## Summary Add a command to `glab` that installs its own bundled agent skill file (`SKILL.md`) into the local environment, so AI coding agents can discover how to use `glab` effectively. ## Motivation In [ai-assist !4848](https://gitlab.com/gitlab-org/modelops/applied-ml/code-suggestions/ai-assist/-/merge_requests/4848), we moved the `developer_unstable` agent to rely on `glab` and native `git` via `run_command` instead of dedicated GitLab API tools. To make this work, a glab "skill" prompt was **hardcoded into the agent's system prompt** as a Jinja snippet — teaching the LLM how to use `glab` commands. This works, but it's the wrong place for that knowledge to live: - **It's static and can drift.** The prompt in ai-assist is decoupled from the glab binary version. When glab adds or changes commands, the embedded prompt doesn't update. - **It's not discoverable.** Other agents (Claude Code, Codex, Gemini CLI, etc.) can't benefit from it — it only exists inside our developer agent flow. - **It couples ai-assist to glab internals.** The agent platform shouldn't need to maintain documentation about CLI tools. The fix is to make **glab self-describing**: glab carries its own skill file and can install it into any agent's skills directory on demand. This follows the [Agent Skills open standard](https://agentskills.io) (`SKILL.md` with YAML frontmatter), which is supported by all major AI coding agents. ### The intended workflow 1. An agent environment (CI job, local dev setup, etc.) has `glab` installed and authenticated. 2. As a bootstrap step, glab installs its own skill: `glab skills install` (or `glab skills install --global`). 3. The skill file lands in `.agents/skills/glab/` (project scope) or `~/.agents/skills/glab/` (user scope). 4. Any agent that scans for skills now knows how to use glab — no hardcoded prompts needed. 5. The embedded glab prompt in ai-assist can be removed. ## Proposal ### What the command does - Writes glab's **bundled `SKILL.md` file** (embedded in the binary via `go:embed`) into the target directory. - The skill describes glab's core workflows and common gotchas that agents consistently get wrong without guidance. - No network calls, no registry — the skill ships with the binary and is always in sync with the installed version. ### Aligned implementation (incremental, from [cli!3177](https://gitlab.com/gitlab-org/cli/-/merge_requests/3177)) Based on discussion and alignment in [cli!3177#note_3298187886](https://gitlab.com/gitlab-org/cli/-/merge_requests/3177#note_3298187886), the command is `glab skills install` under the `skills` namespace: | Flag | Target directory | Use case | |---|---|---| | _(default, project scope)_ | `.agents/skills/glab/` in the current repo | Per-project setup, checked into the repo or gitignored | | `--global` | `~/.agents/skills/glab/` | User-wide, available to all agent sessions | | `--path <dir>` | Custom directory | Explicit override for non-standard setups | | `--force` | — | Overwrite existing skill files | The `--agent <name>` flag (for agent-specific paths like `~/.claude/skills/`) is deferred to a follow-up. The command is marked **experimental** — it may change or be removed as we learn more. ### Naming decision The command uses the `glab skills` namespace (not `glab agent`), as aligned with reviewers in [cli!3177](https://gitlab.com/gitlab-org/cli/-/merge_requests/3177). This avoids confusion with `glab duo` (agent behavior) and aligns with the broader `glab skills` vision in [cli#8252](https://gitlab.com/gitlab-org/cli/-/work_items/8252). ### Complementary subcommands (future, tracked in [cli#8252](https://gitlab.com/gitlab-org/cli/-/work_items/8252)) - `glab skills list` — list all bundled skills - `glab skills get <name>` — print skill content to stdout (agent-native, no filesystem write) - `glab skills install --agent <name>` — install to agent-specific directory with auto-detection ### Out of scope (future work) - Installing third-party skills from GitLab repositories - Bundled skill support in the ai-assist agent framework ([discussed in !4848](https://gitlab.com/gitlab-org/modelops/applied-ml/code-suggestions/ai-assist/-/merge_requests/4848#note_3184449245)) - Progressive disclosure / `activate_skill` tool integration - Agent auto-detection (`--agent` flag) ## Related - [cli!3177](https://gitlab.com/gitlab-org/cli/-/merge_requests/3177) — feat(skills): add `glab skills install` to install bundled agent skill (experimental) - [cli#8252](https://gitlab.com/gitlab-org/cli/-/work_items/8252) — Proposal: Add `glab skills` command for bundled AI agent skills - [ai-assist !4848](https://gitlab.com/gitlab-org/modelops/applied-ml/code-suggestions/ai-assist/-/merge_requests/4848) — feat: Run the developer_unstable on glab and git natively available - [Agent Skills specification](https://agentskills.io)