Codex reads skills from .agents/skills/ and custom agents from TOML files in .codex/agents/. The repository root plugin.json is an Agent Plugins manifest with Codex settings under extensions["com.openai"].
Plugin:
codex plugin marketplace add requie/endostackThen install EndoStack from the Plugins Directory in the Codex app.
Files, one project:
./install.sh --scope project --agents codexFiles, every project:
./install.sh --scope user --agents codexRun 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 |
|---|---|
| Skills | .agents/skills/<name>/ or ~/.agents/skills/<name>/ |
| Command-style skills | Each also gets agents/openai.yaml with allow_implicit_invocation: false, so Codex runs it only when you type $<name>. With --link this file cannot be added and those skills become model-invocable. |
| Subagents | .codex/agents/<name>.toml or ~/.codex/agents/<name>.toml (fields name, description, developer_instructions). Project agents load only in trusted projects. |
| Rules | A marked block in the project AGENTS.md |
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 Codex CLI |
|---|---|
| Spawn a subagent | spawn_agent (unverified) |
| Ask the user | not documented |
| Track a todo list | update_plan (unverified) |
| Run a shell command | shell_command (unverified) |
| Read a file | read_file (unverified) |
| Edit a file | apply_patch |
| Search the codebase | rg and glob_file_search (unverified) |
| Fetch a URL | none; web_search only |
| Load a skill | $<name> or /skills |
| Transcripts | ~/.codex/sessions/<YYYY>/<MM>/<DD>/rollout-*.jsonl (unverified). Files are grouped by date, so filter by the working directory recorded inside each file. |
"(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 agent definition: the model and model_reasoning_effort fields in an agent TOML file. The spawn call itself takes no model. EndoStack's generated agents set neither field, so they run on the session model. To pin a role, add model to the installed TOML file (the uninstaller then keeps that file because you changed it).
Role assignments live in ~/.endostack/models.toml or the project's .endostack/models.toml, written by setup-endostack.
- No ask-user tool is confirmed. Skills ask in plain text and stop until you answer.
- No fetch-by-URL tool. Skills that need a page ask you for it or use
curlwhen the sandbox allows network access. - Up to 6 subagents run at once by default (
[agents] max_threads), and subagents cannot spawn their own by default (max_depth1). Larger fan-outs inswarmandsecurity-reviewrun in waves. - Multi-model panels (
arena,architect,interrogate) run every seat on the session model and vary each seat's brief instead.
Non-interactive mode (used by make-bot-ui for a local bot): codex exec "<prompt>".