<![CDATA[Harbor Engineering]]>https://harborai.hashnode.devRSS for NodeSat, 07 Nov 2026 11:19:17 GMT60<![CDATA[We Audited 481 Public CLAUDE.md and AGENTS.md Files: Here Is What Claude Code, Cursor, and Codex Silently Ignore]]>https://harborai.hashnode.dev/we-audited-481-public-claude-md-and-agents-md-files-here-is-what-claude-code-cursor-and-codex-silently-ignorehttps://harborai.hashnode.dev/we-audited-481-public-claude-md-and-agents-md-files-here-is-what-claude-code-cursor-and-codex-silently-ignoreWed, 28 Oct 2026 15:28:07 GMTMost engineering teams in 2026 use at least two AI coding agents side by side: Claude Code, Cursor, OpenAI Codex CLI, or GitHub Copilot.

To find out what actually happens to repo instruction files in practice, we audited 481 public CLAUDE.md and AGENTS.md files and traced the exact file-loading rules across all 7 major agent harnesses.

You can test your own repo's CLAUDE.md or AGENTS.md in your browser (free, client-side, no login) at https://gethrbr.com/check.

Here are the 5 silent failure modes that break almost every multi-agent repo today:

1. The 7-Agent Instruction Loading Matrix (2026)

File & Location Claude Code Codex CLI Cursor GitHub Copilot
CLAUDE.md (Repo Root) Loaded at launch Ignored (unless added to project_doc_fallback_filenames in ~/.codex/config.toml) Loaded in Agent chat only (Cursor Tab, Inline Edit Cmd+K, and Bugbot never see it) Ignored
AGENTS.md (Repo Root) Ignored the moment any CLAUDE.md exists in the folder or above it (unless CLAUDE.md starts with @AGENTS.md) Loaded at launch (capped at 32 KiB) Loaded in Agent chat only (Tab, Inline Edit, and Bugbot never see it) Loaded at launch
AGENTS.md (Subfolder) Ignored Ignored when Codex starts at the repo root (Codex walks root to cwd and stops there) Loaded when Agent chat works in that folder Ignored (root only)
.cursor/rules/*.mdc Ignored Ignored Loaded across Tab, Inline Edit, and Agent chat (when alwaysApply: true) Ignored

2. Codex CLI Silently Cuts AGENTS.md Mid-Sentence at 32 KiB (32,768 bytes)

Codex enforces a hard 32 KiB budget (project_doc_max_bytes) across the combined AGENTS.md chain. The moment your file crosses byte 32,768, Codex truncates the tail mid-sentence with zero warning in the terminal. Any git, testing, or security rule at the bottom of a long AGENTS.md disappears silently.

3. Prose Pointers (See AGENTS.md) Do Not Auto-Load in Claude Code

If your CLAUDE.md says Read [AGENTS.md](./AGENTS.md) before starting, Claude Code does not load AGENTS.md at session start. It only sees AGENTS.md if the model decides to spend a tool call opening it. To make Claude Code inline AGENTS.md automatically at startup, line 1 of CLAUDE.md must be @AGENTS.md.

4. @path/to/file.md Imports Break Outside Claude Code

While Claude Code expands @path/to/doc.md imports at startup (up to 4 hops deep, burning the full token count on every turn), Codex CLI, Cursor, and Copilot treat @path/... as literal plain text and never open the referenced file.

5. 60% of Repos Ask in Prose for Guardrails That Require a PreToolUse Hook

Across the 481 files we studied, 60% used words like NEVER, MUST NOT, or CRITICAL in prose to ban dangerous commands (rm -rf, git push --force, git reset --hard, --no-verify, or reading .env). Because CLAUDE.md and AGENTS.md are injected as user-message context, compliance degrades as the context window fills. Hard blocks belong in .claude/settings.json (permissions.deny or a PreToolUse hook exiting 2).

Check Your File (and Sync Rules Automatically)

  1. Free Browser Linter: Paste your CLAUDE.md, AGENTS.md, or .cursor/rules/*.mdc into https://gethrbr.com/check to see which agents load it, where it hits byte/line limits, and which prose lines need to become hooks.
  2. Harbor (Shared Memory & Rule Routing for Coding Agents): We built Harbor (https://gethrbr.com/#start) so teams using Claude Code, Cursor, Codex, and Copilot don't have to maintain four drifting markdown files or load 10,000 tokens of static rules on every turn. Join the early access beta at https://gethrbr.com/#start.
]]>