Skip to content

Latest commit

 

History

History
78 lines (53 loc) · 3.95 KB

claude-code.md

File metadata and controls

78 lines (53 loc) · 3.95 KB

EndoStack in Claude Code

Claude Code has native skills, a plugin system with marketplaces, and custom subagents. The plugin install is the best fit: skills arrive namespaced as endostack:<name>, and nothing is copied into your repository.

Install

Plugin (recommended):

claude plugin marketplace add requie/endostack
claude plugin install endostack@endostack

Inside a session the same steps are /plugin marketplace add requie/endostack and /plugin install endostack@endostack.

Files, one project:

./install.sh --scope project --agents claude-code

Files, every project:

./install.sh --scope user --agents claude-code

Run 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.

Where the files land

What Path
Plugin install Claude Code keeps its own copy of the plugin folder dist/claude-code/plugin/, which the marketplace entry in .claude-plugin/marketplace.json points at. Skills load from its skills/ and subagents from its agents/.
Skills (files) .claude/skills/<name>/ or ~/.claude/skills/<name>/
Subagents (files) .claude/agents/ or ~/.claude/agents/
Rules The project installer adds a marked @AGENTS.md import block to an existing CLAUDE.md. It never creates CLAUDE.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.

Tool names

EndoStack skills use neutral verbs. In this agent they mean:

EndoStack verb Tool in Claude Code
Spawn a subagent Agent (older name Task still works)
Ask the user AskUserQuestion
Track a todo list TodoWrite
Run a shell command Bash
Read a file Read
Edit a file Edit; new files with Write
Search the codebase Grep for contents, Glob for file names
Fetch a URL WebFetch
Load a skill Skill, or the user types /<name>
Transcripts ~/.claude/projects/<dir>/<session-id>.jsonl, where <dir> is the absolute workspace path with each / replaced by -. One JSON object per line (unverified).

"(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.

Models per subagent

Per-subagent choice works. The spawn tool takes a model value of opus, sonnet, haiku, a full model ID, or inherit. Map tiers as strong to opus, standard to sonnet, fast to haiku, unless /setup-endostack recorded other names. Reasoning effort is set per subagent with the effort field. Up to 20 subagents run at once by default (CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS).

Role assignments live in ~/.endostack/models.toml or the project's .endostack/models.toml, written by setup-endostack.

What degrades here

  • Claude Code does not read AGENTS.md when a CLAUDE.md exists. The installer adds an @AGENTS.md import to CLAUDE.md so the EndoStack rules block still loads.
  • A plugin root CLAUDE.md is never loaded, so EndoStack ships no always-on rules for the plugin install. Start each session with /tarique-mode when you want the mode.
  • If you install both the plugin and the .claude/skills/ copies, every skill appears twice. Pick one.

Non-interactive mode (used by make-bot-ui for a local bot): claude -p "<prompt>".

Sources