stevengonsalvez/agents-in-a-box

LLM-facing skill teaching agents how to use `ainb fleet ...` orchestration subcommands. Fleet covers multi-session broadcast, ack-gated sequence, blocked-session needs, and the auto-continue daemon. Replaces the deprecated `popa` plugin — the Node code is gone; the orchestration logic now lives in ainb's Rust binary.

Emit actionable host-agent lifecycle events to the ainb notification inbox: Claude/Copilot Notification, Codex PermissionRequest, and Stop/agentStop are delivered to ainb-notifyd over a Unix socket, surfacing as native OS banners, the ainb-tui Inbox screen, and per-session attention markers. Telemetry events are intentionally not hooked. Install via `/plugin install ainb-hooks@agents-in-a-box` (Claude); Codex/Copilot wiring is handled by `ainb notifyd install`.

Caveman mode lifecycle hooks. PostToolUse watermark-tracks output tokens per session and writes .caveman-statusline-suffix so the statusline badge updates live. PreCompact re-injects the full caveman ruleset into the compaction summary so mode survives context compaction. Pair with caveman@caveman for full caveman setup.

Explanatory & brand illustration skill starring Popa — a pink blob with a notepad — drawn as intricate hand-drawn sketchnotes. Digest → shot list → per-image generation → QA → deliver. Invoke `illustration` or `illustration:popa`.

Agent self-improvement + retrieval. Captures learnings via colon-namespaced sub-skills (reflect, reflect:consolidate, reflect:ingest, reflect:recall, reflect-status) and auto-injects relevant prior learnings via SessionStart/UserPromptSubmit hooks. PostToolUse arms low-cost mini-learning capture; Stop enqueues short-session reflection. Philosophy: Correct once, never again; recall everything next time.