Multi-Agent Architecture (v2)

Agent Overview

Agent	Role	Key Action	Dosya
BrainAgent	Otonom karar motoru	decide(), force_produce()	agents/brain.py
Orchestrator	Koordinasyon	plan_week(), daily_check()	agents/orchestrator.py
Planner	Konu secimi	suggest_topic()	agents/planner.py
Creator	Icerik uretimi	create_post(), create_reels_prompt()	agents/creator.py
Reviewer	Kalite kontrol	review_post() (score 1-10)	agents/reviewer.py
Publisher	Platform yayin	publish()	agents/publisher.py
Analytics	Performans takip	fetch_metrics()	agents/analytics.py
BaseAgent	Temel sinif	call_claude_with_retry()	agents/base_agent.py

v2 Engine Layer

V2Scheduler (app/engine/scheduler.py)

Feed aggregation + Brain karar donguleri. v1 scheduler ile paralel calisir.

from app.engine.scheduler import V2Scheduler

scheduler = V2Scheduler(brain_agent=brain, feed_aggregator=aggregator, event_bus=bus)
await scheduler.start()  # Feed loop (30m) + Brain loop (120m) + Expiry loop (6h)

SystemState (app/engine/state.py)

Aktif production tracking, pause state, pool status.

from app.engine.state import SystemState

state = SystemState()
state.is_paused                    # Sistem duraklatilmis mi?
state.is_production_active()       # Aktif uretim var mi?
state.get_full_state()             # Tum durum bilgisi (Brain icin)
state.register_production(key)     # Uretim baslat
state.complete_production(key, t)  # Uretim tamamla

EventBus (app/engine/event_bus.py)

Async pub/sub bilesenleri arasinda loose coupling.

from app.engine.event_bus import EventBus

bus = EventBus()
bus.subscribe("brain_decision", callback)
await bus.publish("feed_updated", {"active_count": 42})

BrainAgent (Otonom Karar Motoru)

Her 2 saatte bir calisir. Claude API ile karar verir.

from app.agents.brain import BrainAgent

brain = BrainAgent()
brain.event_bus = event_bus  # main.py'de set edilir (agents/__init__.py'den degil)

# Karar dongusu
decision = await brain.decide()
# -> {"action": "produce"|"wait", "reason": "...",
#     "opportunity_id": 123, "content_type": "reels",
#     "model_id": "kling-3.0-pro", "visual_style": "cinematic_4k",
#     "hook_type": "question", "voice_mode": false}

# Telegram /force komutu
result = await brain.force_produce(opp_id=42, content_type="reels")

# Son kararlar (Telegram /brain icin)
decisions = brain.get_last_decisions(limit=5)

# Dry-run mode (env: BRAIN_DRY_RUN=true)
brain.is_dry_run  # True = kararlar loglanir, uretim yapilmaz

Brain Karar Sureci

1. Sistem durumu toplanir (SystemState)
2. Quick checks (gece saati, min interval, havuz durumu)
3. Claude API ile karar (produce/wait)
4. Uygun production pipeline tetiklenir

Brain Creative Parameters

• model_id: sora-2, sora-2-pro, kling-3.0-pro
• visual_style: cinematic_4k, 3d_render, neon_cyberpunk, anime, minimalist
• hook_type: question, statistic, bold_claim, problem, value, fear, before_after, list, comparison, local
• carousel_style: tech_blue, energy_green, warm_industrial, dark_premium, clean_minimal
• carousel_layout: data_heavy, storytelling, comparison, step_by_step, tips_list

Feed System

10 RSS feed'den icerik firsatlari toplanir, skorlanir, havuza eklenir.

from app.sources.feed_aggregator import FeedAggregator

aggregator = FeedAggregator()
result = await aggregator.run_feed_pipeline()

Feed -> Opportunity Lifecycle

RSS Feed → discovered → enriched → scored → ready → selected → producing → used
                                                        ↓
                                                     expired (72h)

content_opportunities Tablosu

• source_type: rss, evergreen, calendar, manual
• combined_score: relevance + timeliness + virality potential
• content_type_suggestion: reels, carousel, post, voice_reels
• Brain Agent ready durumundaki en yuksek skorlu firsati secer

v2 Production Pipelines (7 pipeline)

Pipeline	Dosya	Default Model
post	production/post_pipeline.py	-
reels	production/reels_pipeline.py	kling-3.0-pro
carousel	production/carousel_pipeline.py	-
voice_reels	production/voice_reels_pipeline.py	sora-2-pro
long_video	production/long_video_pipeline.py	kling-3.0-pro
news_reels	production/news_reels_pipeline.py	kling-3.0-pro
conversational	production/conversational_pipeline.py	sora-2-pro

Brain Agent dinamik olarak pipeline tetikler:

# BrainAgent.PIPELINE_MAP
pipeline_info = brain.PIPELINE_MAP[content_type]
# -> ("app.production.reels_pipeline", "ReelsPipeline")

v1 Pipeline (hala aktif, paralel calisir)

from app.scheduler.pipeline import ContentPipeline

pipeline = ContentPipeline(telegram_callback=callback)
await pipeline.run_daily_content()      # Telegram onayli
await pipeline.run_autonomous_content() # Otomatik
await pipeline.run_reels_content()      # Video

v1 Pipeline States

IDLE → PLANNING → AWAITING_TOPIC_APPROVAL →
CREATING_CONTENT → AWAITING_CONTENT_APPROVAL →
CREATING_VISUAL → AWAITING_VISUAL_APPROVAL →
REVIEWING → AWAITING_FINAL_APPROVAL →
PUBLISHING → COMPLETED | ERROR

System Startup (main.py)

python3 app/main.py
  1. Database init
  2. EventBus olustur
  3. v2: FeedAggregator + BrainAgent + V2Scheduler
  4. v1: ContentPipeline + Scheduler (paralel)
  5. Telegram bot baslat (brain_agent + feed_aggregator global set)

BaseAgent Pattern

from app.agents.base_agent import BaseAgent

class MyAgent(BaseAgent):
    def __init__(self):
        super().__init__("myagent")  # Loads persona from context/agent-personas/

    async def execute(self, input_data):
        response = await self.call_claude_with_retry(prompt="...", timeout=120, max_retries=3)
        return json.loads(response)

Deep Links

• app/agents/brain.py - Brain Agent (otonom karar)
• app/engine/scheduler.py - V2Scheduler (feed + brain loops)
• app/engine/state.py - SystemState
• app/engine/event_bus.py - EventBus (pub/sub)
• app/sources/feed_aggregator.py - Feed pipeline
• app/sources/feed_config.py - 10 RSS feed
• app/production/ - 7 production pipeline
• app/scheduler/pipeline.py - v1 pipeline (paralel calisir)
• app/agents/base_agent.py - Base class
• app/main.py - v2 entry point