pi

name	pi
description	Ask the pi coding-agent CLI for a second opinion. NOT for routine lookups (use grep/read/recall-memories). NOT a Claude Agent and NOT codex — this is badlogic's `pi` CLI (a distinct model, complements /codex).
when_to_use	pi, ask pi, second opinion from pi, tricky algorithm, unfamiliar library, sanity check, architecture decision, disagreement after reasoning, cross-check codex. NOT for routine lookups
user-invocable	true

Pi

Runs pi -p as a subprocess for a one-shot second opinion. pi is the Pi coding agent (badlogic/earendil, pi.dev) — same shape as codex: one binary, agent loop, its own tools (read/bash/edit/write). Methodology + citations: docs/pi/research-pi.md.

Why alongside /codex: pi answers from a different model than this Claude harness (its default here is codex-serve/gpt-5.5; configurable to Gemini/GPT/Claude), so it's a real second opinion. Use it to cross-check codex or when you want a non-Claude view. NEVER use a raw Agent(...) for a second opinion — use this skill or /codex.

Invoke

pi runs in YOLO / full-access mode by default — no permission prompts, full filesystem + shell. That's the codex --dangerously-bypass-approvals-and-sandbox behavior built in, so NO sandbox flag is needed. The container is the perimeter.

# There is NO reliable static auth check (no `login status`; a
# ~/.pi/agent/settings.json can exist with defaultProvider but NO credentials —
# that alone does NOT mean logged in). The actual `-p` call IS the check:
# a `404 page not found` body or non-zero exit = not authed / endpoint down.

# --no-session: don't persist a rollout file (a batch loop fills the disk otherwise)
# --thinking high: match codex's "newest at high effort"
# </dev/null is REQUIRED — without it pi blocks waiting for more stdin
pi -p --no-session --thinking high "Goal: <X>. Find the flaw in..." </dev/null

ALWAYS treat a non-zero exit OR a body like 404 page not found as pi unavailable (its subscription serve endpoint can be down/expired) — say so and continue WITHOUT crashing the turn. NEVER let a pi failure abort your work.

NEVER pkill -f pi to clean up — pi is a 2-char substring that matches unrelated processes (and your own shell). Kill by numeric PID (ps -eo pid,args | grep -F 'pi -p' | grep -v grep).

Model — newest available, high effort

• Discover with pi --list-models; it prints the providers/models the current auth can reach (here: claude-serve haiku/sonnet + a codex-serve gpt model).
• ALWAYS inherit pi's configured default (defaultProvider/defaultModel in ~/.pi/agent/settings.json) unless you have a reason to switch — it pins the strongest configured model. Add --thinking high (or xhigh) for max reasoning.
• To force a specific tier: --model sonnet:high, or --provider openai --model <id>. --model <provider>/<id> and --model <id>:<effort> shorthands work.
• pi's OUT-OF-BOX default (fresh install, no login) is google/gemini-2.5-flash, which needs GEMINI_API_KEY. For a second opinion, a non-Claude model is a feature — pick the strongest one available.

Auth — no status command

pi has NO non-interactive login status (unlike codex). Two working paths:

• Subscription login — cached in ~/.pi/agent/settings.json (Claude Pro / ChatGPT Plus / Copilot via codex-serve/claude-serve). Presence of that file = probably logged in; confirm only via an actual -p call.
• BYOK env var — ANTHROPIC_API_KEY, GEMINI_API_KEY, OPENAI_API_KEY, etc., or --api-key <key> per run.

If neither is present, tell the user to export <PROVIDER>_API_KEY or run pi once and /login. ALWAYS report "pi isn't configured" and continue. NEVER crash.

Rules

pi is a peer agent — it has tools and sees the repo. Give it the goal and entry points, then let it explore. NEVER pre-chew the answer or walk it through steps. For open-ended tasks give a high-level goal and let pi research it.

Adversarial framing

Like any model, pi is sycophantic — confirming questions get confirming answers.

• ALWAYS attack your own conclusion: "Find the flaw in X", "Why would this break?"
• NEVER ask "is X correct?" / "does this look right?" — primes a yes.

Prompt contents

• ALWAYS state the goal in one line ("Goal: ").
• Targeted questions: hand it entry points (file paths, symbols, error output).
• NEVER paste your reasoning chain or conclusions — biases the second opinion.

To restrict what pi can touch (read-only-ish review): --tools read,grep,find,ls or --no-tools. -a/--approve govern project file-trust, NOT tool gating.

ALWAYS verify pi's claim against the codebase before acting. Discard with a one-line reason if wrong; cite when acting.

Output

pi -p "<prompt>" writes the final text to stdout. --mode json emits a JSON-lines event stream; --mode rpc is for IDE embedding (not for one-shots). Treat the answer as advisory. Cite when acting on it.