tool_github schema exceeds Anthropic 200k token context limit

[joshuadarron]

Problem

Connecting tool_github to an agent causes the prompt to exceed Anthropic's 200,000 token context limit before the agent receives its first message.

Error:

Error code: 400 - {'type': 'error', 'error': {'type': 'invalid_request_error',
'message': 'prompt is too long: 206767 tokens > 200000 maximum'}}

Reproduction

Pipeline: agent_langchain (or agent_rocketride) + llm_anthropic (claude-sonnet-4-5) + tool_github + tool_http_request

Agent instructions are minimal (~14 short strings). The user question is a small JSON object. The token overflow comes entirely from the tool_github schema injection.

A second run without tool_github (same agent, same instructions, same LLM) measured 233k tokens — suggesting the schema may also be cached/accumulated across pipeline sessions.

Workaround

Remove tool_github and use tool_http_request with a URL whitelist (^https://api\.github\.com/.*) instead. Agent instructions specify each GitHub API call explicitly.

Suggestions

• Lazy-load tool schemas — only inject when the agent requests a specific operation
• Offer a slim/scoped mode for tool_github that exposes a subset of endpoints
• Document the approximate token cost of each tool component so users can budget context

[github-actions]