Repository navigation
Claude install exposes full repo under skill discovery root #1202
Description
Problem
Claude installs currently document and support cloning the full gstack repo into ~/.claude/skills/gstack.
That path is also under Claude's skill discovery tree. Because the full repo contains generated/source skill artifacts for multiple hosts, Claude can recursively discover host-specific nested SKILL.md files that were never meant to be Claude skills, for example openclaw/skills/*/SKILL.md.
This is the same class of problem fixed for Codex in #235 / #346: the host skill discovery root should not contain the full source checkout.
Expected behavior
Claude's skill root should contain only Claude-discoverable skills plus a minimal gstack/ runtime sidecar for shared assets.
Desired layout:
~/.gstack/repos/gstack/ # full gstack checkout
~/.claude/skills/<skill>/ # top-level Claude skills only
~/.claude/skills/gstack/ # runtime assets only, no nested host SKILL.md files
Why it matters
Users invoking a normal Claude skill like /office-hours should not see OpenClaw/Hermes/Codex host setup behavior leak into Claude skill discovery. The current layout mixes source checkout, generated host artifacts, and runtime sidecar files in one recursive discovery tree.
Proposed fix
- Migrate direct Claude installs from
~/.claude/skills/gstackto~/.gstack/repos/gstack. - Recreate
~/.claude/skills/gstackas a runtime-only sidecar withbin/,browse/,design/,review/,qa/, andETHOS.mdsymlinks. - Keep actual Claude skills as top-level
~/.claude/skills/<skill>/SKILL.mdentries. - Update install docs and team-mode instructions to clone/run setup from
~/.gstack/repos/gstack. - Add regression coverage proving the Claude sidecar exposes no nested
SKILL.mdfiles.
Activity
skyzer commentedon May 31, 2026
I hit this from the Cursor side and implemented the exact layout proposed here on a real machine (global Claude install, gstack
v1.44→ relocated to~/.gstack/repos/gstack). Sharing confirmation + a few implementation notes, and offering to send the PR.The proposed layout works end-to-end. Moving the checkout to
~/.gstack/repos/gstackand keeping only top-level~/.claude/skills/<skill>/aliases dropped the recursively-scanned catalog from 613 → 77SKILL.mdunder~/.claude/skills. 477 of the removed entries were exactly the gitignored host mirrors (.cursor/.hermes/.kiro/.openclaw/.opencode/.slate/.factory/.gbrain/.agents) — i.e. this is the same root cause as #1694. The top-level aliases and the Codex runtime all re-resolved cleanly to the new path.A few notes for whoever implements it (happy to be that person):
1. The Codex relocation is a drop-in template.
migrate_direct_codex_install()(setup, ~L225) already does precisely this for Codex, so amigrate_direct_claude_install()mirror is the natural shape.2.
/gstack-upgradealready handles the relocated layout for free. Its detection (gstack-upgrade/SKILL.mdL89–94) checks~/.claude/skills/gstack/.gitfirst, then~/.gstack/repos/gstack/.git. Because the relocated sidecar has no.git, upgrade resolves to the moved repo automatically — no upgrade changes needed. One constraint: the sidecar must keep itsbin/symlink, since/gstack-upgradeand ~9bin/scripts hard-code~/.claude/skills/gstack/bin/….3. "Zero nested
SKILL.md" needs care — the Codex sidecar doesn't fully meet it today. The Codex runtime root curatesbrowse/,review/,qa/,design/to exclude theirSKILL.md, but still exposes two:gstack/SKILL.mdandgstack/gstack-upgrade/SKILL.md. Worth making the Claude sidecar truly zero and asserting that in the regression test this issue asks for.4. This and #1694 are really one fix, and it has to span every host runtime root. Per Cursor's skills docs, Cursor "walks the skills root recursively" across
~/.cursor/skills,~/.agents/skills, and (for compatibility)~/.claude/skills+~/.codex/skills— with no.cursorignore/exclude mechanism for discovery. So even after relocating the Claude tree, Cursor (and the cursor-agent CLI noted in #1694, which hard-freezes withNGHTTP2_ENHANCE_YOUR_CALMrather than degrading gracefully) still recursively scans the Codex and other host runtimes. A shared "runtime sidecar with no nestedSKILL.md" builder + one regression test applied across all host roots would close #1202 and #1694 together.I have this working locally and can open a PR mirroring
migrate_direct_codex_installunless someone's already on it.time-attack commentedon Jul 15, 2026
@16francej — recommend closing this issue. The checkout is still exposed under a discovery root, and #1819 only partially addresses the problem. It should close only after a corrected #1819 lands with safe sidecar migration, complete runtime behavior, and tests. Please retain the scanner-exposure and runtime-read requirements.