Phantom Skills

phantom-skills (`pskillctl`)

A self-improving **skills daemon** for the phantom control-server fleet. It watches an agent's real activity, detects when a reusable [Claude Code skill](https://code.claude.com/docs/en/skills) would help, authors it, verifies it actually helps (held-out grading), prunes stale/duplicate ones, and promotes proven skills from a machine's local `~/.claude/skills/` up to a shared registry that syncs back down to every machine.

Every create/prune/promote is **human-gated by default**, with per-action-type autonomous opt-in earned once the gate's approval rate proves the automated standard is trustworthy.

It's a meta-skill: it exists to keep the skill library small, sharp, and shared — so context loads what's needed and nothing that's rotted.

Why

Claude Code loads skills via progressive disclosure — ~100 tokens of each skill's `description` sit in context always, and the full body loads only when the model matches a task to that description. That keeps the window small, but today skills are hand-authored, never pruned, and never shared across machines. A survey of ~25 prior systems (Voyager, TROVE, Ada, ExpeL, Anthropic `skill-creator`, `crune`, …) found nobody ships the full combination: continuous background *watching* + usage-and-success *pruning* + automatic cross-agent *promotion* + held-out *verification*, with a light human-in-loop. That combination maps directly onto phantom-brain's promote-to-shared-memory model.

Architecture

One Go binary, two roles (mirrors `pbrainctl`):

• pskillctl client mcp — per-session stdio MCP server spawned by Claude Code. Exposes the skill_* tools, tails the session transcript, syncs skills into ~/.claude/skills/, and proxies reads/writes to the daemon over HTTP (with a local SQLite write-ahead queue for offline resilience).
• pskillctl server serve — the control-server HTTP daemon. Owns the shared skill registry (Postgres SoR), runs the background pipeline (detect → author