[FEATURE]: Dynamic AGENTS.md resolution has no low-cost trigger for intentional loading #13550

Feature hasn't been suggested before.

• I have verified this feature I'm about to request hasn't been suggested before.

Describe the enhancement you want to request

The AGENTS.md autoloading (added in #10678) is only triggered by the read tool. This means the only way to trigger autoloading of a subdirectory's AGENTS.md is to read an actual file from that directory.

There's no low-cost way for an agent to load folder-level agent documentation without pulling in potentially unwanted file content. Tools like glob and grep that operate on directories don't trigger autoloading, and explicitly reading the AGENTS.md itself doesn't prevent re-injection on subsequent reads (since explicit reads don't populate metadata.loaded the same way).

The root problem for me is that I want a system where an agent can explore the codebase through its localized agent documentation before reading any code, but also have any unloaded documentation loaded passively as a side effect of the read tool (current approach).

A potential solution is to deduplicate explicit AGENTS.md reads against subsequent autoloads.

This functionality can be recreated with plugins, and I had Claude recreate it as a plugin with deduplication of explicit reads here. The main constraint of this approach is that the built-in dynamic AGENTS.md resolution can't be disabled currently, so I'm forced to use a different filename. However, I think I like my output format better for my needs because auto-loaded localized documentation feels semantically different than a "system-reminder," so I may continue using it even if my suggestion here goes through. An option to disable the built-in auto-loading would let me use AGENTS.md as the filenames still with my plugin, so that might be a nice addition as well.