Summary
graphify has no first-class, consistent way to install the skill to the cross-framework .agents/skills location — in particular the user-global ~/.agents/skills that the Agent-Skills ecosystem (npx skills, and frameworks that follow the spec) reads. Today that location is only reachable by accident. Filing as a scoped feature with a concrete design. PR welcome (good first issue for someone familiar with the tools/skillgen platform system).
Originally reported as #1405 (bare graphify install only populates ~/.claude/skills, not ~/.agents/skills).
Current state (why --platform amp isn't enough)
Bare graphify install is intentionally single-platform (claude on macOS/Linux, windows on Windows → ~/.claude/skills). That part is by design. The real gap is that .agents/skills coverage is piecemeal and inconsistent:
| Location |
Reached by |
Scope |
./.agents/skills (project) |
--platform amp, --platform antigravity |
project only |
~/.config/agents/skills |
--platform amp |
amp global |
~/.gemini/config/skills |
--platform antigravity |
antigravity global |
~/.agents/skills (the standard user-global location) |
only --platform gemini on Windows (__main__.py:147) |
accidental side effect |
So:
- There is no platform named
agents/skills (20 platforms today, none of them this).
--platform amp gives ./.agents/skills for project scope but ~/.config/agents/skills for global — not ~/.agents/skills.
- The only path to user-global
~/.agents/skills is the accidental gemini-on-Windows branch.
Proposed fix — add a first-class agents platform
Add graphify install --platform agents (alias skills?) that installs the skill to the generic Agent-Skills locations:
- project (
--project): ./.agents/skills/graphify/SKILL.md
- global (default):
~/.agents/skills/graphify/SKILL.md
It can reuse the existing agents-md skill body that amp/antigravity already ship to .agents/skills — so no new skill content, just a new platform target.
Implementation notes
- Register the platform in
tools/skillgen/platforms.toml (reusing amp's skill bundle / agents-md core) and regenerate — this goes through the skillgen-check CI job, so python -m tools.skillgen + --bless are required, and the per-host coverage audit must pass.
- Add an
agents branch in graphify/__main__.py:_platform_skill_destination returning the project/global paths above.
- Wire install/uninstall dispatch for the new platform.
- Tests:
_platform_skill_destination("agents", project=…) for both scopes; an end-to-end graphify install --platform agents (global) writing ~/.agents/skills/graphify/SKILL.md; and uninstall removing it.
Acceptance criteria
graphify install --platform agents writes ~/.agents/skills/graphify/SKILL.md; --project writes ./.agents/skills/graphify/SKILL.md.
graphify uninstall / --platform agents removes it.
- No change to bare
graphify install behavior (still single-platform claude/windows).
skillgen-check + full test suite green.
Out of scope
- Making bare
graphify install fan out to multiple skill dirs — that would surprise users and double-install; the named platform is the right model.
Filed from a maintainer-side root-cause of #1405. Happy to review a PR.
Summary
graphify has no first-class, consistent way to install the skill to the cross-framework
.agents/skillslocation — in particular the user-global~/.agents/skillsthat the Agent-Skills ecosystem (npx skills, and frameworks that follow the spec) reads. Today that location is only reachable by accident. Filing as a scoped feature with a concrete design. PR welcome (good first issue for someone familiar with thetools/skillgenplatform system).Originally reported as #1405 (bare
graphify installonly populates~/.claude/skills, not~/.agents/skills).Current state (why
--platform ampisn't enough)Bare
graphify installis intentionally single-platform (claudeon macOS/Linux,windowson Windows →~/.claude/skills). That part is by design. The real gap is that.agents/skillscoverage is piecemeal and inconsistent:./.agents/skills(project)--platform amp,--platform antigravity~/.config/agents/skills--platform amp~/.gemini/config/skills--platform antigravity~/.agents/skills(the standard user-global location)--platform geminion Windows (__main__.py:147)So:
agents/skills(20 platforms today, none of them this).--platform ampgives./.agents/skillsfor project scope but~/.config/agents/skillsfor global — not~/.agents/skills.~/.agents/skillsis the accidental gemini-on-Windows branch.Proposed fix — add a first-class
agentsplatformAdd
graphify install --platform agents(aliasskills?) that installs the skill to the generic Agent-Skills locations:--project):./.agents/skills/graphify/SKILL.md~/.agents/skills/graphify/SKILL.mdIt can reuse the existing agents-md skill body that
amp/antigravityalready ship to.agents/skills— so no new skill content, just a new platform target.Implementation notes
tools/skillgen/platforms.toml(reusing amp's skill bundle / agents-md core) and regenerate — this goes through theskillgen-checkCI job, sopython -m tools.skillgen+--blessare required, and the per-host coverage audit must pass.agentsbranch ingraphify/__main__.py:_platform_skill_destinationreturning the project/global paths above._platform_skill_destination("agents", project=…)for both scopes; an end-to-endgraphify install --platform agents(global) writing~/.agents/skills/graphify/SKILL.md; and uninstall removing it.Acceptance criteria
graphify install --platform agentswrites~/.agents/skills/graphify/SKILL.md;--projectwrites./.agents/skills/graphify/SKILL.md.graphify uninstall/--platform agentsremoves it.graphify installbehavior (still single-platform claude/windows).skillgen-check+ full test suite green.Out of scope
graphify installfan out to multiple skill dirs — that would surprise users and double-install; the named platform is the right model.Filed from a maintainer-side root-cause of #1405. Happy to review a PR.