PortPilot Assistant

Translate natural language into portpilot commands with safe defaults.

Workflow

1. Classify intent into one of: scan, who, pick, doctor, free, init, config.
2. Resolve target port/range/path from user text; if absent, apply defaults.
3. For the first portpilot call in a new session, run one permission bootstrap to avoid repeated prompts (details below).
4. Execute read actions directly (scan, who, pick, doctor).
5. For write actions (free, init --force, config migrate), require explicit confirmation and show details.
6. Return concise conclusion first, then command output details.

Intent mapping

• 帮我选个端口 / pick a port:
  • portpilot pick --range 3000-3999 --count 1 --lease-ms 20000 --json
• 扫描当前端口 / scan current used ports:
  • portpilot scan --protocol both --json
• 查看 3000 端口 / check port 3000:
  • portpilot who 3000 --json
• 释放 3000 端口 / free port 3000:
  • First portpilot who 3000 --json, then confirm, then portpilot free 3000 --yes --json
• 扫描冲突 / run doctor:
  • portpilot doctor --json
• 初始化配置 / init port config:
  • portpilot init --dry-run --json (default preview)

Safety policy

• Read actions run automatically.
• Write actions require confirmation.
• Confirmation payload must include: port, pid, command, cwd, startTime, action.
• If dependency failure occurs, run portpilot doctor --preflight --json and return install/fix commands.

Permission bootstrap (one-time)

• On the first portpilot action in a new session, proactively run one lightweight bootstrap command (portpilot doctor --json) with escalation.
• Resolve bundled CLI absolute path first (path-independent):
  • Preferred: $CODEX_HOME/skills/portpilot-assistant/assets/portpilot/bin/portpilot.js
  • Fallback for repo-local usage: <repo>/.claude/skills/portpilot-assistant/assets/portpilot/bin/portpilot.js
• In that one-time escalation request, set:
  • sandbox_permissions: require_escalated
  • prefix_rule: ["node", "<resolved-portpilot-cli-absolute-path>"]
• Explain that this is a one-time approval to avoid repeated prompts for future portpilot commands.
• After approval, run requested commands normally without escalation.
• If bootstrap was not done (or approval was denied), and a read action fails with permission errors (Operation not permitted / netlink denied / lsof denied), rerun once with escalation.
• Write actions still require explicit user confirmation before execution.

Execution notes

• Run bundled CLI from this skill first: assets/portpilot/bin/portpilot.js.
• Fallback to npx only if bundled CLI is missing: npx -y portpilot-cli@1 ....
• Always pass --json for machine-readable parsing in skill scripts.