Agent Resource Roots

mevedel treats .agents/ as the interoperable agent-resource family shared with other AI harnesses and .mevedel/ as its mevedel-native counterpart. Every supported .agents/XYZ resource has a corresponding .mevedel/XYZ override unless an ADR records a deliberate exception. New portable resources prefer .agents/, mevedel-specific resources and overrides use .mevedel/, and existing resources are updated in place.

Unless a subsystem documents a deliberate exception, resource precedence from highest to lowest is workspace .mevedel/, workspace .agents/, global ~/.mevedel/, global ~/.agents/, then explicit defcustom roots. This precedence decides conflict winners; ordered systems such as hooks may execute lower-precedence layers first so higher-precedence layers can refine them. Systems that retain colliding resources under prefixes, such as skills, keep both families instead of discarding the lower-precedence entry.

Standalone hooks are included in the resource-root model. Global .agents/hooks.* files are trusted user files like ~/.mevedel/hooks.*; project .agents/hooks.* files use the same workspace trust-by-hash model as project .mevedel/hooks.*. Plugin hooks may also come from plugins discovered in either plugin family, but those plugins remain inactive until explicitly enabled in a workspace.

When same-named skills are found in multiple roots or bundled skills, mevedel keeps them all and uses the shortest unique visible prefix from scope (global/local) and family (agents/mevedel), preferring family-only prefixes when either dimension is sufficient. Bundled and managed skills keep their existing bundled: and managed: conflict prefixes.

Persistent memory is merged from local .mevedel/, local .agents/, global .mevedel/, then global .agents/, with earlier layers treated as more specific when memories conflict. New broadly portable memory writes prefer .agents/ in the appropriate scope; existing memories are updated where they already live.

The workspace journal is a deliberate exception: .mevedel/journal/ is generated, uncurated evidence and runtime control state, not a portable agent-authored resource. It has no .agents/journal/ counterpart and no global root. Its immutable public records are separate from private .mevedel/state/journal/ records. See ADR 0117.

Skill discovery uses the same local-before-global and .mevedel/-before- .agents/ order for ordinary resource roots, followed by bundled skills and plugin skills. Hook execution visits lower-precedence hook roots before higher-precedence hook roots because later hook rewrites win. Plugin discovery includes local .mevedel/plugins/ and .agents/plugins/ roots, followed by global ~/.mevedel/plugins/ and ~/.agents/plugins/ roots, but discovery never implies activation. /plugin install writes new installs to the interoperable global ~/.agents/plugins/ root. Existing managed global plugins in either family are updated and removed in place.

When the same plugin manifest name appears in multiple roots, mevedel uses root precedence rather than exposing multiple prefixed plugin names: workspace .mevedel/plugins/, workspace .agents/plugins/, global ~/.mevedel/plugins/, global ~/.agents/plugins/, then extra roots. Plugin details expose shadowed duplicates and their source roots.

Workspace instructions remain in ordinary AGENTS.md and AGENTS.local.md files along the workspace directory chain. mevedel does not read CLAUDE.md or introduce .agents/AGENTS.md as a second instruction location. The initial prompt loads the chain through the session working directory; successful Read calls lazily deliver deeper applicable files as host-generated reminders, deduplicated by conversation owner, canonical path, and content hash.