Back to Everything Claude Code

.pi — Pi Coding Agent Integration

.pi/README.md

2.2.08.5 KB
Original Source

.pi — Pi Coding Agent Integration

This directory contains the Pi adapter for ECC — a thin extension that connects the @earendil-works/pi-coding-agent terminal coding agent to ECC's canonical skills, prompts, and lifecycle hooks.

Design Principle

ECC's canonical assets—skills, agents, commands, and hooks—remain the single source of truth. This adapter contains only the integration logic. No copies, no duplication.

What This Provides

  • ECC's skills from ./skills/ — available in Pi as /skill:<name>
  • ECC's commands from ./commands/ — available in Pi as /<name>
  • ECC's engineering rules from ./rules/common/ — injected into Pi's system prompt on every turn, so coding style, testing, security, git workflow, and code-review standards apply in Pi as they do in other harnesses
  • Session lifecycle hooks — ECC's SessionStart and SessionEnd hooks, run through ECC's own run-with-flags.js, so ECC_HOOK_PROFILE and ECC_DISABLED_HOOKS keep working under Pi
  • Session context injection — whatever ECC's SessionStart hook returns as additionalContext is folded into Pi's system prompt for the next turn
  • /ecc-doctor — diagnostic command to verify the integration

Verified against Pi 0.84.1: a global install exposes 285 skills and 94 commands, resolved directly from skills/ and commands/, with no generated copies.

Installation

bash
# Install ECC as a Pi package
pi install git:github.com/affaan-m/ECC

# Or from a local checkout
pi install /path/to/ECC

# Or project-local only
pi install -l /path/to/ECC

# Verify
pi list

Then inside Pi, run /ecc-doctor to confirm skills, commands, and hooks are available.

To uninstall:

bash
pi remove git:github.com/affaan-m/ECC

Option 2: Zero-Install (Existing Claude Code Users)

If you already have ECC installed for Claude Code, point Pi at the same canonical directories from ~/.pi/agent/settings.json:

json
{
  "skills": ["~/.claude/skills"],
  "prompts": ["~/.claude/commands"]
}

This gives you skills and commands directly. It does not include the lifecycle hook adapter or /ecc-doctor — use Option 1 for the full integration.

How It Works

The extensions/index.ts file handles:

  1. Skill and command mounting — Pi reads ./skills and ./commands directly via the pi key in package.json. No transformation is needed: ECC's SKILL.md files already follow the Agent Skills standard Pi implements, and ECC's command frontmatter (description, argument-hint) is already Pi's prompt-template format
  2. Lifecycle hooks — Maps Pi's session_start to ECC's session:start hook (scripts/hooks/session-start.js) and Pi's session_shutdown to ECC's session:end:marker hook (scripts/hooks/session-end-marker.js), both invoked through scripts/hooks/run-with-flags.js so ECC's profile and disable flags are honored
  3. Rule injection — Reads ECC's portable engineering rules from the canonical rules/common/ directory at runtime and appends them to the system prompt inside an <ecc-engineering-rules> block on every turn. Nothing is copied into .pi/. agents.md, hooks.md, and performance.md are excluded on purpose: they describe Claude Code primitives Pi does not have (Task/TodoWrite delegation, Claude hook event types, thinking-budget toggles), so injecting them would point the model at tools that are not there. Language-specific rules under rules/<language>/ are not injected in this first adapter. Set ECC_PI_RULES to 0, false, off, none, or disabled to turn injection off; /ecc-doctor reports the current state and the injected size
  4. Context injection — Parses hookSpecificOutput.additionalContext from the SessionStart hook and appends it to the system prompt on the next before_agent_start, wrapped in an <ecc-session-context> block. Non-JSON hook output is tolerated, not treated as an error
  5. Hook isolation — Failing, missing, or slow hooks degrade to a warning and never terminate the Pi session. Hook execution is bounded by a timeout and an output limit
  6. Package resolution — Resolves hook scripts from the installed package via __dirname, never from process.cwd(), so a global install works from any project directory. Hooks still run in the user's project directory, so project detection stays correct

All hook execution is non-shell (execFile without shell interpretation), so paths containing spaces, tabs, or shell metacharacters are safe.

Scope

Intentionally out of scope for this first adapter (to be added independently):

  • Subagent conversion and chains (need the pi-subagents companion package)
  • Structured approval gates (need @juicesharp/rpiv-ask-user-question)
  • Persistent todos (need @juicesharp/rpiv-todo)
  • Profile-based resource filtering
  • MCP translation — see below; no translation turned out to be necessary

ECC works in Pi without any of these. Skills and commands are fully available today.

These capabilities are provided by existing community Pi packages rather than by anything ECC would need to write. This adapter deliberately does not bundle or auto-install them: bundling would ship third-party code that executes with full user permissions in every ECC install, and would make optional capabilities mandatory. Install whichever you want yourself — /ecc-doctor reports which are present and prints the exact pi install command for the ones that are not.

MCP

Pi core has no MCP surface by design. The community pi-mcp-adapter package adds one, and it reads the standard mcpServers format from .mcp.json and ~/.config/mcp/mcp.json — which is exactly the format ECC already uses in .mcp.json and mcp-configs/mcp-servers.json.

Verified against pi-mcp-adapter 2.21.2: copying ECC's mcp-configs/mcp-servers.json to a project's .mcp.json registers Pi's mcp tool and /mcp command with all 35 ECC servers discovered, alongside this adapter's own /ecc-doctor. No translation layer is needed and no ECC change is required.

bash
pi install npm:pi-mcp-adapter
cp mcp-configs/mcp-servers.json /path/to/project/.mcp.json

ECC neither installs nor depends on that package. Two caveats: the adapter's first run against a new config performs initialization that blocks in non-interactive (-p) mode, so run it once interactively before using it headless; and only server discovery was verified, not live tool invocation, which needs real credentials for each server.

Security

  • Pi extensions run with the same OS permissions as the Pi process
  • This adapter does not auto-commit, push, merge, or deploy
  • Hooks are executed without a shell, preventing command injection
  • Hook failures are isolated and cannot silently authorize blocked operations

Troubleshooting

Skills or commands not showing up

Cause: the package's resources are disabled, or a project-local install has not been trusted. Pi asks before trusting a project folder that carries its own .pi/ resources.

Fix: run pi config and confirm the ECC package's skills and prompts are enabled (<kbd>Tab</kbd> switches between user and project scope). Then confirm the package itself is registered with pi list.

/ecc-doctor not found or reports missing package root

Cause: Extension not loaded or package installed incorrectly.

Fix:

  1. Run pi list to confirm ECC is registered
  2. Restart Pi: exit and reopen the session
  3. Run /ecc-doctor again

/ecc-doctor prints the resolved package root, the skill and command counts it found, the hook runner path, the active hook profile, and which optional companion packages are present. A NOT FOUND line points at the specific path that failed to resolve.

Hooks not firing

Cause: the extension is not loaded, or the hooks are gated off by an ECC hook profile.

Fix:

  1. Confirm pi list shows ECC and that /ecc-doctor reports the hook runner as found
  2. Check ECC_HOOK_PROFILE and ECC_DISABLED_HOOKS/ecc-doctor prints both. A hook listed in ECC_DISABLED_HOOKS is skipped by design
  3. Restart Pi so the extension reloads

Notes

  • The .pi/extensions/ directory is the only place for adapter code
  • Skills and commands are defined in the repo root (skills/, commands/) and referenced by Pi
  • MCP is not bundled, but ECC's MCP configs load in Pi through the community pi-mcp-adapter — see MCP above
  • This adapter was tested against Pi v0.84.1