sonichi / claude-pi

Claude Pi

Delegate work from Claude Code to the local pi CLI (pi.dev). This skill uses whatever
authentication Pi is already logged into on this machine (~/.pi/agent/auth.json — OAuth
for Kimi, Claude Pro/Max, or ChatGPT). It does not copy or export secrets.

Usage: /claude-pi [prompt]

ARGUMENTS: $ARGUMENTS

When to Use

• "Use pi on this repo" / "ask kimi"
• Need a second opinion from a Kimi-family model (default: kimi-coding/kimi-for-coding)
• Need a non-Anthropic, non-OpenAI, non-Google perspective on a design or diff
• Need a scripted, non-interactive agent pass (pi -p) in the current workspace

Guardrails

• Default provider/model is kimi-coding/kimi-for-coding; override with --provider/--model.
• Use --read-only for analysis passes — it restricts Pi to its read tool (no bash/edit/write).

• Without --read-only, Pi has full tools (read, bash, edit, write) in the working directory —
  same trust level as a default Codex workspace-write run.

• Keep Pi in the target repo via --cd; prefer --mode json when another tool consumes the output.

Quick Checks

bash "$SKILL_DIR/scripts/pi-run.sh" --check

Common Commands

# Read-only analysis
bash "$SKILL_DIR/scripts/pi-run.sh" --read-only -- "Trace how tasks flow from voice input to execution"

# Different provider/model (any Pi is logged into)
bash "$SKILL_DIR/scripts/pi-run.sh" --provider kimi-coding --model k3 -- "Review the repo structure"

# Machine-readable output
bash "$SKILL_DIR/scripts/pi-run.sh" --mode json -- "Summarize risks in src/startup.sh"

# Allow edits when the user asked for implementation help
bash "$SKILL_DIR/scripts/pi-run.sh" -- "Implement a safer startup preflight for missing services"

If Invoked As A Slash Command

• If ARGUMENTS is empty, explain the available modes and suggest --read-only for analysis.
• If ARGUMENTS is present, run:

bash "$SKILL_DIR/scripts/pi-run.sh" -- "$ARGUMENTS"