb_index._claims)} claims indexed\n"
|
|
f"• {len(kb_index._entities)} entities tracked\n"
|
|
f"• {len(kb_index._positions)} agent positions\n"
|
|
f"• {stats['merged_claims']} PRs merged\n"
|
|
f"• {stats['contributors']} contributors"
|
|
)
|
|
|
|
|
|
def _load_agent_config(config_path: str):
|
|
"""Load agent YAML config and set module-level variables."""
|
|
global BOT_TOKEN_FILE, RESPONSE_MODEL, TRIAGE_MODEL, AGENT_KB_SCOPE
|
|
global LEARNINGS_FILE, MAX_RESPONSE_PER_USER_PER_HOUR
|
|
|
|
with open(config_path) as f:
|
|
cfg = yaml.safe_load(f)
|
|
|
|
if cfg.get("bot_token_file"):
|
|
BOT_TOKEN_FILE = f"/opt/teleo-eval/secrets/{cfg['bot_token_file']}"
|
|
if cfg.get("response_model"):
|
|
RESPONSE_MODEL = cfg["response_model"]
|
|
if cfg.get("triage_model"):
|
|
TRIAGE_MODEL = cfg["triage_model"]
|
|
if cfg.get("learnings_file"):
|
|
LEARNINGS_FILE = f"/opt/teleo-eval/workspaces/main/{cfg['learnings_file']}"
|
|
if cfg.get("max_response_per_user_per_hour"):
|
|
MAX_RESPONSE_PER_USER_PER_HOUR = cfg["max_response_per_user_per_hour"]
|
|
if cfg.get("kb_scope", {}).get("primary"):
|
|
AGENT_KB_SCOPE = cfg["kb_scope"]["primary"]
|
|
|
|
logger.info("Loaded agent config: %s (scope: %s)", cfg.get("name", "unknown"),
|
|
AGENT_KB_SCOPE or "all domains")
|
|
return cfg
|
|
|
|
|
|
def main():
|
|
"""Start the bot."""
|
|
parser = argparse.ArgumentParser()
|
|
parser.add_argument("--config", help="Agent YAML config file")
|
|
parser.add_argument("--validate", action="store_true", help="Validate config and exit")
|
|
args = parser.parse_args()
|
|
|
|
# Load agent config if provided
|
|
agent_cfg = None
|
|
if args.config:
|
|
agent_cfg = _load_agent_config(args.config)
|
|
if args.validate:
|
|
logger.info("Config valid: %s", args.config)
|
|
return
|
|
|
|
# Load token
|
|
token_path = Path(BOT_TOKEN_FILE)
|
|
if not token_path.exists():
|
|
logger.error("Bot token not found at %s", BOT_TOKEN_FILE)
|
|
sys.exit(1)
|
|
token = token_path.read_text().strip()
|
|
|
|
agent_name = agent_cfg.get("name", "Rio") if agent_cfg else "Rio"
|
|
logger.info("Starting Teleo Telegram bot (%s)...", agent_name)
|
|
|
|
# Initialize persistent audit connection (Ganymede + Rhea: once at startup, not per-response)
|
|
global _audit_conn
|
|
_audit_conn = sqlite3.connect(PIPELINE_DB, timeout=30)
|
|
_audit_conn.row_factory = sqlite3.Row
|
|
_audit_conn.execute("PRAGMA journal_mode=WAL")
|
|
_audit_conn.execute("PRAGMA busy_timeout=10000")
|
|
try:
|
|
from lib.db import migrate
|
|
migrate(_audit_conn)
|
|
logger.info("Audit DB connection initialized, schema migrated")
|
|
except Exception as e:
|
|
logger.error("Audit DB migration failed — audit writes will fail: %s", e)
|
|
|
|
# Prebuild KB index at startup so the first query doesn't pay the 29s rebuild cost
|
|
logger.info("Prebuilding KB index...")
|
|
kb_index.ensure_fresh(max_age_seconds=0) # force immediate build
|
|
logger.info("KB index ready: %d claims, %d entities",
|
|
len(kb_index._claims), len(kb_index._entities))
|
|
|
|
# Build application
|
|
app = Application.builder().token(token).build()
|
|
|
|
# Command handlers
|
|
app.add_handler(CommandHandler("start", start_command))
|
|
app.add_handler(CommandHandler("stats", stats_command))
|
|
|
|
# Tag handler — messages mentioning the bot
|
|
# python-telegram-bot filters.Mention doesn't work for bot mentions in groups
|
|
# Use a regex filter for the bot username
|
|
app.add_handler(MessageHandler(
|
|
filters.TEXT & filters.Regex(r"(?i)(@teleo|@futairdbot)"),
|
|
handle_tagged,
|
|
))
|
|
|
|
# Reply handler — replies to the bot's own messages continue the conversation
|
|
reply_to_bot_filter = filters.TEXT & filters.REPLY & ~filters.COMMAND
|
|
app.add_handler(MessageHandler(
|
|
reply_to_bot_filter,
|
|
handle_reply_to_bot,
|
|
))
|
|
|
|
# All other text messages — buffer for triage
|
|
app.add_handler(MessageHandler(
|
|
filters.TEXT & ~filters.COMMAND,
|
|
handle_message,
|
|
))
|
|
|
|
# Batch triage job
|
|
app.job_queue.run_repeating(
|
|
run_batch_triage,
|
|
interval=TRIAGE_INTERVAL,
|
|
first=TRIAGE_INTERVAL,
|
|
)
|
|
|
|
# Transcript dump job — every 1 hour
|
|
app.job_queue.run_repeating(
|
|
_dump_transcripts,
|
|
interval=3600,
|
|
first=3600,
|
|
)
|
|
|
|
# Audit retention cleanup — daily, 90-day window (Ganymede: match transcript policy)
|
|
async def _cleanup_audit(context=None):
|
|
try:
|
|
_audit_conn.execute("DELETE FROM response_audit WHERE timestamp < datetime('now', '-90 days')")
|
|
_audit_conn.commit()
|
|
logger.info("Audit retention cleanup complete")
|
|
except Exception as e:
|
|
logger.warning("Audit cleanup failed: %s", e)
|
|
|
|
app.job_queue.run_repeating(
|
|
_cleanup_audit,
|
|
interval=86400, # daily
|
|
first=86400,
|
|
)
|
|
|
|
# Run
|
|
logger.info("Bot running. Triage interval: %ds, transcript dump: 1h", TRIAGE_INTERVAL)
|
|
app.run_polling(drop_pending_updates=True)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main()
|