GitHub skill-group FAQ

Practical answers for GitHub agent-skill-group, Agent Skills, MCP, and plugin questions.

This FAQ explains what agent-skill-groups is and how to test it safely.

What is a GitHub skill-group?

In this project, a GitHub skill-group is a scenario profile for local SKILL.md Agent Skills, including skills installed from GitHub repositories. The point is to keep a small core profile active and load specialized skill groups only when a task needs them.

What is a GitHub agent-skill-group manager?

agent-skill-groups scans local Agent Skills, suggests groups, writes managed AGENTS.md or CLAUDE.md snippets, renders session-load and session-unload packs for the current conversation, backs up state, switches profiles, and restores skill directories without editing skill contents.

Is it a skill, an MCP server, or a plugin?

The core project is a Python CLI. It also includes a lightweight Agent Skill entry and a Copilot-compatible plugin package so agent runtimes and directories can discover the workflow. It is not an MCP server because it does not expose MCP tools or resources.

How do I test it safely?

Install the latest release wheel and run the one-command demo. The demo uses a sample directory, not your real skill roots.

python -m pip install https://github.com/go165/agent-skill-groups/releases/download/v0.3.52/agent_skill_groups-0.3.52-py3-none-any.whl
agent-skill-groups demo --json

The expected scriptable JSON shape is published at examples/demo-output.json.

Which runtimes are supported?

The CLI includes presets for OpenAI Codex, Claude Code, OpenCode, and generic Agent Skills layouts. Custom runtime roots can also be defined in groups.json.

Can it load a group in the current conversation?

Yes. Run agent-skill-groups session-load --config groups.json <group> and feed the generated context pack to the current agent. This reads SKILL.md files from both active roots and the disabled pool, so it works without moving directories.

If the host already supports native progressive disclosure, use --context-mode summary to emit only group and skill metadata, or --context-mode native-only with Codex direct activation to emit no skill bodies at all. Use full mode only when the current conversation needs complete skill instructions pasted in.

To stop applying that group later in the same conversation, run agent-skill-groups session-unload --config groups.json <group> and feed the generated stop pack to the current agent. Current-session unload cannot erase prior model context; it changes behavior from that point forward.

profile remains useful for native skill discovery after a runtime reload, rescan, or new session.

Can it rewrite Codex native skills in an already-running CLI conversation?

No user-space manager can erase or replace the native skills block after Codex has already sent that block to the model in the current running CLI conversation. Use session-load and session-unload when the current conversation needs a group immediately.

agent-skill-groups codex-native --json
agent-skill-groups codex-native --probe-app-server --json
agent-skill-groups codex-refresh --config groups.json --runtime codex ctf-web --json
agent-skill-groups codex-activate --config groups.json --runtime codex --direct --context-mode native-only ctf-web --json
agent-skill-groups codex-direct --config groups.json --runtime codex ctf-web --json
agent-skill-groups codex-clear --config groups.json --runtime codex ctf-web --json

These probes report native discovery signals such as codex debug prompt-input and app-server skills/list with forceReload. codex-refresh applies a filesystem profile and immediately verifies it with app-server skills/list forceReload, reporting matchedSkills, missingSkills, and native errors. codex-direct updates app-server skills/extraRoots with a direct group root without moving skill directories. codex-clear sends an empty skills/extraRoots list, force-reloads native metadata again, and reports remainingManagedGroupSkills when another native root still exposes requested group skills. codex-activate --direct --context-mode native-only updates Codex native metadata without returning full skill bodies; summary and full are available when the current conversation needs emitted context. They prove fresh metadata scan support where Codex exposes it; they do not retroactively rewrite model-visible history.

Related Links