Web Fetch Summarization

When the agent fetches a web page (documentation, API references, etc.), the raw HTML converts to 10K-50K+ tokens of markdown. Most of it is navigation, sidebars, and boilerplate. GitAuto uses Claude Haiku 4.5 as a summarization layer to extract only the relevant information before passing it to the main reasoning model.

Why This Exists

The main reasoning model (Claude Opus) costs $5/$25 per million input/output tokens. Claude Haiku costs $1/$5. Passing a 30K-token web page directly to Opus costs ~$0.15 in input tokens alone. Running it through Haiku first and returning a 1K-token summary costs ~$0.03 for Haiku input + ~$0.005 for Haiku output + ~$0.005 for Opus input on the summary. That's roughly an 80% cost reduction per web fetch.

How It Works

1. The agent calls web_fetch with a URL and a prompt describing what information to extract.
2. GitAuto fetches the page, strips unnecessary HTML elements (nav, footer, ads, scripts), and converts the main content area to markdown.
3. The markdown and the extraction prompt are sent to Claude Haiku, which returns a focused summary containing only the requested information.
4. The summary (not the full markdown) is returned to the main model's conversation.

Two Tools, Not One

Not every URL needs summarization. JSON API responses, raw text files, and configuration files should be returned as-is. GitAuto provides two tools:

• web_fetch - Fetches HTML pages, converts to markdown, summarizes with Haiku. For documentation, articles, and web content.
• curl - Fetches raw content with no processing. For JSON APIs, text files, and anything where exact content matters.

Why the Model Cannot Solve This

The main model is smart enough to ignore irrelevant content on a web page. But by the time it sees the content, you have already paid for the input tokens. Asking the model to "focus on the relevant parts" does not reduce the cost - the full page is already in the context window. The filtering must happen before the tokens reach the expensive model, which is an application-layer decision, not a model capability.