A technical analysis reveals that both Codex and Claude Code exhibit inconsistent behavior when loading instructions from nested directories within a monorepo. Across 76 test runs, neither CLI reliably loaded all instructions, with Claude Code sometimes failing to load AGENTS.md and Codex not loading files in subdirectories by default. The analysis proposes simple fixes, including creating symbolic links for Claude Code and configuring fallback filenames for Codex, to ensure consistent instruction loading. AI
IMPACT Highlights potential issues in how code agents interpret and load instructions from complex project structures.
RANK_REASON Analysis of specific tool behavior and proposed fixes.
Read on dev.to — Claude Code tag →
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
