Claude Code Silently Skips AGENTS.md When Telemetry Is Disabled

Claude Code 2.1.277 shipped support for AGENTS.md, a shared instruction file meant to replace the tool-specific CLAUDE.md. But the feature never activates for users who disable telemetry. The loader is packaged as a built-in plugin whose availability is decided by a remote feature flag (tengu_agents_md_mod) that defaults to false. When either CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC or DISABLE_TELEMETRY is set, or when the client runs against Bedrock, Vertex, or a third-party gateway where the flag can’t resolve, the plugin stays off and the local file is ignored — even though reading a markdown file from the working directory requires no network at all.

The author confirmed the behavior with a canary-word test and a string search through the binary, finding that no warning is ever printed. The session simply starts, the model answers without the project instructions, and nothing signals that a file was skipped. Per-repo overrides in .claude/settings.json don’t help; only a session-level —settings flag that clears both variables restores the feature. The practical workaround is a one-line CLAUDE.md containing @AGENTS.md, since the import mechanism doesn’t consult the flag — reintroducing exactly the extra per-repo file that AGENTS.md was supposed to eliminate.

The core objection is one of design principle: a privacy setting should never quietly disable unrelated local behavior. The gate disproportionately affects the users most likely to adopt AGENTS.md — those careful about outbound traffic and teams on enterprise gateways — and the silent failure leads people to blame the model for ignoring instructions the file never delivered. The author argues that a flag that can’t be fetched should fall back to documented behavior or at least emit a startup warning, and adds a wishlist for a global AGENTS.md and native support for shared agent skills.