Usefulness Ratings for Agentic Work

Schema Analysis

Response Size Analysis

30-Day Trend Summary

GraphQL vs REST: Key Architectural Insight

Today revealed a critical reliability split:

GraphQL-based tools (immune to REST rate limits):

• list_discussions — always works, compact
• list_label — always works, scales with count
• search_* tools — always work, well-optimized

REST-based tools (vulnerable to rate limits):

• list_pull_requests, list_workflows, list_code_scanning_alerts — all hit 403 today
• These tools share the same REST rate limit pool

Recommendation: For high-frequency agentic polling, prefer GraphQL-based tools. Use REST tools only for non-time-sensitive operations.

Recommendations

High-value tools (rating 4-5, reliable):

• search_repositories — best for repo discovery, no bloat
• list_discussions — best for compact event stream
• get_file_contents — best for content retrieval with caching support (SHA)
• list_issues — good but limit body size with perPage=1

Tools needing improvement:

• list_code_scanning_alerts — needs pagination support urgently; one call can consume 34K tokens
• list_workflows — per_page parameter should be respected
• get_me — 403 across all 5 days; not useful in this integration context

Context-efficient tools (low tokens, high rating):

• list_discussions: 160 tokens, rating 4 — ideal for lightweight polling
• search_repositories: 220 tokens, rating 5 — ideal for repo routing

Context-heavy tools (high tokens, use sparingly):

• list_code_scanning_alerts: up to 34,584 tokens — only call when alerts are expected
• get_file_contents on README: 2,100–3,500 tokens — cache by SHA to avoid re-fetching