siva010505/helio2

Purpose and Execution Surface

This is a Python agent pipeline for producing and operating video content: research, scripting, visual planning, assembly, publishing, and analytics-driven iteration. Primary runnable surfaces are:

• src/pipeline.py — likely programmatic pipeline entry.

• scripts/run_daily_pipeline.py — daily end-to-end execution.

• scripts/dry_run.py — non-production validation path.

• .github/workflows/daily_pipeline.yml — scheduled/CI automation boundary.

Orchestration and Pipeline Control

The central control plane is an orchestrator that sequences specialized agents and carries state between stages:

• src/agents/orchestrator.py

• src/pipeline.py

• src/config_loader.py

Configuration is externalized in config.yaml, with environment-provided secrets expected through .env.example. The orchestration boundary separates workflow decisions from individual content-generation, rendering, and publishing capabilities.

Content Intelligence and Script Generation

Topic discovery and editorial development are decomposed into specialized agents:

• src/agents/research_agent.py

• src/agents/longform_research_agent.py

• src/agents/deep_dive_script_agent.py

Supporting stages refine output for format and performance: script_agent.py, punch_agent.py, caption_agent.py, seo_agent.py, and scoring_agent.py under src/agents/. This suggests a flow from research evidence to structured script, engagement edits, metadata/captions, and quality scoring.

Media Planning and Production Assembly

Video creation is divided between creative direction, asset planning, narration, thumbnails, and final assembly:

• src/agents/visual_director_agent.py

• src/agents/visual_planner_agent.py

• src/agents/assembly_agent.py

Additional production boundaries include voice_agent.py, thumbnail_agent.py, and compilation_assembler_agent.py. Local media inputs/branding reside under assets/, including assets/video/intro.mp4; stock-media lookup is isolated in src/services/stock_search.py.

Persistence, LLMs, and External Integrations

Persistent pipeline/agent state is modeled behind a database layer:

• src/db/models.py

• src/db/db.py

• src/db/init_db.py

The checked-in data/agent.db indicates a SQLite-style local datastore. Model-provider interaction is centralized in src/llm_client.py, preventing agents from individually owning LLM transport. Publishing is an external-service boundary handled by:

• src/agents/upload_agent.py

• src/youtube/oauth.py

Feedback Loop and Operations

Post-publication analytics feed back into future content decisions:

• src/agents/analytics_agent.py

• src/agents/evaluation_agent.py

• scripts/run_analytics_pull.py

scripts/run_analytics_feedback.py operationalizes the feedback pass, while run_daily_pipeline.py combines production execution with recurring automation. The resulting architecture is a staged, stateful agent system with explicit boundaries for LLM calls, media sourcing, database persistence, YouTube authentication/upload, and analytics ingestion.