Skip to content

plugins.md

Latest commit

 

History

History
114 lines (94 loc) · 5.15 KB

plugins.md

File metadata and controls

114 lines (94 loc) · 5.15 KB
Raw

Plugins

infer plugins installs Claude Code-format plugins — content packages that bundle Agent Skills and an always-on instruction ruleset — and surfaces them through infer's native primitives. The Claude Code plugin layout is the closest thing to a de-facto packaging standard across agent CLIs, so plugins published for that ecosystem (e.g. ponytail) install into infer unchanged.

infer plugins install DietrichGebert/ponytail

What gets installed

A plugin repo is mapped onto native infer features — content only:

Plugin component infer feature
skills/<name>/SKILL.md Native skills system, new plugin scope (autocomplete, /name invocation, infer skills list)
AGENTS.md (repo root) Injected into the system prompt as a labeled PLUGIN INSTRUCTIONS (<name>) section while the plugin is enabled
hooks.yaml (repo root) Native Infer command hooks (disabled by default; enable with infer plugins enable-hooks <name>)
.claude-plugin/plugin.json Name/version/description metadata (optional — falls back to the repo name)
hooks/, commands/, agents/ Detected and reported, never executed or installed

Only that mapped subset is downloaded. infer deliberately does not adopt executable plugin models (e.g. OpenCode's JavaScript plugins): a plugin can add instructions and skills, but it never runs code on your machine by default. The one exception is strictly opt-in: a plugin-root hooks.yaml in infer's native command-hook format is installed disabled, must be enabled per plugin with infer plugins enable-hooks <name> (the master hooks.enabled switch still applies on top), merges additively after your own hooks.yaml (a plugin can never override or remove a user hook), and every command is gated by the per-mode bash allow-list exactly like a model-proposed command — off-list commands are skipped and reported, never run. Hook points are validated at install time and each hook is listed in the install summary before you confirm. Claude Code hooks/*.js remains detected-only, never executed.

Because a plugin's AGENTS.md becomes always-on prompt content, install prints a summary of what the plugin provides and asks for confirmation (--yes skips it; non-interactive stdin requires --yes).

Commands

infer plugins install <owner/repo | github-url | local-path> [--ref <ref>] [--yes] [--overwrite]
infer plugins list [--format text|json]
infer plugins update [<name>]        # re-fetch from the install source
infer plugins enable <name>
infer plugins disable <name>         # skills unloaded, instructions removed
infer plugins enable-hooks <name>    # opt in to plugin command hooks (off by default)
infer plugins disable-hooks <name>   # opt out of plugin command hooks
infer plugins remove <name>

Sources: owner/repo (optionally owner/repo@ref), a github.com URL (optionally /tree/<ref>), or a local directory (./my-plugin) for plugin development. Without a ref, HEAD (the default branch) is used. Downloads go through the GitHub tree + raw APIs — no git binary needed; set GITHUB_TOKEN/GH_TOKEN for higher rate limits or private repos.

Storage and layering

Plugins are userspace-only: content lives under ~/.infer/plugins/<name>/ and the registry is ~/.infer/plugins.yaml, created on first install (never seeded or overwritten by infer init).

  • Skills precedence: project (.infer/skills) → .agents/skills → user (~/.infer/skills) → plugins. A local skill with the same name always overrides a plugin's.
  • Instructions merge, never replace (matching the AGENTS.md standard): base prompt → custom_instructions → your project AGENTS.md → each enabled plugin's AGENTS.md, in registry order.
  • Instruction files are read verbatim (no environment-variable expansion) and capped at max_instructions_lines (default 399, per the AGENTS.md standard) with a max_instructions_chars backstop (default 8000); truncation is always explicitly marked.
  • In Claude Code mode, plugin rulesets ride --append-system-prompt; your project AGENTS.md is not appended there because the claude CLI reads it natively.

Configuration

~/.infer/plugins.yaml:

---
enabled: true                  # master switch (INFER_PLUGINS_ENABLED)
# dir: /custom/path            # storage override (INFER_PLUGINS_DIR)
# max_instructions_lines: 399
# max_instructions_chars: 8000
plugins:
  - name: ponytail
    source: DietrichGebert/ponytail
    version: 4.8.4
    enabled: true
    # hooks_enabled: true       # per-plugin opt-in for the plugin's hooks.yaml

Related toggles:

  • agent.skills.enabled: false stops plugin skills from loading (plugin instructions still inject; disable the plugin or set INFER_PLUGINS_ENABLED=false to remove everything).
  • agent.agents_md.enabled / max_lines / max_chars control the native injection of your own project AGENTS.md (INFER_AGENT_AGENTS_MD_*).

Verifying what the model sees

infer skills list                          # plugin skills appear with scope "plugin"
infer debug agent system_prompt | grep -A3 "PLUGIN INSTRUCTIONS"