Idea
Expand agr to support additional dependency types beyond skill. Two candidates:
1. Instruction files (type = "instructions")
A directory containing either an AGENTS.md or CLAUDE.md file. These provide project-level or agent-level instructions without being executable skills.
dependencies = [
{handle = "username/repo/instructions", type = "instructions"},
]
2. Ralph loops (type = "ralph")
A directory containing a RALPH.md file following the Ralph loop format from Ralphify. These define structured agent loops.
dependencies = [
{handle = "username/repo/my-loop", type = "ralph"},
]
Open questions
- Where should each type be synced to? Skills go to
skills/ — do instructions go somewhere else?
- Should
agr sync handle all types uniformly, or does each type need its own sync logic?
- Validation: what makes a valid instruction dir vs a valid ralph dir?
- How does this affect
agrx? Should it support running ralph loops ephemerally?
Why
Supporting more types makes agr a general-purpose agent resource manager rather than just a skill manager. Instruction files and ralph loops are both gaining traction in the ecosystem.
Idea
Expand
agrto support additional dependency types beyondskill. Two candidates:1. Instruction files (
type = "instructions")A directory containing either an
AGENTS.mdorCLAUDE.mdfile. These provide project-level or agent-level instructions without being executable skills.2. Ralph loops (
type = "ralph")A directory containing a
RALPH.mdfile following the Ralph loop format from Ralphify. These define structured agent loops.Open questions
skills/— do instructions go somewhere else?agr synchandle all types uniformly, or does each type need its own sync logic?agrx? Should it support running ralph loops ephemerally?Why
Supporting more types makes
agra general-purpose agent resource manager rather than just a skill manager. Instruction files and ralph loops are both gaining traction in the ecosystem.