Comparable systems survey

Collected 2026-09-23 by four research agents: two web surveys and two code inspections of local clones under related-systems/, updated the same day. Claims marked unverified come from search summaries or inference. File paths are relative to each clone.

Harness mechanisms

Harness	Loose skills	Plugins	No-copy route
Claude Code	~/.claude/skills, project .claude/skills, managed dir, .claude/skills inside --add-dir dirs; symlinked skill dirs supported; no setting for extra dirs	Marketplace plugins copied to ~/.claude/plugins/cache per version. command source in link mode (v2.1.229+) runs a command that prints a plugin dir and uses it in place, re-run once per session; not on Windows. Local-directory marketplaces and --plugin-dir load in place. Components outside the plugin root are rejected, symlinks included	Link-mode plugin; symlink
Codex	.agents/skills up to repo root, ~/.agents/skills, /etc/codex/skills; [[skills.config]] only enables/disables; follows symlinks	Copied to ~/.codex/plugins/cache, including local plugins; issue #18863: cache install drops symlinks	Symlink only
OpenCode	skills array in opencode.json takes absolute paths and HTTP catalogs (v2 docs)	—	Config path
Gemini CLI	~/.gemini/skills, ~/.agents/skills, workspace dirs; /skills link symlinks	Extensions installed to ~/.gemini/extensions; extensions link symlinks a dev dir	Symlink
Goose, Cursor	Fixed dirs; no configurable path found	—	Symlink at best

Requests for configurable skill paths were closed without visible maintainer rationale: Claude Code #22902, #77766 (stale); Codex #13074 (not planned). Codex #22869 appeared open. The Agent Skills spec (agentskills.io) does not mandate locations and lists user-configured paths as optional; it does not address packaging or versioning.

Sources: code.claude.com/docs/en/skills, /plugins-reference, /plugin-marketplaces; learn.chatgpt.com/docs/build-skills; opencode.ai/v2/docs/skills/; geminicli.com/docs/cli/skills/; agentskills.io/client-implementation/adding-skills-support.

Systems inspected in local clones

System	Ships	Channel	Copied vs in place	Drift handling
GBrain	Bun CLI + MCP server, ~80 skills, shared convention docs	Claude and Codex plugins over a generated, version-stamped plugin/skills tree (scripts/generate-plugin-tree.ts); skillpack scaffold into harness skill dirs; stub mode	Plugin: in place, skills reach ../conventions/*.md relatively. Scaffold: copies, additive, "local edits win". Stubs: frontmatter-only SKILL.md telling the agent to call gbrain skill <slug> or MCP get_skill — "always current" (src/core/skillpack/harness-bridge.ts:76-78,291-312)	sha256 per file at install (bridge-state.ts:10-16) → three-way: local edit vs upstream change; skillpack reference --apply-clean-hunks merges clean upstream hunks. MCP launcher via ${CLAUDE_PLUGIN_ROOT} is Unix-only
Ars Contexta	~249-note methodology library, meta-skills, skill templates, hooks	Claude Code plugin only	Library read in place via ${CLAUDE_PLUGIN_ROOT}/...; working skills generated into project .claude/skills by /setup	generated_from version stamp; /upgrade uses git status and an LLM judgment per skill; keep/merge/replace with archive; rejects hashing deliberately
nvk/llm-wiki	One skill with ~22 references, commands, Python CLI, hooks	Claude and Codex plugins; OpenCode fetches a raw GitHub URL each session	In place from plugin cache; relative reference links; ${CLAUDE_PLUGIN_ROOT}/bin/llm-wiki	Harness plugin update; README documents manual cache repair; replaced a symlink with a copy because the Codex cache broke it
Basic Memory	Python CLI + MCP server; Claude and Codex plugins; shared skills	bm install claude-code/codex wraps the harness plugin install; skills also shipped as wheel package data (pyproject.toml force-include → basic_memory/data/...), located with importlib.metadata.distribution().locate_file, and copied to user-level Tau/Pi dirs	Hooks run uv run --script and import the installed package	Install plan compares previous bytes, refuses changes made after preview (cli/commands/install.py); schemas seeded once, customizations win
claude-obsidian	Python core, 17 skills, hooks, templates	Claude Code plugin; setup-multi-agent.sh symlinks each skill into other harnesses' dirs	Symlinks to one canonical tree; skills resolve their own install root	Never replaces existing files (CONFLICT); --check verifies link targets
OKF Harness	npm packages, one skill with references	Claude and Codex marketplaces; npx @okf-harness/setup	User-level host skill + per-workspace rendered skills	Frontmatter markers: installed / version-drifted / unmanaged-conflict; workspace pins exact runtime version
hyalo	Rust CLI, skills, rule, managed CLAUDE.md/AGENTS.md section	hyalo init writes templates embedded in the binary; Codex plugin	Copies; removes project copies when the Codex plugin is used	hyalo:managed marker; refuses to overwrite unmarked files or write through symlinks
ai-modules	Skills, agents, commands, hooks for six harnesses	Claude and Codex plugins; deployment.sh copies	Copies ("every deployed artifact is a copy")	Log of deployed paths for uninstall; no edit detection
memwiki, ai-memex-cli	npm CLIs with embedded instruction text / skill templates	init copies into project	Copies	Blind overwrite

Systems found on the web only

• beads: a SessionStart hook runs bd prime, which prints instructions from the installed binary; MCP only as fallback, citing 10–50k tokens of MCP schemas versus 1–2k for bd prime.
• spec-kit: uv tool CLI plus specify init copies templates; CLI and project files upgrade separately; managed files overwritten only if unchanged since install.
• antfu/skills-npm: skills inside npm packages, symlinked from node_modules; names "skills and tools update separately" as the problem.
• pixi-skills (pavel.pink, 2026-02-09): skills in conda packages pinned to the library version, symlinked into the project.
• Vercel npx skills: one canonical .agents/skills, other harness dirs symlinked to it; copies where symlinks fail.
• gentle-ai #4646: proposes release-time content-hash manifests and a skills verify command after stale copies broke command references.

MCP as a channel

Not the common answer for instructions. Recorded drawbacks: context cost of tool schemas (Task Master ~21k tokens), server instructions capped or dropped by some clients, and poor adherence (Serena's docs report Claude Code largely ignoring its instructions). The trend runs toward skills or CLI plus hooks for knowledge, MCP for tool access. GBrain's get_skill is the exception worth examining.

Language-native distribution

Several systems ship agent content inside their language package: Basic Memory (wheel package data), OKF Harness and memwiki (npm), hyalo (compiled into the Rust binary), skills-npm (npm), pixi-skills (conda). None makes a harness read skills from the package location directly; all copy, symlink, or print from it.