Gemini CLI installs EndoStack as an extension (gemini-extension.json at the repository root) and reads skills, subagents, and TOML commands. Google has moved consumer accounts to Antigravity CLI; Gemini CLI still works with paid Gemini API keys and Code Assist Standard or Enterprise.
Extension (recommended):
gemini extensions install https://github.com/requie/endostackFiles, one project:
./install.sh --scope project --agents gemini-cliFiles, every project:
./install.sh --scope user --agents gemini-cliRun the installer from a clone of the EndoStack repository. Add --dry-run to see every file it would write, --link to symlink instead of copy, and --uninstall (with the same --scope) to remove exactly what it wrote. On Windows use install.ps1 with the same options.
| What | Path |
|---|---|
| Extension install | Skills from the extension's skills/; context from GEMINI.md, which maps EndoStack's verbs to Gemini tools. |
| Skills (files) | .agents/skills/<name>/ or ~/.agents/skills/<name>/ (both win over .gemini/skills/) |
| Subagents (files) | .gemini/agents/ or ~/.gemini/agents/ |
| Commands (files) | .gemini/commands/<name>.toml or ~/.gemini/commands/<name>.toml, one per command-style skill; {{args}} carries what you type after the command |
| Rules | A marked block in AGENTS.md, plus an @AGENTS.md import block in an existing project GEMINI.md. Gemini reads AGENTS.md directly only if context.fileName in settings.json lists it. |
The installer records every file it writes, with a hash, in .endostack/install-manifest.json (project) or ~/.endostack/install-manifest.json (user). It never overwrites a file it did not write.
EndoStack skills use neutral verbs. In this agent they mean:
| EndoStack verb | Tool in Gemini CLI |
|---|---|
| Spawn a subagent | a tool named after each subagent (for example tarique-agent); @<name> forces delegation |
| Ask the user | ask_user |
| Track a todo list | write_todos |
| Run a shell command | run_shell_command |
| Read a file | read_file |
| Edit a file | replace; new files with write_file |
| Search the codebase | grep_search, glob |
| Fetch a URL | web_fetch |
| Load a skill | activate_skill (asks you to confirm before the skill body loads) |
| Transcripts | ~/.gemini/tmp/<project-hash>/chats/, one folder per project. The hash function and file format are not documented. |
"(unverified)" marks a name or path taken from a third-party source or prior knowledge rather than the vendor's docs. Where a verb has no tool, skills fall back as they describe: plain-text questions, a checklist in the reply, or doing subagent briefs one after another.
Model choice is per subagent definition (the model field). EndoStack's generated agents leave it unset, so they run on the session model unless you add one.
Role assignments live in ~/.endostack/models.toml or the project's .endostack/models.toml, written by setup-endostack.
- Subagents cannot spawn subagents. Fan-outs that a delegate would start run from the main chat instead.
- Parallel subagents are described in Google's launch post but not in the reference docs. If calls run one at a time, skills still finish, only slower.
- Every skill activation asks for your confirmation.
- The extension install also loads the canonical
agents/*.mdfiles, whosetoolsvalue is prose rather than a tool list. If Gemini rejects them, use the files install, which writes Gemini-format agents to.gemini/agents/.
Non-interactive mode (used by make-bot-ui for a local bot): gemini -p "<prompt>".