I build agent infrastructure, and I audit the tooling that AI coding agents depend on. Every bug report I file ships with a repro, a root-cause read, and a minimal tested diff.
- cloudflare/cloudflare-docs — PR #32985: fixes llms.txt hub link generation (27 broken links found across Cloudflare's live machine-consumed llms.txt hubs; opened on the docs team's invitation)
- upstash/context7 — PR #3063: fixes Codex config re-setup silently dropping pinned MCP args (multi-line TOML array parsing)
- mensfeld/llm-docs-builder — reported #165; fix shipped in v1.0.0: mailto/tel/ftp links corrupted into dead URLs in docs-for-LLM transforms
- raphaelstolt/llms-txt-php — reported #11; fix shipped in a803a34: invalid XML generation from unescaped markdown text nodes
- 14 more proof-led reports across repomix, gemini-cli, firecrawl, mcpdoc (langchain-ai), clawsec (prompt-security), helicone, and others
AI coding agents waste your time when they work from stale docs and deprecated API patterns. I build a repo-specific context pack (AGENTS.md / CLAUDE.md / .cursorrules add-on) containing:
- current-doc links pinned to the versions your repo actually uses
- deprecated-pattern traps — the exact API shapes agents hallucinate for your stack
- verification commands that force agents to prove work instead of claiming it
$5 flat → get it here — pay in USDC/MATIC, delivery by X DM. Sample deliverable: FreshContext pack walkthrough