Connect AIWG to Pi Coding Agent

Prompt-first procedure: Describe the outcome you want in your agent conversation. The agent should select and load the appropriate AIWG assets, explain material changes, request any needed approval, and report verification evidence. Exact commands and flags appear only in the CLI reference.

Upstream baseline: Pi Coding Agent 0.85.0, commit

`47236c84450656043dd8fb21c8513d1421505ae3`,

verified 2026-09-04.

Pi Coding Agent is the minimal agent harness published at pi.dev. AIWG's `pi` provider targets that harness directly; `pi.dev` is not a separate provider ID.

Pi is an experimental AIWG deployment provider. AIWG installs declarative context, skills, prompt templates, and a reviewed trust-gated policy extension; it does not install Pi packages or credentials. Start with Install, Connect, and Verify for the general AIWG setup journey.

Install, preview, and deploy

Pi 0.85.0 requires Node.js 22.19 or newer and is published as `@earendil-works/pi-coding-agent`.

Use AIWG to complete this documented outcome: Pi 0.85.0 requires Node.js 22.19 or newer and is published as @earendil-works/pi-coding-agent.
Have it inspect the current state, explain the plan, ask before material
changes, and report the result with verification evidence.

The deployment command refreshes AIWG's shared project context and prints a verification result. Follow any reload notice in that output before starting a Pi session.

Here, AIWG's `the provider option pi` selects the coding-agent integration. It is not Pi's own the provider option option, which selects an LLM backend such as OpenRouter. Pi documents its model flags in the coding-agent README.

Trust and verification

Project-local settings, extensions, skills, and prompts are trust-gated; context files such as `AGENTS.md` load independently of that decision. In interactive Pi, run `/trust` and restart after approving the project. Non-interactive modes do not show the trust dialog, so every automated invocation must choose explicitly between the approve option and the no-approve option; see Pi's project trust documentation.

Use AIWG to complete this documented outcome: Project-local settings, extensions, skills, and prompts are trust-gated; context files such as AGENTS.md load independently of that decision. In interactive Pi, run /trust and restart after approving the project. Non-interactive modes do no
Have it inspect the current state, explain the plan, ask before material
changes, and report the result with verification evidence.

Try one small task immediately after trust and verification:

Review this project's README and getting-started docs for unclear positioning,
missing setup steps, or unsupported claims. Return the three highest-priority
fixes with file references and a recommended next edit.

Success means Pi names the intended project, loads `AGENTS.md` and the trusted AIWG project resources, and produces a concrete review you can inspect.

For a bounded, non-interactive check that loads trusted project resources:

PI_CODING_AGENT_SESSION_DIR="$(mktemp -d)" \
  pi the approve option the mode option json -p "List the AIWG skills available in this project"

This may call the selected model and consume tokens. Use the no-approve option when testing without project-local resources.

Installed resources

Resource	Location	Purpose
Project context	`AGENTS.md`	Directs Pi to `WORKSPACE.md` and `AIWG.md`
Portable skills and roles	`.agents/skills/*/SKILL.md`	Pi's shared Agent Skills surface
AIWG-managed skills	`.pi/.aiwg/skills/`	Receipted AIWG projection
Prompt templates	`.pi/prompts/*.md`	User-facing AIWG commands
Policy bridge	`.pi/extensions/aiwg-bridge.ts`	Fail-closed headless command policy

Pi discovers project skills in `.pi/skills/` and `.agents/skills/`, as documented in Skills. Prompt templates live in `.pi/prompts/`; AIWG flattens generated prompt names because Pi's prompt-template discovery is non-recursive.

Refresh and remove

Use AIWG to complete this documented outcome: Refresh and remove
Have it inspect the current state, explain the plan, ask before material
changes, and report the result with verification evidence.

the agent-owned remove operation operates on an installed framework or addon, not on a provider alone. Repeat it for each installed item you intend to remove. AIWG only removes receipted files and preserves operator-owned `.pi/settings.json`, prompts, skills, extensions, packages, sessions, and Pi's trust decisions.

For user scope, session handling, OpenRouter examples, limitations, and troubleshooting, see the Pi operational reference.

External Ralph can launch Pi with `the provider option pi`; pass a Pi-native `provider/model` identifier using the model option, and optionally the thinking option and the tools option. Session ingestion supports authorized Pi v3 JSONL roots; see Pi session acquisition.