Auto‑Blog Agent: Automating Weekly GitHub‑to‑LinkedIn Content

Architecture Overview

graph LR
A[GitHub Actions] --> B[Python Agent]
B --> C{Sync Phase}
C --> D[Fetch Repos]
C --> E[Get Week]
C --> F[Get Processed]
B --> G[Async Phase]
G --> H[Process Repo]
H --> I[LLM Call]
H --> J[Save Blog]
H --> K[Save Project]
H --> L[Post LinkedIn]
H --> M[Mark Processed]

Core Components

• agent.core.github – Uses httpx to call the GitHub REST API. It fetches public repos, filters by creation date, and extracts the README. The README is truncated to config.github.max_readme_length characters to keep the prompt size reasonable.
• agent.core.llm – Wraps OpenRouter httpx calls. It sends a system prompt and user prompt, then validates the JSON with Pydantic v2. If validation fails it retries with the fallback model.
• agent.core.database – A thin wrapper over supabase-py. All read/write operations are synchronous; we use asyncio.to_thread to keep the event loop unblocked.
• agent.core.linkedin – Posts to LinkedIn via the ugcPosts API. Errors are caught and the raw LLM output is stored in agent_processed_repos.raw_llm_output for manual recovery.

Async Strategy

async def _process_repo(repo):
    readme = await to_thread(fetch_readme, repo)
    llm_output = await to_thread(llm_call, readme)
    blog = BlogPostInsert.from_llm_output(llm_output)
    project = ProjectInsert.from_llm_output(llm_output)
    await to_thread(save_blog, blog)
    await to_thread(save_project, project)
    try:
        await to_thread(post_to_linkedin, llm_output.linkedin_post)
    except Exception as e:
        log.error("LinkedIn post failed", exc_info=e)
    await to_thread(mark_repo_processed, repo.id, status="success")

Trade‑offs

• Using httpx sync client keeps dependencies minimal but forces to_thread. The overhead is negligible for a handful of repos.
• The JSON contract guarantees deterministic downstream processing. It also makes unit tests trivial.
• The audit table records every repo even in dry‑run mode; you need to clear it before a real run.

Future Work

Adding a retry back‑off for LinkedIn, a cache for LLM responses, and a UI to trigger manual runs would make the agent even more robust.