pi-add-dir

Add external directories to your pi session

Install · Usage · How it works

Add directories from outside your current working directory to a pi session. Their AGENTS.md, CLAUDE.md, and skills are automatically loaded into context every turn — so the agent understands both projects at once.

Skills from external directories are registered natively as /skill:name commands via pi's resources_discover event.

---

Install

pi install npm:pi-add-dir

Or from git:

pi install https://github.com/itisbryan/pi-add-dir

Then /reload in pi.

Usage

Commands

Command	Description
/add-dir <path>	Add an external directory to this session
/add-dir	Interactive mode — shows smart suggestions based on project structure
/suggest-dirs	Show all directory suggestions with relevance scores
/remove-dir [path]	Remove a directory (interactive picker if no path, tab-completion supported)
/dirs	List all added directories with their detected context

Examples

/add-dir /Users/me/other-project
/add-dir ../shared-library
/add-dir ~/Desktop/design-system
/dirs
/remove-dir /Users/me/other-project

LLM Tools

Two tools are available for the agent:

Tool	Description
add_directory	Add an external directory (loads its AGENTS.md, skills, etc.)
search_external_files	Search for files across all external directories by name/pattern

The agent can request adding a directory on its own:

"I need to reference the shared library at /Users/me/libs/core — let me add it to the session."

And search across external dirs when @ file picker isn't available:

"Let me search for config files across the external directories."

Smart Suggestions

When you run /add-dir without arguments, the extension analyzes your project structure and suggests relevant directories:

• Workspace members — npm, pnpm, Cargo, Go, Python (uv), and monorepo packages
• Local dependencies — file:/link:/portal: in package.json, path: in Gemfile/Cargo.toml
• Git submodules — paths from .gitmodules
• Sibling projects — related projects alongside your cwd
• TypeScript project references — references in tsconfig.json
• Docker Compose services — build.context paths
• Gradle project modules — include() in settings.gradle(.kts)
• Maven multi-module — <modules> in pom.xml
• uv Python workspaces — [tool.uv.workspace] members in pyproject.toml
• .NET solutions — project references in .sln files
• PHP Composer — path repository references in composer.json
• Flutter/Dart — path: dependencies in pubspec.yaml
• Swift PM — .package(path:) local dependencies in Package.swift
• Elixir — {:dep, path: "..."} in mix.exs
• Context-rich directories — prioritizes dirs with AGENTS.md, CLAUDE.md, or skills

Directories with context files get higher relevance scores, making the most useful suggestions appear first.

Widget

When directories are added, a widget appears above the editor:

📂 2 external dirs │ other-project, shared-library  (/dirs to manage)

The widget automatically truncates to fit your terminal width.

How It Works

┌─────────────────────────────────────────────────────────┐
│  pi session (cwd: /my-project)                          │
│                                                         │
│  /add-dir /other-project                                │
│     │                                                   │
│     ├─► Scans /other-project for:                       │
│     │     AGENTS.md, CLAUDE.md                          │
│     │     .pi/skills/, .agents/skills/, .claude/skills/ │
│     │     .pi/extensions/ (detection only)              │
│     │                                                   │
│     ├─► Persists to session (survives restart)          │
│     │                                                   │
│     ├─► Registers skills via resources_discover         │
│     │   (auto-reloads pi to activate /skill:name)       │
│     │                                                   │
│     └─► Every turn: injects found context files         │
│         into the system prompt via                      │
│         before_agent_start event                        │
│                                                         │
│  Agent now knows both projects' rules & conventions     │
│  Agent can read/edit/write files in /other-project      │
│  using absolute paths (always worked — now it has       │
│  the context to do it intelligently)                    │
│                                                         │
│  Skills from /other-project work as /skill:name         │
│  commands, just like local skills                       │
└─────────────────────────────────────────────────────────┘

What gets injected

For each added directory, the extension reads:

File	Location(s) checked
AGENTS.md	<dir>/AGENTS.md, <dir>/.pi/AGENTS.md
CLAUDE.md	<dir>/CLAUDE.md, <dir>/.pi/CLAUDE.md
Skills	<dir>/.pi/skills/*/SKILL.md, <dir>/.agents/skills/*/SKILL.md, <dir>/.claude/skills/*/SKILL.md

Context files are appended to the system prompt on every turn (cached — filesystem is only re-scanned when directories change).

Skills are registered natively with pi via the resources_discover event, so they appear as /skill:name commands with full autocomplete support.

What's persisted

Added directories are stored in the session via pi.appendEntry(). When you /resume a session, the directories are automatically restored.

A temp file (/tmp/pi-add-dir-<hash>.json) is also maintained so resources_discover can read the directory list before the session is fully loaded. This file is automatically synced with session state and cleaned up on session_shutdown.

Extension detection

When adding a directory that contains .pi/extensions/, the extension detects them and shows actionable instructions:

Found 2 extension(s) in other-project/.pi/extensions/.
   To enable them, add to your settings.json:
   { "extensions": ["/path/to/other-project/.pi/extensions"] }
   Then /reload to activate.

Extensions cannot be loaded dynamically at runtime (pi platform limitation), but the extension tells you exactly how to enable them.

Performance

Context injection is cached — the filesystem is only re-scanned when directories are added or removed, not on every turn. This keeps the before_agent_start hook fast.

Auto-reload behavior

When you add or remove a directory that contains skills, pi automatically reloads to register/unregister those skills as /skill:name commands. This is a brief operation — your session state is preserved.

Limitations

Works ✅	Platform limitation ❌
AGENTS.md / CLAUDE.md loaded into system prompt	@ file fuzzy search doesn't include external dirs¹
Skills registered as native /skill:name commands	External .pi/extensions/ are not auto-loaded²
Agent can read/edit/write any path	Can't change ctx.cwd at runtime³
Persists across session restarts
File search across external dirs via LLM tool
Extension detection with setup instructions

Notes:

1. The @ file picker is a TUI-level feature that can't be extended from an extension. Use the search_external_files tool as a workaround — the agent can call it to find files across all external directories.
2. Pi doesn't support loading extensions dynamically at runtime. When extensions are detected, you'll see instructions to add them to settings.json manually.
3. ctx.cwd is read-only in the pi extension API. The agent uses absolute paths for external directories, which works well in practice.

Changelog

See CHANGELOG.md for release history.

License

MIT